Energy-Sustainable IoT Connectivity: Vision, Technological Enablers, Challenges, and Future Directions

Technology solutions must effectively balance economic growth, social equity, and environmental integrity to achieve a sustainable society. Notably, although the Internet of Things (IoT) paradigm constitutes a key sustainability enabler, critical issues such as the increasing maintenance operations, energy consumption, and manufacturing/disposal of IoT devices have long-term negative economic, societal, and environmental impacts and must be efficiently addressed. This calls for self-sustainable IoT ecosystems requiring minimal external resources and intervention, effectively utilizing renewable energy sources, and recycling materials whenever possible, thus encompassing energy sustainability. In this work, we focus on energy-sustainable IoT during the operation phase, although our discussions sometimes extend to other sustainability aspects and IoT lifecycle phases. Specifically, we provide a fresh look at energy-sustainable IoT and identify energy provision, transfer, and energy efficiency as the three main energy-related processes whose harmonious coexistence pushes toward realizing self-sustainable IoT systems. Their main related technologies, recent advances, challenges, and research directions are also discussed. Moreover, we overview relevant performance metrics to assess the energy-sustainability potential of a certain technique, technology, device, or network, together with target values for the next generation of wireless systems, and discuss protocol, integration, and implementation issues. Overall, this paper offers insights that are valuable for advancing sustainability goals for present and future generations.


I. INTRODUCTION
T HE vision of a sustainable society is relentlessly pursued by industry, academic research, and regulatory bodies following the famous United Nations' sustainable development goals introduced in 2015 1 . In general, sustainability is stimulated by simultaneously developing the economy, promoting social equity, and protecting the integrity of the environment for current and future generations [1], as captured by the so-called sustainability triangle illustrated in Fig. 1. Therefore, associating sustainability solely with environmentally-conscious (also referred to as "green") practices is a widespread misconception since sustainability comprises Economic, Societal (Equity), and Environmental factors (also known as the 3 E's of sustainability).
Quantifying/measuring and developing standardized metrics and benchmarks for sustainability is quite a challenging endeavor because i) sustainability is a subjective concept that can be assessed differently by various stakeholders; ii) the interdependencies, trade-offs, and feedback loops between the multiple interconnected systems and 1  performance corners (economic, societal, and environmental) must be captured; and iii) pursuing specific long-term perspectives while measuring and tracking progress in the short term may not be easy and viable in general. This is why, despite numerous attempts over the years, e.g., [2]- [9], none of them have gained significant traction. Regardless of the diverse assessment criteria and methods, sustainability enablers are generally wellestablished and include Internet of Things (IoT) technologies [10]- [26]. Specifically, IoT solutions promote i) economic development (by enabling automation, facilitating accurate decision-making via massive data collection and real-time processing, optimizing resource usage and efficiency, improving maintenance processes, and reducing systems' downtime); ii) social equity and well-being (by providing enhanced safety and security measures against potential safety hazards or security breaches and by facilitating flexibility and barrier-free experience to the end users via the ubiquitous provision of services and the exploitation of remotecontrolled or completely autonomous systems); and iii) environment protection (by optimizing energy usage and reducing waste/pollution); thus advancing sustainability goals.

A. ENERGY-SUSTAINABLE IOT ECOSYSTEM
Undoubtedly, one of the critical aspects of selfsustainable IoT systems is energy. Self-sustainable solutions must be energy-sustainable, which entails green 3 energy autonomy when considering solely the operation phase of an IoT product lifecycle. Energy-sustainable IoT operation is precisely the focus of this work, although our discussions sometimes extend to other sustainability aspects and IoT lifecycle phases. Readers interested in specific discussions on sustainability at planning, manufacturing, deployment, maintenance, and disposal phases are encouraged to refer to [11], [26], [33], [38]- [43].
The main energy processes that are present in an energy-sustainable IoT ecosystem are illustrated in Fig. 1 together with the sustainability triangle and discussed below.
Energy provision (EP) refers to the charging process(es) exploiting green energy sources. Cur- 2 Refer to https://www.statista.com/statistics/1183457/ iot-connected-devices-worldwide. 3 Green energy refers to energy from renewable sources.  [10] X X green mobile edge computing (MEC) • survey of the state-of-the-art on joint radio-and-computational resource management in MEC • discussion on MEC challenges and future research directions, including MEC system deployment, cache-enabled MEC, mobility management for MEC, green MEC, and privacy-aware MEC [11] X X energy harvesting (EH), smart grids, cloud computing, RFID, machine learning (ML) • review of smart manufacturing and sustainable energy industry • discussion of prominent technologies to support sustainable industrial IoT [12] X X EH, RFID, wake-up radio (WuR), cloud computing, data/communication reduction • survey of potential technologies for greening the IoT and relevant research directions [13] X X smart microgrids, IoT for waste & energy management, ML, low-power wide-area networks, hardware (HW) and software (SW) duty cycling • review of the state of the art research and advances on IoT technologies in/for energy management, environment monitoring, transportation, e-health, and smart cities [14] X X EH, low-power wide-area networks, Bluetooth low energy (BLE), MEC, smart grids • review of applications of IoT in energy systems, especially in the context of smart grids • discussion of IoT-enabling technologies and challenges in the energy sector, including privacy and security [15] X X X EH, wireless ET (WET), smart grids, MEC, unmanned aerial vehicle (UAV)-assisted communications • overview of advances/prospects on the utilization, redistribution, trading, and planning of harvested energy in future wireless networks • review of wireless power and information transfer technologies [16] X X EH, WET • proposal of a self-sufficient IoT architecture that adopts only single-and double-hop energy and data transitions to enable efficient energy sharing and reduced data traffic • discussion of enablers and identification of future research directions [22] X X green radio-frequency (RF) WET, distributed ledger technology for energy trading, ultra-low power receivers, enhanced energy transmitters, metasurface-assisted RF-WET • discussion of RF-WET for enabling self-sustainable IoT • overview of the main architectures, challenges, and techniques for efficient and scalable RF-WET • outline of key research directions for realizing RF-powered 6G IoT [23] X X RFID, EH, TinyML HW, ML, MEC • overview of enablers, architectures, technologies, energy models, and strategies for green IoT • discussion of effective behavioral change models and strategies to create energy-awareness among users and IoT service providers [25] X X EH, energy-aware TinyML, MEC • discussions on the incorporation of tinyML algorithms and application tasks on battery-less IoT devices • study of several inference strategies, including local/cloud computing • testing a battery-free person detection prototype [26] X X X EH, RF harvesters, WET, backscattering • overview of HW-related research trends, application use cases, and enabling technologies for sustainable IoT systems • high-level discussion of eco-friendly manufacturing, sustainable WET, and low-power wireless connectivity solutions rently, the information and communication technology industry accounts for 1 − 1.5% of the global electricity use [44], a figure that could rise to nearly 4% considering the increasing traffic demands at the transmission networks and the computation requirements at the data centers [45]. Consequently, the implementation of sustainable energy in this sector will have a significant impact on the environment considering that the current average global carbon intensity factor 4 is approximately 0, 441 kg of CO 2 emissions per kWh. Furthermore, certain EP implementations enable the production of electricity at a local level, thereby making electricity (and also in-  APD] Section VI: Energy-related performance metric Section VII: Conclusions Fig. 2: Integration of the EP, ET, and EE processes throughout the structure of the paper. The terms in brackets refer to KPIs later discussed in Section VI, whereas associated challenges and research directions of the illustrated technologies are summarized in Section VII. formation and communication technology services) accessible to remote communities.
Energy transfer (ET) refers to the intentional movement of energy from one device/system to another. This is required, for instance, when green energy sources are not available for direct exploitation (or provide insufficient energy). In such cases, another device/system powered by green sources can transfer such required energy. The global market projections indicate a significant increase in the value of ET technologies, reflecting the growing interest in this field. As per [46], the global market of ET is projected to surge from USD $5, 705.1 million in 2020 to $35, 226.4 million by 2030, with a resulting average annual growth rate of 21.3%. Consequently, technological breakthroughs in ET will generate new business opportunities, enable previously infeasible IoT use cases, and facilitate the ubiquity of ET services across the globe.
Energy efficiency (EE) refers to the ability of a device/system/process to perform its intended function with minimal energy and thus is a measure of how effectively energy is used to achieve a desired outcome [47]. Obviously, EE directly impacts the environment and economy sustainability corners as it pushes to minimize EC and processing, i.e., toward more responsible energy usage. Notably, an ambitious goal for the sixth generation (6G) of wireless systems is to realize 10−100-fold EE gains with respect to the current fifth-generation (5G) networks [48]. Notice that EE designs are often lowcomplexity/cost, thus may promote sustainability in other IoT lifecycle phases as well.
The harmonious coexistence of the above processes ultimately pushes to zero dirty-EC, thus, realizing truly energy-sustainable IoT systems.

B. CONTRIBUTIONS AND ORGANIZATION OF THIS WORK
The realization of energy-sustainable IoT connectivity depends critically on the technological advancements in the EP, ET, and EE processes, and their holistic integration. Notably, the related stateof-the-art discussions are sparse in the literature as most works focus on a specific enabling technology. Only a few other works, such as those listed in Table I, have a wider scope, although still limited in terms of technology enablers, recent advances, key performance indicators, challenges, and/or research directions, and their link to sustainability. Moreover, wake-up accuracy WuR/WuS wake-up radio/signal they generally lack clarity regarding the role or classification of the discussed technologies/approaches toward supporting (self)-sustainable IoT connectivity, e.g., by associating them with the corresponding EP, ET, and EE processes.
Our work aims to address the aforementioned research gap by making the following contributions: • We provide a fresh look at energy-sustainable IoT. Moreover, we identify EP, ET, and EE as the three main energy-related processes whose harmonious coexistence pushes toward realizing self-sustainable IoT systems. This is mostly covered in Section I. • We discuss the main technologies, recent ad- vances, and challenges and associated research directions for EP (Section II), ET (Section III), and EE (including communication (Section IV) and learning/computation (Section V)-related aspects processes to support sustainable IoT connectivity. In addition, we summarize the key relevant challenges and research directions per technology in Section VII. • We present a set of energy-related performance metrics to assess the absolute/relative performance potential of a certain technique, technology, device, or network. They are classified as i) energy conversion and transfer metrics, ii) energy storage and consumption metrics, iii) EE ratio (EER) metrics, and iv) others. Moreover, we overview several related key performance indicators (KPIs) for the next generation of wireless systems. This is comprehensively covered in Section VI. Fig. 2 illustrates how the structure of the paper covers the integration of the EP, ET, and EE processes. Finally, Table II lists the acronyms used throughout this article in alphabetical order.

II. ENERGY PROVISION (EP)
The increasing demand for a reliable energy supply to power the IoT has made EH technologies appealing to reduce/eliminate the need for batteries. There are readily available energy sources in our environment such as sunlight, wind, heat, electromagnetic (EM) radiation, and many others. Table III summarizes the advantages, limitations, and attractive use cases for the EH technologies covered in this section.
To power devices' electronics, EH circuits must transform these forms into electric energy. The architecture of a generic EH device, illustrated in Fig. 3, consists of three main components: • transducer, which turns a physical variable (or its variations) into electricity; • power management unit, which accommodates the output signal of the transducer to power the devices' electronics and/or charge the storage element. One important function of this block is to dynamically match the output impedance of the transducer to a variable input impedance of a rectifying circuit or other device's peripherals; • energy storage component, e.g., a battery or a supercapacitor, which buffers the variations of the ambient energy. We refer the reader to the Section VI-VI-B for more details on this component. In general, the performance of ambient EH technologies is severely limited by the uncontrollability and unpredictability of ambient sources. To address this, some designs incorporate multiple transducers to enable simultaneous EH from different sources, e.g., [49]- [52]. This boosts the system's reliability, although increasing the hardware's complexity and cost.

A. LIGHT-BASED EH
Light-based EH systems exploit the photovoltaic effect to generate electricity from light using photovoltaic cells (PVs). Notice that sunlight can provide an average power density (APD) of 100 mW/cm 2 outdoors (only during the daytime), whereas combined artificial light and indirect sunlight can illuminate indoors with an APD of 100 µW/cm 2 [53]. However, these reference values can change depending on the geographic location, weather conditions, and indoor working hours. Increasing the conversion efficiency of PVs under different ambient light conditions is key for reducing the footprint, increasing the cost-effectiveness, and accelerating the adoption of light-based EH systems. In this front, multijunction PVs, which, as shown in Fig. 4a, are constructed by stacking multiple p-n junctions, become appealing as each layer is optimized for harvesting at different wavelength ranges, thus, increasing total harvested energy [54]. However, multi-junction PVs are challenging and expensive to manufacture, which motivated researchers to propose the replacement of some silicon layers with perovskite PVs. Perovskite-based multi-junction cells facilitate tuning the optical properties of the PVs for improved response in the solar spectrum and lower the manufacturing costs considerably [55]. Unfortunately, perovskite PVs' instability in real-life conditions significantly degrades their conversion efficiency over the long-term (compared to silicon PVs), which makes them (currently) a less favorable option for widespread use [56].
Concentrator photovoltaics (CPVs) are an alternate solution to boost the harvested energy. This system, as shown in Fig. 4b, relies on a built-in light concentrator to focus the incoming light from multiple directions on a very small semiconductor area [57]. This eliminates the need for complex electromechanical tracking mechanisms, which increase the operational expenditure, aka OpEx, and limit the possible use cases due to the resulting heavier design and moving parts. Note that a high concentration of sunlight increases the operating temperature of CPVs, thus decreasing their efficiency. Fortunately, one can leverage the excessive heat for EH using the appropriate transducer (discussed in the next subsection) [58]. In general, the seamless integration of lightbased EH with the corresponding application is required. This is favored by CPVs, which allow the miniaturization of light-based EH systems to the millimeter and sub-millimeter scale with reduced manufacturing costs [59]. Moreover, thin-film PVs improve the integration capabilities by allowing the deployment of light-based EH systems on complexshaped surfaces such as cars' roofs and umbrellas [60]. In this regard, the recent introduction of organic PVs permits printing PVs on practically any surface [61], [62] without impacting their mechanical properties or transparency [63]. Despite their current low efficiency, the abundance of required raw materials, low-cost and scalable manufacturing process, and the possibility of using biodegradable materials, make these cells a green choice for future EH implementations.

B. HEAT-BASED EH
Heat-based EH exploits temperature changes in the surrounding environment [64]. Specifically, thermoelectric generators (TEGs) turn spatial gradients of temperature on the device's surface into electric energy. TEGs work on the principle of the Seebeck effect, due to which a pair of dissimilar materials exposed to different temperatures generate electricity [65]. The basic TEG's architecture is illustrated in Fig. 5a and consists of a set of n-and p-type semiconductors connected electrically in series, and thermally in parallel. To maximize the output power, one side is connected to a heat sink which increases the temperature difference with respect to the side exposed to the heat source. In practice, an array of thin-film TEGs becomes more appealing as one can accommodate more devices per unit volume, thus increasing the harvested energy [66]. Moreover, the miniaturization of the TEGs makes heat-based EH difficult since a very small surface area may not suffice to capture spatial temperature gradients.
Pyroelectric generators are an alternate implementation of heat-based EH that can transform temporal temperature variations into electricity. In principle, pyroelectric materials have an asymmetric crystallographic structure whose charges' disposition changes in response to a temperature variation [67]. Fig. 5b illustrates the basic architecture and operating principle of a pyroelectric generator. When the material is heated, its lattice structure expands creating more space for the charges to move, which weakens the polarization. On the contrary, the entire structure shrinks when the material is cooled, thus reinforcing the lattice asymmetry and increasing the spontaneous polarization of the material. In either case, the temperature variations will cause a flow of electrons between the electrodes to compensate for the new charges' disposition in the crystal.   Heat-based EH technologies are especially useful for waste heat recovery, e.g., in automotive applications [68], for powering wearables (using the temperature changes in the human body) [69], and in industrial facilities [70].

C. MICROBIAL FUEL CELLS
Microbial fuel cells (MFCs) exploit bioelectrochemical conversion to harness the energy resulting from the metabolic processes of microorganisms to generate electricity. MFCs are useful in such scenarios where the devices interact with organic matter, such as compost, wastewater, and ponds [71].
Notice that MFCs require a continuous supply of fresh materials to prevent the microorganism from depriving the available organic matter and hence stopping the electricity generation. To solve this problem, plant MFCs leverage the plants' root exudates and rhizodeposits, resulting from the photosynthesis process, to continuously feed the microorganisms [72], as illustrated in Fig. 6. Plant MFCs can be regarded as a form of converting the sunlight stored as chemical energy in plants into electricity. Factors such as plant species, weather conditions, soil nutrients, and microbial diversity heavily determine the lifetime and output power of this technology [72].
Further development directions focus on boosting the output power and lowering the manufacturing costs of current MFCs. In this regard, the authors in [73] have shown how stacking multiple MFCs boost the output power compared to individual MFCs. Moreover, the authors in [74], [75] provide an extensive discussion on how to build cost-effective MFCs as well as key properties of the components such as high conductivity, large surface area, porosity, durability, and corrosion resistance.

D. VIBRATION-BASED EH
Vibration-based EH (VEH) is the process of converting mechanical energy in the form of vibration (including acoustic vibrations), impact, deformation, or friction into electric energy. As Fig. 7 shows, there are four main types of mechanical-based EH: • Piezoelectric energy harvesters, which exploit the direct piezoelectric effect. The latter states that certain materials, such as quartz, generate electrical charges when subjected to mechanical stress [76]. The output power depends on the relative directions of the electric field and the stress/strain for the direction of polarization of the piezoelectric material. • Triboelectric energy harvesters, which operate thanks to the triboelectric effect. The latter states that certain materials generate electrical charges when they come into contact and separate from another material. Due to their relatively small form factor and high APD, triboelectric energy harvesters are widely used in wearables to harness energy from human motion [49]. • Electrostatic energy harvesters, which are variable capacitor structures. Specifically, their capacitance changes when the plates' overlap area or the gap distance varies in response to an external force. The basic operating principle consists of exciting the capacitor with an external power supply when the capacitance is maximum. Then, the energy is harvested when the potential energy stored in the capacitor increases in response to a smaller capacitance value [77]. Due to the high polarization voltage required in this method, some electrostatic energy harvesters incorporate electrets, which are quasi-permanent electric dipoles that can hold electrical charges for years. • Vibration EM energy harvesters, which exploit Faraday's law of induction. Such a law states that the relative movement between a wire coil and a permanent magnet produces an electromotive force.
VEH presents a significant challenge: the resonant frequency of the EH device must match the input environmental frequency to maximize the output power [50]. However, ambient energy frequency can vary considerably, necessitating the adjustment of the resonant frequency to ensure a reliable energy supply. In the literature, researchers have explored various approaches to achieve this, including manual mechanical tuning [78], mechanical self-tuning [79], and electronic self-tuning [80]. Another solution is to design the device to resonate at multiple frequencies, thereby expanding its frequency response spectrum. To achieve this, hybrid energy harvesters [51] or arrays of multiple harvesters [81] can be used, with each device resonating at a different frequency.
Another challenge of VEH is the relatively low frequency of ambient vibrations with respect to the resonant frequency of the devices. To address this issue, frequency up-conversion mechanisms have been proposed. For instance, by applying a strong external force for a brief period, e.g., using a freemoving mass, the system may vibrate at its natural resonant frequency [82]. Alternatively, one can lower the harvester's natural resonant frequency by reducing the stiffness of its moving parts. This can be achieved, for instance, by replacing them with fluids [52].
Finally, ambient VEH is often the result of multiple external forces acting in different directions. Thus, allowing devices to harvest energy from multiple directions, e.g., using two-dimensional [83] and three-dimensional [84] EH devices, can boost the total output power.

E. FLOW-BASED EH
Flow-based EH leverages the kinetic energy of naturally or artificially originated fluid flows, such as water streams and wind currents, to generate electricity. Notice that in most use cases, the underlying physical principles of VEH are also applicable to flow-based EH. Therefore, this subsection focuses on wind-based EH and, in particular, on their form factor, blade design, and energy-efficient design for medium-to large-scale electricity generation.
Traditional wind turbines leverage the kinetic energy of the wind to turn the rotor of an electric generator. Wind turbines are usually deployed in farms for large-scale electricity generation. However, massive IoT deployments are more likely to appear in cities where installing a huge wind turbine is not always feasible. Thus, research and industry efforts have been focused on building compact and cost-effective designs of wind-based EH systems. A significant issue with traditional wind turbine design is the potential threat to wildlife posed by the rotating blades. To address this, Halcium has created the PowerPods 5 , which are portable wind turbines designed for residential areas. The blades of the PowerPods are contained within the pod, as shown in Fig. 8a, making them safe for people, pets, and wildlife in close proximity. The shell of the PowerPods captures wind from multiple directions and channels it through small exits, increasing the wind speed and therefore the harvested energy. Different from traditional wind turbines, PowerPods utilize vertical-axis blades whose low aerodynamic noise pollution and ability to operate under unstable wind flow (without needing a yawing mechanism) 5 For more information please check https://www.halcium.com/ make them appealing for urban environments [85]. Vertical-axis turbines are also the basis of the socalled Savonius turbines (shown in Fig. 8b) which are reliable and cost-effective systems that can operate under turbulent wind flows and stormy weather [86]. Notice that traditional wind turbines require a braking mechanism to operate under such conditions as high wind speeds can cause the system to shatter becoming a hazard for the nearby areas.
Another area of development in wind-based EH involves blade-less turbines. An example of this technology is the Vortex Bladeless 6 , shown in Fig. 8c, which leverages the vortex shedding phenomenon [87]. Vortex shedding occurs when a fluid flow separates from a solid surface and creates vortices. In the case of the Vortex Bladeless, the vortex shedding effect causes a vertical mast to oscillate in resonance with the wind flow. One advantage of this design is the reduced maintenance costs, as there are no moving parts that can wear out through friction. However, the system can experience mechanical fatigue and stress at the base of the oscillating mast. Additionally, the smaller footprint of the Vortex Bladeless compared to traditional wind turbines allows for the installation of multiple units in the same area, potentially compensating for their individual lower performance.

F. RF-EH
RF-EH technology utilizes a rectenna to convert EM waves into direct-current (DC) power, which can be harnessed to power and/or charge small electronic devices and batteries. The rectenna is composed of a receiving antenna and a rectifier plus a matching network to maximize the power transfer efficiency (PTE). The main sources of ambient RF energy are classified as [88]: i) static, which corresponds to energy transmissions that remain relatively stable over time, e.g., from television and radio transmitters, allowing long-term predictability of the energy supply, thus favoring the network planning, and ii) dynamic, which supplies timevarying energy over a certain region, either because of fluctuating transmit power levels or mobility, e.g., from WiFi/mobile access points or devices, thus should be adaptive and allow operation over a wide range of frequencies.
The main challenge of ambient RF-EH circuits lies in how to guarantee an appropriate performance given mostly unpredictable EM conditions, in which the receive signal parameters, e.g., carrier frequency, bandwidth, polarization, and antenna orientation determine the amount of harvested energy. Hence, key desirable features of RF-EH circuits include performing EH i) from multiple bands, which can be accomplished by using wideband, multi-band, or tunable receivers [21], ii) from multiple spatial directions, for which omnidirectional antennas are preferred, iii) from randomly polarized signals, which is mostly implemented by some sort of circularly polarized antenna; and iv) from a wide range of input power levels, in which case the RF-EH adjust the sensitivity and saturation levels to obtain the most from the ambient energy source. Figs. 9a-c illustrate the basic architecture of some of the aforementioned receivers.
Multi-antenna RF-EH circuits, as shown in Fig. 9a, Fig. 9d, and Fig. 9e, are another alternative to boost the amount of harvested energy. In such case, the incoming signals can be combined in i) the RF domain; ii) in the DC domain, or iii) hybridly in both domains [89]. Multi-antenna rectennas also improve the spatial selectivity of the antenna in the direction of the maximum incident signal, if properly adjusted. However, for ambient RF-EH the optimal receive beamforming is difficult to attain due to the non-dedicated nature of the transmissions. To cope with this problem, several low-complexity solutions, including codebook-based beamforming, have been proposed in [89] such that an ambient RF-EH device sweeps a phase shift table in an initial phase seeking the configuration that yields the best performance for exploitation in a second phase.
Magnetostrictive antennas are an alternative circuit implementation to rectennas for RF-EH. In such a case, the transducer is built with multiferroic structures with the capacity of sensing the magnetic component of the EM waves and, in response, produce vibrations. Then, a mechanically connected piezoelectric EH converts the vibrations into electrical energy. This allows realizing an ultracompact receiver, which is up to two orders of magnitude smaller than state-of-the-art rectennas [90]. Metamaterials-aided designs are another solution for realizing ultra-compact RF-EH implementations, due to the sub-wavelength periodicity of the unit cells within the lattice structure. Metasurfaces-aided RF-EH boosts the conversion efficiency as it allows wider beamwidths, polarization-independent operation, built-in and less lossy matching networks, and higher antenna gains [91].

III. ENERGY TRANSFER (ET)
WET refers to the intentional transmission of energy using dedicated transmitters to power EH devices and even other transmitters. In contrast to ambient EH, WET provides a controllable and predictable energy supply and the tools for increasing the end-to-end power conversion efficiency. In Table IV, we compare the current development state of the WET technologies discussed in this section. In the following subsections, we cover popular implementations of WET and outlook wireless energy trading in the context of microgrids.

A. LASER-BASED WET
The most common implementations of laserbased WET are laser power beaming and distributed resonant beam charging. In the former case, as shown in Fig. 10a, a laser diode at the transmitter steers an optical beam towards a PV cell at the receiver, thus using a similar working principle as in conventional solar power systems. Meanwhile, in the latter, the receiver bounces back a portion of the incident energy towards a gain medium at the transmitter, which amplifies the light and initiates a resonant beam [92], as illustrated in Fig. 10b.
Although the above two implementations are regarded as long-range WET technologies, laser power beaming provides superior coverage in the order of several kilometers [93]. Some authors consider it as key for intra-satellite WET applications [94], charging UAVs [95], and energy transmission from solar-powered satellites to terrestrial stations [96]. However, the performance of the laser power beaming heavily depends on the line-of-sight (LOS) and the atmospheric conditions. Besides, laser power beaming raises safety concerns due to the potential harm to the living species' tissue. That is why some researchers have proposed lowering the transmission frequency to provide a hazardousfree system at the cost of high maintenance cost and lower efficiency (and therefore trading efficiency for safety) [93], whereas others have proposed monitoring the periphery of the beam path for intruders [97].
Differently, resonant beam charging is inherently safe as any interruption of the link causes the resonance to stop immediately. In addition, the system is capable of supporting mobile users without the need for a tracking system as long as LOS conditions hold, thanks to its self-aligning capabilities which have been experimentally validated (see [98]  simultaneous resonating beams also enables native multi-UE support, as reported in [99]. However, broadcasting energy to multiple receivers is typically more inefficient, often motivating time division multiple charging protocols [100]. To overcome the LOS limitations, UAVs [92] or supporting reflectors [101] can relay energy transmissions to charge those devices out of the coverage of the main transmitter.

B. ACOUSTIC WET
This technology leverages acoustic waves to charge EH devices by typically equipping transmitters and receivers with piezoelectric transducers [102]. Acoustic WET works in any medium capable of propagating pressure waves, e.g., metal, air, and human tissue, but it is particularly convenient in EM wave absorption-prone mediums, such as Faraday shielding structures [103] and water. Besides, for the same operating frequency, acoustic WET transmitters/receivers have a more compact form factor and achieve a higher directivity than those implementations based on EM WET [104].
Similar to other WET technologies, the channel attenuation, which varies significantly with the spe- cific medium, affects severely the system performance. Notably, one cannot increase deliberately the intensity of the acoustic waves since, depending on the operating frequencies, it may cause hearing impairments, body heating, and other unpleasant effects in humans and animals [105]. Notice that the attenuation increases with the impedance mismatch of the medium between the transmitter and the receiver since traveling waves may encounter different materials in their path. This has motivated the use of glue or electromagnets to ensure a firm connection of both the transmitter and receiver to the matching layers in the path (when solid) and therefore reduce the propagation losses [106]. Moreover, when receivers are deployed inside a Faraday shielding structure but not connected to it, one can rely on hybrid WET approaches in which the last section of the transmission path relies on a different WET technology, such as inductive coupling [107].
Phased acoustic arrays, illustrated in Fig. 11, which are composed of multiple piezoelectric transducers, also aid in overcoming medium attenuation by focusing the energy toward the receiver direction using sound beams. In fact, transmit phased acoustic arrays outperform single transducers of the same size, with an increasing performance gap as the number of transducers increases [104], [108]. Further improvements can be achieved by also equipping the receiver with phased acoustic arrays and hence allowing different combining techniques to improve the transfer efficiency, similar to a traditional MIMO-like wireless link [109]. Notice that the achievable performance of a phased acoustic array is also a function of the transducers' geometry and arrangement, and the array's aperture and diameter [110].
Finally, acoustic WET can also provide heterogeneous quality of service (QoS) to the end users. For instance, acoustic transmitters can use Lamb waves to create a pattern of peaks and valleys in solid structures. Therefore, by changing the operating frequency, the resulting vibrations charge the devices deployed in different locations of the same structure with a different intensity [111]. Besides, adaptive acoustic beamforming also serves to discriminate which devices to charge depending on their locations and energy demands [112]. This might be especially useful in sensor networks embedded in structures such as buildings or bridges.

C. INDUCTIVE COUPLING-BASED WET
This technology exploits the inductive coupling phenomenon. The simplest setup consists of two wire coils (transmitter and receiver) coupled by a magnetic field such that the oscillating magnetic field in the transmitter's coil passes through the receiver's coil inducing an alternating current. This forms a highly efficient air-gap transformer whose performance depends on the operating frequency and the mutual inductance between both coils. However, this basic setup performs poorly when both coils are misaligned or too separated.
One can extend the charging coverage by tuning both coils to resonate at the same frequency. This is commercially known as magnetic resonant coupling, which compared to the basic inductive coupling, is more resilient to coils' misalignment. The most common magnetic resonant coupling methods include an external capacitor to compensate for the internal inductive reactance and extra coils for tuning and impedance matching. The first implementation is easier to realize, but the second one achieves a higher conversion efficiency given that there are no power losses in external resonators [113].
Magnetic resonant coupling allows a single transmitter to charge multiple receivers simultaneously. Here, the internal resistance of the resonators can be adjusted to control the mutual inductance among coils and hence boost the network performance. For instance, increasing the internal resistance of the nearest receivers may increase the harvested energy at the most distant receivers; however, at the cost of a higher transmit power [114].
Deploying multiple transmitters also improves the system's performance. On the one hand, the optimal deployment of the transmitters guarantees uniform power coverage and ensures a minimum available energy at the receivers regardless of their locations [115]. On the other hand, coordinated transmissions from multiple transmitters can result in a constructive combination of the magnetic fields at the receivers' locations, which is known as distributed magnetic beamforming [116].

D. CAPACITIVE COUPLING-BASED WET
This technology exploits the capacitance coupling phenomenon. The basic architecture consists of two pairs of transmit and receive plates each coupled by an electric field. One pair forwards the displacement current while the other provides the return path to close the circuit. Capacitive coupling systems come in two different flavors depending on how the plates are arranged. The horizontal capacitive coupler has the transmitter's (and receiver's) plates placed side by side in the same plane, while the vertical capacitive coupler has both the forward and return paths overlapped in the direction of the electric field. Although the former architecture has the highest coupling, it comes at the cost of being bulky and less reliable due to the large number of required components. For a fixed device size, one can enhance the coupling of the system by combining the geometry of the basic architectures and adding more plates [117].
Some critical challenges in capacitive coupling systems include the high excitation frequency/voltage due to the large capacitive reactance of the coupling and the control of the fringing electric fields, i.e., the non-uniform field at the edge of the plates, within safe limits. To cope with these issues, resonant matching networks can be added at both the transmit and receive sides. They compensate for the reactive losses in the system, hence reducing the required excitation voltage. Further, they can be designed as a transmit voltage gain stage to drop the displacement current of the plates, hence reducing the fringing field to the safety limits. Herein, the size of the inductors in the matching network is key to realizing compact designs. As an example, one can exploit the successive impedance transformation that provides multistage matching networks [118]. The required amount of gain and compensation of such a matching network depends on the ratio between the load and excitation currents, and the misalignment/distance between couplers [119].
Different from inductive coupling, capacitive coupling systems are more tolerant to the plates' misalignment, reduce the risk of interference to neighboring networks, and do not induce eddy currents in nearby metallic objects. Moreover, they are lighter, easy to integrate, and more mechanically robust. That is why they are regarded as promising for powering medical implants, vehicles, consumer electronics, and rotary electric machines [120], [121].

E. RF-WET
RF-WET technology relies on intentional RF transmissions to charge RF-EH devices. The main challenge of this technology is its limited coverage due to the channel attenuation and the regulations on the maximum radiating power of the RF energy transmitters, hereinafter referred to as power beacons (PBs). For this reason, directive antennas are preferred to increase the incident RF power in the desired direction without increasing the transmitter's radiating power. Notice that when serving multiple users, the PB can utilize an omnidirectional antenna to provide basic service guarantees and a directional antenna (for out-band) energy transmissions to meet more specific user equipment (UE) requirements [122]. However, in case the PB is equipped with non-reconfigurable antenna(s), a mechanical sweeping of the service area may be needed to charge those devices otherwise located in the minimum of the radiation pattern [123].
Another potential technique to extend RF-WET coverage is channel state information (CSI)-based energy beamforming. The main challenge here resides in the often unavoidable cost of instantaneous CSI acquisition, especially when powering a massive number of devices with strict HW constraints, which may null or even reverse the gains from accurate CSI-based transmit strategies. This has motivated alternative strategies relying on statistical CSI [124], received energy feedback [125], and devices' position information [123]. Conveniently deploying the PBs is also key to overcoming channel attenuation, banning blind spots, and homogenizing the incident RF power according to the network requirements [22], [126], [127]. Notably, distributed single-antenna PBs offer better service than a single PB equipped with the same total number of antennas due to the reduced charging distance [127] and the reduced number of EH devices to be charged per PB [22]. Robotic WET, in which the PBs move [128] or fly [129], becomes also appealing not only to reduce the link distance but also to power the devices deployed in hard-to-reach places or to meet temporary service requirements during emergencies. Other potential technologies to boost the coverage of RF-WET are also discussed in Sections IV-IV-B, IV-C.
The next generation of wireless systems envisions extremely large antenna arrays to compensate for the attenuation at high-frequency bands, increase reliability, and reduce interference. Consequently, many future IoT deployments may operate in the radiating near-field (or Fresnel region) of the transmit antennas, which contrasts with the traditional assumption of far-field operation. As Fig. 12 illustrates, under such operating conditions, a PB can focus the energy in a particular location as opposed to what happens in far-field conditions, where the energy is steered towards a certain angular direction. As a consequence, near-field RF-WET will generate less RF pollution, thus interference, in both angle and distance domains and a reduced human RF exposure [130]. Moreover, the conversion efficiency achievable in the near-field can be significantly high even when the energy receiver is not in the focal point of the transmitter's antenna [131].
Given all the above facts, RF-WET is usually regarded as a short-distance solution to charge lowpower devices. However, some experiments have shown energy transmissions over distances greater than 1 km distance with a peak incident power of several Watts at the rectenna [132]. Long-distance RF-WET could also help bypass the complicated infrastructure of the electric network to power hardto-reach locations or to provide an energy supply during emergency situations. 7

F. ENERGY TRADING & MICROGRIDS
The idea of energy transmitters relying entirely on the power grid for operation contradicts the vision of sustainable networks since most of the massgenerated electricity comes from burning fossil fuels. Hence, the exploitation of renewable sources to power the self-sufficient energy transmitters becomes an appealing solution on the road toward sustainable networks [133]. However, a key challenge lies in the unpredictable relationship between electricity generation and demand, which calls for energy trading mechanisms.
The above vision aligns with the emergence of microgrids, which are becoming increasingly popular to provide reliable and sustainable electricity in remote areas. A microgrid is a local energy network that comprises renewable electricity generators (as those discussed in Section II), energy storage, fuelbased generators, and controllers/inverters. Notice that fuel-based generators can be used as a backup source of electricity when renewable sources are not enough to meet the energy demand while the controllers/inverters ensure that energy flows efficiently between the various microgrid components.
While energy trading with the main grid serves as a backup for renewable-powered WET systems [134], the main goal is to achieve a cost-effective and reliable operation of the network using as little energy as possible from fuel-based generators. This is not only more environmentally friendly but also reduces energy production and distribution costs. Moreover, direct energy trading among energy transmitters can also be possible using wired and wireless links that bypass the hierarchy of the power grid distribution network. In the case of wireless energy trading, static energy transmitters can exchange energy using, for instance, RF-WET and laser power beaming, to transfer energy over large distances. Moving/flying energy transmitters can benefit from these but also from near-field WET technologies, such as inductive coupling and acoustic WET, which provide high PTE, to trade energy with static transmitters. The idea here is that those transmitters cannot be equipped with large solar panels or wind turbines, so it is more beneficial for them to trade with other ground infrastructures with superior ambient EH capabilities [16], [135]. Finally, energy transmitters can trade their surplus energy using the grid distribution network [136]. This scenario becomes useful in the case when there is no direct connection between the parties, e.g., when the seller and the buyer transmitters are operated by different companies or deployed so distant from each other that a direct connection becomes infeasible. Some tools to model/study the uncertainty of energy availability and demand in heterogeneous IoT networks include game theory [137], [138], reinforcement learning [139], and others surveyed in [15]. Notably, scalable algorithms are needed when the number of trading energy transmitters increases [140]. Also, transmitters participating in the trading may be managed by different operators, which poses a challenge in terms of privacy and security. This motivates the use of distributed ledger technology, which can provide transparency to the transactions and protect the ledger against possible forging by energy transmitters/receivers behaving maliciously [141]. In other situations, the operators may not want to disclose their generation capacity and demand, hence potentially limiting the information exchange during the trading process [140].

IV. ENERGY EFFICIENT COMMUNICATION TECHNOLOGIES
Once the energy is available for operation through EP/ET mechanisms as those outlined in Sections II and III, the device(s) and/or the network(s) must ensure its efficient usage. Indeed, every technique and technology conceived for performance improvements in terms of coverage, throughput (THP), dependability, and other KPIs, can be in principle leveraged for EC reduction as well, hence, EE. That is the case, for instance, of massive MIMO [142], [143], UAV-based connectivity [144]- [146], satellite-assisted communications [147], [148], cooperation and diversity mechanisms [149]- [151]. Nevertheless, there are some technologies/techniques that are natively conceived for energy-limited/efficient operation, and these constitute the scope of this section.
Notice that the primary sources of EC in a device or network are i) the utilization of active circuit components such as transistors and power amplifiers, which require a power source to function, and ii) the running applications. Therefore, to enable energy-efficient and/or low-power operation, the use of active techniques and devices must be limited when possible while relying on efficient passive architectures. This comes with several challenges as active devices, although consume energy, can facilitate high-performance computing and communication, while passive devices (e.g., resistors, capacitors, inductors) are much less flexible and may dissipate energy in the form of heat as they interact with electric signals. This motivates the use of semi-passive/active architectures, which include some active techniques and components to efficiently support the application demands for which they are deployed, and/or tunable operational models as enabled by the WuR technology. These approaches are discussed next through key technologies. Specifically, backscatter communication (BC), metasurface-aided communication, radio stripes, and WuR technologies are overviewed in Sections IV-IV-A, IV-B, IV-C, and IV-D, respectively. Meanwhile, the ML approach, focused on intelligently reducing the EC burden of computation/communication application tasks at the devices and networks, is discussed in Section V. 8

A. BACKSCATTER COMMUNICATIONS
BCs are passive and involve backscatter tag(s) reflecting the signal from nearby transmitter(s) and modulating it by adjusting its amplitude, frequency, and/or phase via impedance mismatching tuning. At the receiver, the backscattered signal is processed to extract the information added by the backscatter tag. The typical components and types of BC systems are illustrated in Fig. 13. Notice that the transmit- ter(s) may be dedicated or non-dedicated, the latter leading to the so-called ambient BC systems. There are multiple implementation options for powering the backscatter tags, e.g., relying on EH from ambient/dedicated RF signals [155]- [157], solar energy, vibration, and/or thermal gradient [156], [158]. In general, the energy buffer can be as small as a capacitor of several nanofarads, or as large as a digitally-controlled supercapacitor or rechargeable microbattery. The nature of the energy buffer depends on the application. A capacitor might suffice for ultra-low power applications, in which case it is not often regarded as a buffer but simply as another stage element within the circuitry. In other cases, a large and intelligently controlled energy buffer is required to enable substantial gains in range and operating time that could not have been attained otherwise. Such energy storage topology influences also the type of modulation/demodulation of the backscatter signals, which can be digital or analog as in conventional communication systems, and may be activated by a wake-up code detector [159]. Table V lists representative state-of-the-art BC systems along with their distinctive features and performance figures.
1) Key BC Topologies: As shown in Fig. 13, there are two key BC topologies: monostatic and bistatic. In traditional or monostatic backscatter systems, the transmitter and receiver are integrated, e.g., in RFID, while they are separate in bistatic systems. 9 The latter implementation offers [180]: 9 Note that the term "backscatter" may not be accurate for the bistatic architecture since the signal does not necessarily scatters back, but toward the receiver. • Temporal flexibility: The tag has additional time slots to transmit data rather than being restricted to waiting for a single reader's protocol-bound inquiry. This allows the backscatter tag (functioning as a sensor node) to transmit the sensory data immediately upon availability, which might be crucial in many sensing scenarios. • Spatial flexibility: The position of the receivers can be optimized. Moreover, the spatial deployment of backscatter tags can be quite flexible in urban and metropolitan areas with high ambient RF APD, while in other scenarios, dedicated transmitter(s) can be strategically placed in optimal locations to balance the scalability and performance of backscatter tags. • Technology flexibility: A variety of excitation signals can be exploited since the transmitters can be available ambient RF sources, e.g., TV or frequency modulation (FM) radio towers, cellular base stations (BSs), and Wi-Fi access points, from anywhere, while several modulation schemes may be supported. Due to the above, bistatic BC technology has become increasingly popular in recent times. Notice that the transmitter(s) and receiver(s) may not be completely independent entities in bistatic systems, but they may cooperate. For instance, the authors in [171] propose for them to share link settings and metrics to improve the BC link performance. Nevertheless, cooperation is not possible in the case of ambient BC systems.
2) Use cases: The applications of BC technology are numerous as illustrated in Fig. 14   particularly useful in scenarios where sensors must be instrumented in an invasive manner because it can potentially eliminate the need for replacing batteries and the accompanying extra high cost. That is the case of applications such as i) structural monitoring, e.g., to ensure safe and reliable operation of railways, pipelines, dams, bridges, and aircraft, and ii) implantable health-care monitoring [180]. • Ubiquitous localization and sensing: The astonishing spatial diversity offered by the massive proliferation/deployment of backscatter (battery-free) devices (e.g., attached to everyday objects) can be exploited to extend networks' localization and sensing capabilities as envisioned by the 6G paradigm [181]. • Smart world: By incorporating backscatter tags into everyday objects, not only network sensing capabilities are naturally extended, but also each of these objects is digitized and becomes a source of information and/or control unit with added value. An appropriate choice between (non-passive) radios and BC for an application depends on the operating distance, data rate requirement, and power budget [182]. There is no one-size-fits-all solution. In general, BC is more energy-efficient but may not cope with stringent requirements in terms of coverage and rate. Indeed, BLE radio is likely the best option currently to support 1 Mbps at 20-40 m, while BC fits better for applications requiring around 100 kbps at 5-10 m. However, this also a matter of deployment topology, cost, and other KPIs such as reliability, latency, and security as required by the target application.
3) Challenges and Research Directions: Many state-of-the-art BC designs as those implementing LoRa and BLE backscatter in Table V are quite mature and deployable. However, there are still a number of challenges that require attention in the coming years to fully realize the potential of BC technology and enable its seamless integration with 5G and beyond generations, thus paving the way for the most advanced and futuristic applications. A compilation of such key challenges and associated research directions is presented below.
• Imperceptible integration to everyday objects: Fabricating BC miniaturized electronic components and integrating them into small, and diverse, form factors is challenging, and so is the imperceptible integration into everyday objects. In the near future, the exploitation of meta-materials with favorable electrical properties (see Section IV-IV-B) may be appealing to achieve small computational materials that can communicate using backscatter. • High-frequency operation: Most current BC systems operate at well-established microwave frequencies. However, expanding the operation to higher frequencies (e.g., mm-wave and THz) may bring substantial benefits since more antennas per unit area can be packed, which allows producing very directional transmit, receive, and backscattered beams while achieving longer communication ranges. Unfortunately, ambient RF energy availability is more limited in such a high-frequency operation regime, which may prevent exploiting ambient BC. Instead, dedicated RF signals/sources are required together with novel low-complex protocols that efficiently handle the issues related to beam search, especially for flexible bistatic BC systems 10 . Notice that current high-frequency BC prototypes are monostatic, e.g., [183]- [186], or non-flexible bistatic, e.g., [187], thus, much further research is needed towards realizing flexible high-frequency BC systems suited to real-life applications. • Wide-band and frequency-agnostic designs: Wide (ultra-wide)-band BC designs (including advanced modulation schemes) are needed for supporting high data rates applications, e.g., sensing and interaction related to augmented reality, which constitutes a challenging research direction. In addition, frequencyagnostic BC systems capable of operating across different protocols and frequency bands with the potential to operate universally across locations, countries, and applications, would be extremely appealing. In this regard, HW innovations, as well as low-power algorithms, are required to dynamically identify which frequency bands have the strongest signal. • Security: Due to the limited power and complexity of BC systems, guaranteeing secure communications is extremely difficult. Notice that i) the accurate identification of a fake ambient BC tag, and ii) the mitigation of interference generated by a tag maliciously backscattering ambient signals to a nearby reader, remain open problems in the literature. A promising direction lies in designing quantum backscattering mechanisms [188]. • Enhanced RF-EH sensitivity: RF EH, as the charging source, facilitates small form-factor battery-free backscattering implementations. Unfortunately, the RF-EH sensitivity is orders of magnitude worse than that of a BC receiver, which limits the connectivity range. Improving the RF-EH sensitivity, preferably up to two orders of magnitude, is a fundamental research and engineering challenge. 11 Techniques exploiting leakage power reduction, technology scaling, and sub-threshold operating using voltage scaling may be key to achieving this goal. • BC networks: Most of BC research and prototyping is focused on the physical layer with piecemeal evaluation. A full-layer design constitutes a challenging next step for maturing the technology and realizing scalable, integrated, and practical BC networks with the capabilities for realizing, e.g., carrier sense, network management, and polling of devices.

B. METASURFACE-AIDED COMMUNICATIONS
Metasurfaces are surfaces composed of metamaterials with sub-wavelength thickness. The so-called metamaterials have special properties when interacting with EM radiation, and thus metasurfaces may support several functions as shown in Fig. 15 [189], [190]: • Reflection/refraction of the incident RF waves to a given reflecting/refracting direction; • Absorption, by which the reflected/refracted signals corresponding to an incident RF wave are nulled; • Beamforming, by which the incident RF waves are focused toward a given direction/location. Collimation is the complementary operation; • Polarization change of the reflected RF waves with respect to the incident ones. For example, 11 Under free-space propagation conditions, every 6 dB improvement in sensitivity roughly doubles the operating distance [182]. incident RF waves are transverse electric polarized, and reflected RF waves are transverse magnetic polarized; • Splitting, by which multiple reflected or refracted RF waves are created from the incident RF waves; • Analog processing includes mathematical operations at the EM level, e.g., the RF waves refracted by a metasurface can be the first-order derivative or the integral of incident RF waves. These functions have motivated the interest in metasurfaces for improving the performance of various wireless systems, e.g., for passive signal or energy relaying [191] and wireless sensing [192].
Notice that metasurfaces are low-cost fullypassive devices with zero-EC, which can be engineered to statically perform one or several of the above functions. However, more recently, the research community and industries are pursuing a more dynamic approach, where the operation of the metasurfaces is SW-controlled in real-time, thus, leading to intelligent metasurfaces. In the following, we briefly discuss the main configurations of intelligent metasurfaces in the state of the art: i.e., intelligent reflective surface (IRS) and large intelligent surface (LIS), together with associated research challenges and relevant research directions. A list of representative prototypes and experiments with metasurfaces in the last years is presented in Table VI along with their main distinctive features.
1) IRS: As illustrated in Fig. 16, IRS (also known as reconfigurable intelligent surface -RIS) is a metasurface composed of a large number of N subwavelength-spaced passive scattering metamaterial elements. Each scattering element can be controlled in an SW-defined manner by the socalled IRS controller, which can be embedded or separated, to properly tune the EM properties of the output signals given a set of incident RF signals impinging the scattering elements. Therefore, IRS is a promising technology to dynamically control the radio propagation environment and improve the wireless system performance in a cost/energyeffective manner. Specifically, an IRS can reflect the incident signals so they are added constructively in the desired direction to increase the signal power (so-called passive beamforming) or destructively for mitigating undesired interference, either for high data rate, dependable, secure, NLOS, or widecoverage communications, and RF-WET (including joint communication and RF-WET). Conventional IRSs, so-called passive IRSs, e.g., [193]- [197], [200]- [203], [205]- [211], are com-posed of fully passive reflecting elements and the only active power consumption comes from the controlling HW (IRS controller and related active circuitry). This allows low-power/cost energyefficient implementations that do not incorporate additional RF radiation into the environment, which is undoubtedly appealing in terms of sustainability.
Unfortunately, the effective coverage of passive IRS may be seriously limited since the reflected signal suffers high product-distance path loss, which critically constrains the signal power at the receiver(s). This can be addressed by either equipping the IRS with an increasingly massive number of passive elements and/or conveniently deploying the IRS closer to the transmitter/receiver to reduce the cascaded channel path-loss. Unfortunately, these options are not always viable due to, e.g., deployment difficulty (limited space and unavailable site) and enormous training overhead for the CSI acquisition required to optimize the operation of the passive IRS elements in real-time. This has motivated the introduction of active IRS [212], [213].
In active IRS, the reflecting elements incorporate reflection-type amplifiers, which are implemented without high-cost and power-hungry RF chains and that can simultaneously alter the signal's phase and amplify its amplitude to enhance the signal power at the receiver. This requires modestly higher power consumption and HW cost compared to conventional IRS.
Interestingly, the achievable signal-to-noise ratio scales as O(N ) and O(N 2 ) when using active and conventional IRS, respectively [213]. The limited gain in the case of active IRSs is due to the amplification of the noise power introduced by the active design. Nevertheless, an active IRS implementation still provides superior rate performance (subject to a given total power budget constraint) in comparison to a conventional IRS when N is moderate thanks to the additional signal power amplification gain [213]. Table VII illustrates the main features of conventional and active IRSs with competing relaying technologies. All in all, there is no onesize-fits-all solution (as usually in engineering), and the use of one technology over another depends on the specific network's characteristics, constraints, and performance requirements.
2) LIS: LIS is a metasurface equipped with RF circuits and signal processing units and composed of a virtually infinite number of elements to form a spatially continuous transceiver aperture. This structure can be used to transmit/receive communication signals across the entire surface by leveraging the hologram principle [214]. 12 For instance, a LIS architecture may be comprised of multiple waveguides, e.g. microstrip. Each waveguide may embed a large set of radiating metamaterial elements whose frequency response can be externally and individually adjusted by varying the local electrical properties. Each microstrip is fed by one RF chain, and the input signal is radiated by all the elements located on the microstrip, as shown in Fig. 17 [216]. The figure also illustrates an example of transmitting a signal using a single microstrip with multiple elements.
3) Challenges and Research Directions: Some key challenges and associated research directions for maturing metasurface-assisted communication technology and making it a reality in future sustainable networks are briefly discussed in the following: • Low-cost/energy control: The use of tunable reflecting elements with discrete amplitude/phase shift levels favors cost/energy-effective implementations. However, this may significantly limit the beamforming/reflection capabilities of the metasurface, especially in the case of limited form-factor implementations. Therefore, further studies are required to unveil such underlying trade-offs. Moreover, further research is required on the metasurface controller circuitry, which interfaces with all the tunable reflecting elements and constitutes the only EC source in passive IRS implementations. Specifically, low-cost/energy-effective control mechanisms must be developed to connect and communicate with massive tunable elements, and thus agilely and jointly control their EM behaviors on demand. • CSI acquisition: A tunable passive beamforming/reflection requires accurate CSI for optimum performance, which is challenging to acquire in practice. The two main approaches proposed in the literature for passive IRS (without any active RF chain or reflection-type amplifier), but that still require further research in terms of performance trade-offs, are [191]: i) estimate the concatenated (TX-IRS and IRS-RX) channel with some known IRS reflection patterns, or ii) exploit feedback from the TX/RX pertaining to their received signals that are reflected by the IRS (no explicit channel estimation). Meanwhile, CSI acquisition for active IRS is generally more challenging mainly because explicit CSI of the separate TX-IRS and IRS-RX links, instead of the cascade CSI, is needed due to the amplification noise. The research to address this issue is still incipient, with only a few attempts, e.g., [213]. In any case, approaches exploiting limited CSI as those relying, e.g., on positioning information [217], may be often appealing. • Advanced metasurface implementations: The most commonly investigated/prototyped metasurface applications are those related to passive beamforming/reflection. However, as illustrated in Fig. 15, there are several other metamaterial functions, which are so far only incorporated into non-flexible/configurable implementations. In the next years, SW-controlled engineering solutions must be developed exploiting these metamaterial functions for real-life applications. For this, accurate physics-and EMcompliant models are needed.
• Data-driven optimization: Accurate modeling/ analysis and optimum design/implementation/ deployment of IRSs/LISs are challenging due to the inherent complexity of such systems. This calls for efficient data-driven methods, e.g., based on deep learning, reinforcement learning, and transfer learning [218]. Moreover, low-complexity/energy ML mechanisms, the so-called TinyML (see Section V-V-C), seem appealing for incorporation into the metasurface controller circuitry for more cost/energyeffective implementations.

C. RADIO STRIPES
Massive MIMO technology was introduced to address the challenging 5G performance requirements, especially in terms of network coverage, capacity, and THP. However, the required high manufacturing and operating expenses, as well as the increased power consumption, make the development and deployment of truly large-scale antenna arrays extremely difficult. This motivates the research on more affordable and low-power MIMO architectures that can scale with the number of antennas more sustainably. The metasurface-aided communication architectures overviewed in Section IV-IV-B constitute one active research front in this direction. Indeed, IRSs can be deployed to efficiently assist already-deployed MIMO networks (thus, avoiding the need to install new active HW); while LIS are basically low-power/cost massive MIMO that can replace traditional MIMO HW in many scenarios. However, the deployment of IRS/LIS alone may not fulfill all the use cases' constraints/requirements, at least in a near future, for which alternative lowpower/cost MIMO architectures are still needed. That is the case of the radio stripes system discussed in the following.
Radio stripe systems enable cost/energy-effective distributed massive MIMO implementations [219]. In such a system, antenna elements and circuitmounted chips (including power amplifiers, phase shifters, filters, modulators, and A/D and D/A converters) are serially located inside the protective casing of a cable or a stripe, which also provides synchronization, data transfer, and power supply via a shared bus. Fig. 18 illustrates the structure of a radio stripe system.
Unlike traditional massive MIMO BSs, radio stripes [22], [219]: i) allow imperceptible and flexible installation in existing construction elements and alleviate the problem of deployment permissions, ii) support native system resiliency to failures, and iii) facilitate low power consumption due to the inherent low-complex and distributed architecture functionality. Moreover, additional HW, such as temperature/vibration sensors and microphones/speakers, can be deployed in the radio stripes to provide additional features/services, e.g., fire/burglar alarms, earthquake warnings, indoor positioning, and climate monitoring/control. Due to the above, radio stripes technology is undoubtedly attractive for supporting energy-efficient (and sustainable) networks. Applications are numerous, e.g., to: i) facilitate high spatial multiplexing and low EC in indoor communications [220], ii) increase the coverage and end-to-end PTE of RF-WET [22], [221], and iii) support ultra-reliable lowlatency communications (URLLC) in industrial IoT networks [222]. In general, although the last few years have witnessed significant advances in this technology (e.g., see [223]- [225] and references therein), more advanced/efficient resource allocation schemes, circuit implementations and prototypes, and distributed processing architectures to avoid costly signaling between the antenna elements, are still pose open research challenges.

D. WAKE-UP RADIO (WuR)
In the context of machine-type communications, duty cycling has become an attractive technique for extending the lifetime of devices with limited battery capacity. The mechanism consists of turning off the radio component of the device, pausing its MCU, and using a timer to activate the device periodically. However, by turning the main radio off, the device becomes incapable of exchanging data which may result in excessive communication delays [226]. This motivates the adoption of WuR which allow activating the devices' main radio on-demand when there is data to communicate. Since its power consumption is several orders of magnitude lower than that of the traditional main radio (WuR's average power consumption is at the magnitude of 1000 times lower than that of the main radio [227]), the WuR can be kept always on, in contrast to the duty cycling operation. Furthermore, although the EC for operating a clock is relatively low, it is still non-zero. However, by employing WuRs, we can eliminate this EC entirely. Fig. 19 illustrates an exemplary architecture of a WuRcapable device, which includes a low-power radio that receives and detects the WuR signal to activate the main radio for communication. The specific components and their inclusion in different archi- tectures can vary based on the design requirements and specific use cases.
A typical WuR setting is illustrated in Fig. 20. The main radio of the device remains deactivated (OFF) until it is required for communication, or until a special packet known as the Wake-up signal (WuS) is received by the WuR, which generates an interrupt signal to the main MCU to switch it ON. Subsequently, the main radio can exchange data packets with the other node in a conventional manner [228]. An illustrative WuS packet is also illustrated in Fig. 20. Firstly, the frame header consists of the wake-up preamble and start frame delimiter, a standard byte pattern agreed upon between the transmitter and the receiver. The preamble is used for synchronization whereas the start frame delimiter indicates the start of the frame that contains relevant information. Secondly, the address field contains the destination node ID for identifying the intended receiver. Thirdly, the payload contains application data, commands, or extra instructions specified by the UE or application. Finally, the error detection frame, using cyclic redundancy check, aka CRC, is to check data integrity [228].
Among the many benefits of using WuR, we summarize the most important ones below: • Sustainable EC: reduces the device EC in three aspects, i) removes unnecessary idle listening, ii) removes energy wasting related to startup/power-down, and iii) combining with uplink reference signaling, even high-speed UEs can reside in the sleep state for long periods, while not increasing the handover failure rate [229], [230]. • Short buffering delay [230]. • Synchronization assistance [229], [230]. unsaturated traffic scenarios [229], [230].
• Less signaling overhead [230]. 1) WuS standarization: The timeline depicted in Fig. 21 provides an overview of the key aspects and features associated with WuS in each 3GPP Release. Within the 3GPP standardization process, WuS was initially introduced in Release 15 [231], [232] as a paging signal sent over the physical downlink shared channel that "wakes up" a UE from an idle state so that it can prepare to receive data. Concerns towards enabling energy-efficient techniques resulted in a feature for 5G called lowpower WuS, in Release 16 [233], [234]. In what follows we primarily focus on standardization of WuR within 3GPP. Release 16 and Release 17 have been updated [235], [236] to include improved cross-slot scheduling. This means that the network can inform a device when there is a guaranteed minimum time interval between downlink transmissions of packets. As a result, there is a significant reduction in unnecessary RF operations. In addition, group focus WuS is defined, allowing the network to wake up a configurable group of UEs (instead of all UEs) by configuring the WuR of each UE to listen for a specific pattern or sequence in the WuS that is unique to their assigned group, thereby reducing the EC [235]. Enhancements in new features addressing the power consumption were released, such as a physical downlink control channel (PDCCH)based WuS that the network transmits before active on-duration within discontinuous reception cycles. This allows for UEs to avoid PDCCH monitoring during on-durations within which the network is anyway not transmitting any data to the UE. It was anticipated that this may enable between 15% and . Release 17 specified power saving enhancements for reduced-capability (RedCap) devices. However, as RedCap is expected to play a significant role in many applications, solutions enabling energyefficient WuR in RedCap are still missing. Release 18 promises to provide further solutions to energyefficient RedCap [237]. The current set of work items (WI) and study items (SI) for 3GPP Release 18 radio access networks (RAN) features is split into four RAN working groups. Specifically, the first group (RAN1), is responsible for the specification of the physical (PHY) layer including physical channels and modulation, PHY layer multiplexing and channel coding, PHY layer procedures and measurements as well as PHY layer-related UE capabilities. Within the 5G-Advanced Release 18 scope, low-power WuS and WuR constitute a new SI in RAN1, which aims at studying power-saving schemes that do not require existing signals to be used as WuS. The following study objectives are to be covered: i) evaluation methodology for lowpower WuS/WuR for power-sensitive, small formfactor devices including IoT use cases (such as industrial sensors, controllers) and wearables; ii) evaluation of low-power wake-up receiver architectures; WuS designs to support wake-up receivers; iii) L1 procedures and higher layer protocol changes to support the WuS; and iv) UE power saving gains compared to the existing Rel-15/16/17 UE power saving mechanisms and their coverage availability, and latency impact [238]. In addition, the work on WUR standardization is also being carried by IEEE, e.g., as a part of IEEE 802.11ba standard [239].
2) State-of-the-art: WuR appears as a promising technique for achieving a lifespan of beyond 10 years [242]. Therefore, migration to WuR-enabled IoT devices is foreseen [242], enabled by WuR's overwhelming saving capacity superiority [227]. In this regard, investigations have been addressed regarding the implementation of WuR in IoT. Authors in [242] present a prototype implementation of a WuR-enabled IoT via BLE targeted at an IoT scenario where battery-powered massive IoT devices do not support direct 3GPP connections. Real-life results show that the system meets the over 10 years lifetime target while satisfying the latency requirements for 5G IoT devices. On the other hand, the benefits of using passive WuR in wireless EH networks are highlighted in [241], whereas efforts to improve WuS in interference-free OFDM-based systems are made in [243]. WuS mechanisms with wake-up scheduling optimization have been also proposed. In [244] the authors use operational parameters, determined by BS at the beginning of the session, to save energy.
Meanwhile, in [229] the authors exploit an accurate traffic forecasting model to optimize the wakeup parameters, achieving up to 35% EC reduction. Additionally, [245] proposes a super-regenerative WuR solution to improve EE in human-body communication. The WuR operates at a sufficiently low data rate, e.g., 1.25 kbps, in order to obtain high sensitivity while keeping the EC low (only 40 µW). Likewise, in [246] the authors show the EE of using WuR in wireless body area network applications with event-driven traffic and propose a WuR capable of receiving small control commands besides WuS. Moreover, authors in [247] shed some light on Bluetooth Low Energy (BLE) compatible sensor devices enriched with a WUR to save energy. Results demonstrate that the WuR approach can be more EE than standalone BLE in low-latency applications (under 2s).
WuR proposals have been increasing in the literature, the different schemes vary depending on the Cross layer (interaction between layers to optimize MAC layer performance) [226] Wake-up circuit Duty cycled ("duty-cycled" WuR, which is activated periodically to sense the radio channel and send WuS)

Continues cycled (WuR's circuit is always ON) [240] Communication initiator
Transmitter Initiated (the transmitter first sends a WuS to activate the desired node and then sends data) Receiver Initiated (receiver sends a WuS asking for data, bidirectional WuS is also supported) [226] Addressing scheme Broadcast (the WuS is received by all the neighboring nodes of the sender) Identity-based (only the desired next hop node is wakening up) [226] Frequency band In-band with the main radio (main radio has predictable dormant periods, cheaper as there is no need for a separate antenna) Out-of-band (unpredictable dormant periods, it reduces interference, increases signal capacity, but may increase cost and complexity.) [228] Power source Active (battery-powered, immediate responsiveness but consumes extra power, ideal for real-time monitoring, rapid event detection, and instant data transmission.) Fully-passive (no internal power supply is needed, while hybrid combines EH with a small backup battery for responsiveness, limited operational range, and minimal EC.) [241] protocols used, the type of circuitry, and the application, among the most important classifications.
Here, we present the main classification of WuR summarized in Table VIII to shed some light on the various characteristics and wide application range.
3) Challenges and trade-off: However, using a WuR brings two main impacts on the devices' performance. On the one side, the problem of missdetection of the WuS appears. In this situation, the device does not receive the page occasion scheduled within the WuS for information exchange. As a result, the device misses the chance to wake up and has to wait until the next page occasion, increasing the latency of packets/information. On the other hand, due to the inherent simplicity of the WuS, the problem of false alarm needs to be addressed. In a false alarm event, the device/WuR receives a page occasion needlessly when no information is intended to transmit/receive. Therefore, special attention to these problems is needed when using WuR [229].
Next, since many unknowns still remain unaddressed when discussing WuR, we summarize some interesting research directions and challenges to pursue: i) WuS may complicate radio resource management and device scheduling in the network due to sleep patterns, reducing potential power savings. In this context, employing advanced MLbased scheduling algorithms that take into account the sleep patterns of devices may serve as a viable approach, facilitating optimized resource allocation and reducing power consumption [229]. ii) The problem of HW complexity and cost. If the WuR were to utilize a different frequency band than the main radio, the HW complexity and cost of devices would increase. In this sense, in-band operation and RF integrated circuit (RFIC)-embedded WuR implementation could be a possible solution However, this approach complicates resource management and reduces the available spectral resources for transferring application data. iii) beamformed WuS at mmWaves and mobility management is still an open challenge since beam sweeping for WuS is required to reach a desired device. The network should be able to optimize the number of beams in a single WuS burst utilized for waking up the device. iv) Applying WuR brings some trade-offs between EE and the different constraints, depending on the application scenario, like latency, reliability, and robustness [230]. Therefore, more research should be directed in this direction, especially when dealing with massive low-power IoT scenarios.

V. ENERGY-EFFICIENT MACHINE LEARNING
Native support of ML in 6G is essential for dealing with the increasing complexity and automation of networks while improving their performance [248]- [251]. ML techniques can help address issues such as increasing traffic demands, real-time QoS requirements, and resource allocation. However, these benefits usually come at high computational and memory requirements. Therefore, energy- efficient algorithms are paramount for network sustainability [252]. Several ML characteristics affect EC. In general, larger models with more parameters require more energy to be trained and exploited. In addition, training time (TRNT) and inference time (INFT) directly affect the EC at the training and inference phases, respectively. Therefore, the trade-offs between power cost and performance reward require special attention.
As shown in Fig. 22, ML approaches are broadly classified as supervised learning (SL), unsupervised learning (UL), and reinforcement learning (RL). In SL, the model is trained on a labeled dataset that contains the correct outcome for the corresponding input. SL can be used for tasks such as predicting energy usage based on historical data or identifying energy-efficient products based on specific features [253]. Meanwhile, UL does not require labeled data and therefore can be used, for instance, to identify patterns in energy usage data, detect anomalies that may indicate energy waste, or reduce the dimensionality of the data [254]. In RL, effective solutions are learned over time given constraints imposed by the inputs and without attempting to find hidden categories or structures. RL is suitable for solving problems with multiple optimal solutions [255], such as optimizing energy usage in buildings or predicting optimal times to charge electric vehicles. Notice that RL algorithms can be computationally intensive, but they have the potential to improve energy efficiency over time by learning from experiences and making adjustments accordingly [256].
The field of ML in wireless communications can be approached from three distinct perspectives: i) the network side, ii) the edge, and iii) the de-vice side, as illustrated in Fig. 23. Network-based computing treats mobile devices as data collectors sending data to cloud servers. The drawback of this scenario is the introduced overhead and potentially severe latency [257]. Notice that ML algorithms can be complex, data-hungry, and computationally costly on the network side. Contrary, ML at the device side must be simpler (i.e., of lower power/cost), while a balance/trade-off between these two extremes is achievable by implementing ML at the edge. Indeed, ML at the edge takes advantage of local processing and data storage capabilities while communicating with the network and devices. Specifically, pre-trained models are offloaded from the cloud to individual devices, such that they can make inferences locally. This allows for more efficient and effective use of ML in wireless communications. However, it can only support tasks that require lightweight computations [257]. Next, we discuss each of the approaches.

A. ML AT THE NETWORK SIDE
Currently, the radio access network (RAN) accounts for 73% of the total power usage in modern cellular networks [258]. Therefore, intelligent resource management is imperative to maximize EE, and thus minimize EC. Specifically, the current trend is to replace rule-based heuristics with optimal parameters configured through the knowledge acquired by data-driven approaches [258].
In this regard, ML models play a crucial role in enabling intelligent networks to characterize their environment, predict system changes in real-time, and react accordingly [252].
1) ML algorithms: Choosing the most appropriate ML algorithm is critical for solving EE problems on the network side. The choice of algorithm should be based on the specific problem, network architecture, and available data. A comprehensive approach that considers all aspects of the network, including hardware, software, and communication protocols, is needed to design energy-efficient solutions [259]- [261]. Therefore, it is essential to consider a holistic approach when selecting an ML algorithm for improving EE on the network side. This approach should consider the specific EE problem, the architecture and the characteristics of the network, the available data, and the strengths and weaknesses of each ML algorithm [259]- [261].  For instance, decision trees are suitable for routing optimization, while support vector machines are effective for network anomaly detection and prediction. Genetic and clustering algorithms are well-suited for optimizing the placement of network and edge resources [262]- [264], while clustering algorithms are also useful for grouping network devices, e.g., based on their EC patterns [264].
Among the numerous categories for ML techniques, deep learning (DL) is one of the most widely adopted algorithms at the network side [254]. Specifically, deep SL (DSL) and deep RL (DRL) are commonly used on the network side due to the availability of labeled data and the effectiveness and flexibility of the models. Both can be used to optimize EC by adjusting the network resources based on the network traffic load.
Combining feature extraction with prediction, DL models classify, predict and make accurate decisions more effectively than traditional ML algorithms [254]. This is because DL architectures such as multi-layer perceptrons, convolutional neural networks [265] and, more recently, transformers [266] can estimate complex mappings between input and labels in the training data [265] all while efficiently utilizing hardware-based accelerators such as graphical processing units (GPU) for both training as well as inference [267].
By optimizing the network resources based on the network traffic load, ML and DL, in particular, can significantly reduce EC and improve the sustainability of the network [264]. The benefits include the distribution of processing, avoiding redundant capacity in hotspots, and the efficient marshalling of big data, generated in-network or at user devices.
However, DL has its drawbacks. First, it requires large amounts of training data, whose curation and labelling may be costly and face privacy concerns. Second, DL algorithms are largely blackboxes with low interpretability and explanability. Third, DL models may require dedicated ML accelerators for efficient operation [258], [268]. All in all, the benefits of employing DL may be outweighed by the costs in many cases.
2) Use cases and applications: In 6G networks, ubiquitous ML algorithms are essential for providing efficient and personalized intelligent services. However, this poses a challenge in terms of data management and EC [269]. Indeed, EC can increase considerably unless energy-efficient approaches are used, as shown in [270]. Thus, there is an urgent need for more lightweight, flexible, and adaptive solutions with respect to environmental dynamics to minimize the EC of practical networks [258]. This has motivated the adoption of DRL for solving a variety of wireless communication problems [271].
Specifically, DRL has emerged as a potential tool to pave the way for data-driven optimization in newgeneration systems [258].
Using ML models for forecasting network environment can help DRL algorithms converge faster to optimal operational policies, resulting in faster adaptation to changing conditions and potentially improving energy efficiency in network operations [258]. The better understanding of network consumption that can be achieved through ML models can also help to make more informed decisions regarding network trade-offs, such as balancing performance and EE.
New DRL schemes have been proposed to manage advanced sleep modes in BSs [272]. Specifically, in this approach, the sleeping level length is set by the BS in a sequential manner. When the cell becomes idle, the BS starts in the deepest level of sleep and gradually switches to higher levels of activity. At each stage, the BS decides the number of slots during which the current sleep mode status will be kept. Most relevant components must thus only be active (consuming energy) when handling actual data [258]. In this scheme, the BS is the agent, and the environment is the network, including the traffic load, available energy, and the state of other network devices. The BS takes actions (i.e., sets the sleeping level length) based on its observations of the environment (i.e., traffic load and energy budget), and the RL algorithm provides feedback (i.e., rewards or penalties) based on the EC and network performance. The goal of the RL algorithm is to maximize the reward by finding the optimal sleeping level and duration for the BS, while maintaining the required network performance.
Demand forecasting, i.e., predicting how much time and resources will be spent on applications, is a key problem in data center management. Notice that good forecasting techniques can lead to minimizing EC by scheduling jobs efficiently [273]. Networklevel data usually exhibit significant spatio-temporal variations, which can be utilized for network diagnosis and management, UE mobility analysis, and public transportation planning [257]. In this context, DL has the potential to improve EE in a variety of settings by improving demand forecasting, optimizing resource allocation, and identifying patterns and trends in data that can be used to reduce EC.
Establishing a data collection path model is another solution for minimizing EC. Specifically, data collection can avoid visiting needless nodes and collecting unreliable data, resulting in outperforming traditional data collection methods in both energy and delay. In that direction, proactive caching can also contribute to energy minimization using forecasted lookup patterns to jointly optimize computation offloading policies and caching decisions [274].
3) Open RAN and virtualization as key technological enabler: Open interfaces are vital to support operators to swiftly introduce novel services, and enable operators to tailor the network to their own requirements. In this sense, RAN is evolving towards the concept of Open-RAN (O-RAN), which focuses on openness and intelligence [275]. O-RAN brings new business opportunities and encourages local 5G innovations. O-RAN aims at decoupling the RAN components from their underlying SW and HW components, enabling operators to cover more users in a cost-effective, secure, and energy-efficient manner [276].
In O-RAN, the RAN is disaggregated into three main building blocks: i) the radio unit (RU), ii) the distributed unit (DU), and iii) the centralized unit (CU) [275] as shown in Fig. 23. Also, the O-RAN ALLIANCE has defined different interfaces within the RAN including those for i) fronthaul between the RU and the DU and ii) midhaul between the DU and the CU. Another feature of O-RAN is the RAN intelligent controller, which adds programmability to the RAN and the ability to introduce new services and features [275]- [277].
The open interfaces/protocols of O-RAN enable a seamless integration of ML algorithms and models to enhance network intelligence. This enables ML for optimizing various aspects of the RAN, such as radio resource management, interference mitigation, and network planning. In addition, O-RAN supports real-time data processing and analysis, which is essential for ML algorithms that require fast and accurate decision-making. This enables intelligent and adaptive management of the RAN, resulting in better network performance, improved UE experience, and lower operating costs [277], [278].
Notice that network functions virtualization (NFV) enables network functions to be implemented as SW applications running on virtualized infrastructure, resulting in significant energy savings. It improves EE in several ways, such as reducing the required number of physical network devices, dynamically scaling network functions based on demand to avoid over-provisioning, and optimizing energy usage across the network through centralized management and orchestration [279]. Furthermore, NFV allows for the implementation of EE features like sleep modes and power management in network devices, leading to even further energy reduction while maintaining network performance. NFV has the potential to improve EE in networking and reduce EC [280]. In general, O-RAN and NFV are two complementary technologies that can work together to improve the EE and flexibility of wireless networks.
The O-RAN initiative is still in its infancy, with lot of work in progress that is expected to evolve. Hence, it is important that future research activities specifically focus on practical and real-world trials with respect to the virtualized RAN concepts [281], which heavily rely on self-organization and other approaches based on artificial intelligence.

B. ML AT THE EDGE
ML at the edge (popularly known as edge ML [248] or edge intelligence [282]) refers to the training and use of ML models across the computing continuum: on UEs, on edge nodes, and on cloud servers, rather than only in cloud-based centralized setups. ML at the edge has numerous applications across a wide range of industries, including healthcare, transportation, manufacturing, and retail [250].
Overall, ML at the edge is a rapidly evolving field with numerous opportunities for innovation and impact [283], [284]. It is often studied from two viewpoints: first, ML on edge refers to adapting ML methods for the distributed edge environment, while ML for edge is the use of ML methods for the benefit of the edge environment [248], [285].
As an example of ML on edge, distributed learning and inference allow the efficient distribution of ML computations across the computing continuum. This allows optimizing the overall processing time, resulting in lower EC of in/near-sensor devices [252]. Moreover, ML on edge can reduce latency, allow localized filtering of unwanted data, and increase system uptime as data is locally processed. On the other hand, edge ML can introduce various benefits, including the application of a predictive approach in troubleshooting. With edge ML, realtime data can be analyzed by ML models to proactively identify potential issues or anomalies before they escalate into critical problems. This proactive identification enhances operational efficiency and minimizes downtime. As a result, this approach could include analysis aimed at reducing the number of unnecessary offloads and task allocations, optimizing resource utilization, and improving overall system performance [286], [287].
However, some shortcomings need to be addressed: i) complexity due to coordination issues related to IoT constraints like processing power, memory, and delay in real-time applications; ii) heterogeneity, opportunism, and geographical distribution of computing resources; iii) fluctuating or intermittent connectivity; iv) security and privacy, both in the wireless as well as the fixed networking environments; v) standardization-related concerns like interoperability of IoT with ML integration, and vi) accuracy and latency issues in real-time application [250], [282].
1) ML algorithms: ML can be used for multiple purposes in the context of edge computing. For example, edge computing can filter data, with only relevant data getting transmitted between the user devices, the edge nodes, and the cloud. This results in substantial savings in terms of bandwidth and cost of data transmission [23]. Moreover, advanced ML techniques can be utilized, for example, to optimize computation tasks, make offloading decisions on a wireless device, and to identify the best scheduling solutions for working and sleeping time, thus lowering EC and enhancing EE [269].
There are several ML approaches that can be applied at the edge, depending on the application requirements and constraints. Among these approaches are distributed learning, referring to a family of methods such as federated learning (FL) for distributing the learning process as well as the training data across a number of nodes [288], and transfer learning, a method for transferring knowledge between ML models in different domains [23]. For example, FL and other variants of distributed learning have emerged as possible solutions for solving complex operational decisions at the edge side [269] (Fig. 23).
FL has several benefits, including privacy preservation, reduced communication costs, and the ability to adapt to the computational capabilities of the devices. However, implementation of FL requires special attention to [269]: • Expensive communication and synchroniza-tion. Two key aspects to be considered to further reduce communication overhead are: i) reducing the total number of communication rounds, and ii) reducing the number of gradients in each communication round. • Security/privacy/robustness issues. FL must: i) provide protection against malicious attacks and privacy-enhancing techniques, ii) tolerate heterogeneous HW, and iii) support robust aggregation algorithms. • Model size (MDS). The FL MDS might be unsuitable for the device due to its large size or might be unable to meet the network's time requirements, such as real-time demands from the UE. Thus, efficient training and inference are necessary for massive and heterogeneous networks. 2) Use cases and applications: Edge computing is a rapidly evolving field, and the trend is towards moving cloud functions to the network edges [289]. Edge computing benefits the advancement and implementation of 5G and beyond networks with applications such as augmented/virtual reality [290], low-complexity IoT [291], Internet of vehicles [292], and video stream analysis [293]. Delay-sensitive augmented/virtual reality applications can be migrated to edge servers to guarantee a high quality user experience with a timely response. For low-complexity IoT tasks, edge servers can reduce the hardware complexity and increase device lifespan by processing tasks that are originally done locally. Additionally, edge cloud networks close to vehicles can improve transportation safety, reduce traffic congestion, and provide value-added services. By utilizing the capabilities of MEC networks, video playback can be expedited, and user experience can be enhanced. Moreover, smart meters can be used to detect electricity consumption and gather data into a central controller to facilitate realtime power control, leading to improved systematic energy efficiency.
Moreover, the integration of ML and distributed ledger technologies (e.g., blockchain) can drastically improve the privacy and security of transactions in 5G and beyond networks [294], [295]. Smart cities, health care and transportation can all benefit from the trustless computing platform provided by blockchain. ML can enhance the security and efficiency of blockchain by providing intelligent data analysis and prediction capabilities. Moreover, ML can be used to optimize blockchain performance by predicting the demand for resources and dynamically allocating them [269]. Overall, the integration of ML and blockchain technology has the potential to transform the way we manage data transactions, resulting in EE [296].
Spatial and temporal correlation of IoT traffic data is exploited in dual prediction and data compression techniques leading to reduced EC and bandwidth. Dual prediction techniques use the correlation between the current and previous data to predict future data values. This prediction can be used to reduce the amount of data that needs to be transmitted, which leads to reduced EC and bandwidth usage. On the other hand, data compression techniques use the correlation between the different sensor nodes to compress the data before transmission. This reduces the amount of data that needs to be transmitted and hence the EC and bandwidth requirements [10].
At the edge, communication-efficient FL and efficient training aims at reducing the required communication overhead of the training process while maintaining or improving the model's performance. Federated parallelization is an efficient training technique of FL for parallelizing the training process across multiple devices to accelerate the training. Efficient training can also be conducted via distillation which consists of training a small model that mimics the behavior of the original model; therefore, significantly reducing the amount of required training data [269].
3) MEC -A key technological enabler: MEC, or multi-access edge computing, is an architecture for mobile networks. MEC was introduced to address the latency issue during mobile cloud computing offloading, pushing computing and storage resources to the edge with the aim of bringing those resources as well as applications and services near the endusers. MEC is characterized by two key features, namely, low latency and high workload capacity, stemming from proximity to users and their devices [297]. However, MEC has unique design considerations, such as complex wireless environments and MEC servers' inherently limited computational capabilities [10].
Computation offloading is one of the most significant features in MEC. Offloading may significantly prolong the lifespan of IoT devices by delegating computation tasks to edge devices as long as the communication overhead remains reasonable [264]. Specifically, the amount of energy that a mobile node can save by offloading an application depends on the number of computation instructions and communication data. If the computation instructions are much larger than the communication data, it is more energy-efficient to offload the computationintensive application to the server. However, if communication is expensive, it is better to carry out the application at the mobile node itself. The condition also depends on the bandwidth available for communication, where a large bandwidth can save communication time and improve EE [264]. Fig. 24 illustrates when offloading can save energy [298].
There are three categories of energy-efficient offloading methods: i) computation-based methods which involve partitioning the offloading application program and offloading computation-intensive parts, ii) communication-based methods which involve reducing the amount of communication by aggregating data or compressing it [298], and iii) hybrid/joint optimizations methods which involve collaboratively executing tasks on the mobile node and the cloud to minimize EC while ensuring the total execution of tasks [299]. The study in [299] has demonstrated that computation offloading was shown to reduce EC and increase battery life up to 50% for practical applications.
Notice that selecting the locations for the infrastructures of the MEC, specifically servers, is the first step towards constructing the MEC system. To make the server-site selection cost-effective, the system planners and administrators should account for the deployment cost and the demand for computation [262], [263], [297].
Green MEC is an emerging technology that combines the benefits of MEC with energy-efficient computing to create a more sustainable and ecofriendly approach to mobile computing [300]. The need for Green MEC arises from the growing awareness of the environmental impact of mobile computing and the increasing demand for sustainable solutions. As mobile networks and devices become more ubiquitous and essential to our daily lives, their EC and carbon footprint also increase. Green MEC provides a way to reduce this impact by optimizing the use of resources and energy-efficient computing techniques. Unfortunately, designing a green MEC is more difficult than creating a green wireless communication system. The amount of computational resources that need to be managed in order to have a satisfactory computation performance is greater, which makes traditional green radio techniques less practical [300].
In a traditional MEC system, the EC of a server is typically fixed, regardless of its utilization or workload. This means that even if the server is underutilized, it still consumes a significant amount of energy, leading to wastage and higher energy costs. Additionally, MEC systems may not be designed with EE as a primary consideration, leading to suboptimal performance and higher EC. One method for designing green MEC systems is dynamic right-sizing, where the EC of a MEC server is dependent on its utilization, and a server should consume energy proportional to its workload [301]. To achieve energy-efficient servers, the processing speeds of lightly-loaded edge servers can be reduced. Another method is geographical load balancing (GLB) [300], which utilizes spatial variations in workload patterns, temperatures, and electricity costs to direct the flow of workload between different data centers. The third method is renewable energy-powered MEC systems [300], which utilize clean energy sources such as wind and solar power to reduce the carbon footprint of MEC systems. These methods can help to reduce the EC of MEC systems and provide a carbon-neutral energy supply, leading to more sustainable and environmentally friendly network operations [300].

C. ML AT THE DEVICE: TINYML
TinyML is a field of ML focused on computationally efficient and HW-constrained algorithms for deployment on low-power IoT devices [302]. TinyML-equipped IoT devices can energyefficiently execute a broader set of tasks and make autonomous decisions without continuously relying on cloud/network services. This will surely reduce network traffic and latency in decision-making, and increase privacy. Moreover, TinyML has the potential to enable ML applications in intermittent connectivity scenarios because of its on-device data processing and autonomous decision-making features.
1) Use cases and applications: In the following, several use cases and applications motivating the deployment of ML at the device side are discussed.
• Energy Management: It is a promising solution to enable i) energy-neutral operation (ENO) in always-on IoT devices utilizing EH [303] and ii) efficient allocation of PV energy from a single-panel or off-grid system to multiple tasks [304]. In this regard, TinyML models can forecast future EH values to help devise a proactive ENO strategy for IoT devices. Besides, when the number of EH samples becomes insufficient to elaborate a forecast on the incoming ambient energy, energy management strategies can still benefit from TinyML by using the current battery state and previous EH measurements [303]. Again, PV power prediction is necessary for proper energy management/distribution among tasks. The study in [304] employs TinyML to perform PV power prediction and suggests that it can also serve as an indicator for measuring the effects of aging on the power-generating capacity of solar panels. • Radio access technology (RAT) selection: TinyML has the potential to jointly ensure EE on a multi-RAT IoT device and maintain the QoS of the data transmitted by the device. By considering situational characteristics, such as the available energy on a device and the size/urgency level of data to be transmitted, a TinyML-based multi-RAT IoT device can intelligently select a particular RAT for transmission at a given time instance [305]. • Complex event processing (CEP): CEP iden-tifies complex event patterns in real-time data streams from multiple sources using predefined logic rules to detect specific event sequences indicating certain conditions 13 . However, large IoT deployments pose a challenge for CEP in performing sequence matching over raw data due to unexpected events or outliers not covered by predefined rules, motivating the use of ML-based CEP. Nowadays, the data privacy and latency issues faced by centralized IoT are empowering the idea of performing ML-based on-device CEP in IoT. Interestingly, authors in [306] designed a framework that puts together TinyML and CEP for machine safety monitoring in a distributed IoT network. • Object detection running in EH IoT devices: Data privacy and latency requirements in various object detection applications, such as intelligent video surveillance and number-plate recognition, are encouraging the idea of ondevice processing in IoT. While performing data analysis over the captured image data and making a decision from it requires an ML algorithm; TinyML makes implementing an ML algorithm on a resource-constrained IoT device possible. Meanwhile, [25] has used TinyML to detect a person over a battery-less IoT device. Although owing to the dynamic nature of the EH environment, the following aspects must be taken into account for performing on-device processing: i) the available energy on a device, ii) the total time required for performing the task, iii) energy still needed to be collected to complete the task, and iv) the deadline to complete the task. The study in [25] also suggests the combination of TinyML and batteryless IoT devices for applications where devices generally require a long lifetime and are hard to reach. • Predictive Maintenance: Traditionally, predictive maintenance tasks, e.g., for detecting and preemptively solving the impending failures that a system might face, are performed on the cloud. Notably, TinyML makes it possible to perform predictive maintenance procedures over an MCU-based sensor device. For instance, authors in [310] implemented TinyML to perform anomalous behavior detection tasks over the sound recordings of the ToyADMOS dataset [311], where spectrograms (images generated from audio) are provided as input to the TinyML model. It is shown that optimizing the audio sampling rate used to form a spectrogram can lead to a decrement in the spectrogram dimensions, which further leads to a reduced INFT, required memory, and EC. In addition, [312] introduces TinyML for detecting various kinds of faults in a solenoid valve, which is an electro-mechanical actuator. Here, the transient response of the drive current of the solenoid valve is used as input to the TinyML model. This encapsulates information about the solenoid valve's electro-mechanical action, from which predictions about the present working condition can be made. • Tiny Robots: Legged robots use imitation learning [313] along with RL to learn their walking gaits. However, such procedures are not functional for tiny robots, which are lowcost resource-constrained robots that can potentially be used in search & rescue operations, military reconnaissance, space robotics, and routine equipment monitoring [314]. To address this, authors in [315] proposed exploiting TinyML techniques, such as graph freezing and float16 quantization, to shrink the trained walking gaits of the neural network (NN) model size by 8×. Graph freezing made all NN variables constant and float16 quantization changed the floating point weights from 32-bit to 16-bit. 2) Enabling techniques: Next, we discuss the techniques to support energy-efficient ML models. They are illustrated in Fig. 25, while representative state-of-the-art works using them are listed in Ta [307], [317] Smart health-care, image classification, object detection

ROMANet
Compared to the baseline scheme SmartShuttle [318], the number of DRAM accesses is 12%, 36% and 45% smaller for the AlexNet, VGG-16, and MobileNet, respectively. Moreover, the DRAM access energy is 12%, 36%, and 46% smaller, while the numbers are 12%, 34%, and 45% smaller in terms of DRAM operations higher accuracy compared to the MobileNetV2 [322] and ResNet-18 [323] frameworks, while utilizing relatively less SRAM and Flash. All in all, MCUNet may provide energy-cummemory efficient NN architectures that can be implemented on an MCU. • Parallel ultra-low power (PULP): PULP is an architecture for IoT processors that provides SW-level acceleration for TinyML algorithms. PULP brings into play both data-level and thread-level parallelism to deliver a steady performance irrespective of the operating voltages, and allows an energy-efficient operation by enabling the ML computations within a few milliWatts. As shown in Table IX, PULP uses 3.5 times less energy than the ARM Cortex-M4 processor and also exhibits a significantly lower INFT. For instance, authors in [324] adopted PULP-based MCU to facilitate the parallel run in non-neural ML algorithms, while a fast NN (FANN)-on-MCU framework supporting both ARM Cortex-M series MCUs and PULP-based MCUs is used in [320]. The latter framework can be used with both fixed-point and floating-point NN models. • Model compression: Given the often limited computational resources and power of UEs and edge nodes, the ML models used need to be both small and efficient. This necessitates the use of model compression techniques, a crucial subfield of TinyML. Model compression is a collection of techniques aimed at reducing the size of a ML model, thereby enhancing its efficiency in terms of memory usage and computational requirements [325]. These techniques are integral to the successful implementation of TinyML, as they enable the creation of models that are compact enough to fit on the device and efficient enough to run with limited power. Several model compression techniques are commonly employed in TinyML: i) quantization, reduces the precision of the numbers used in the model, significantly decreasing the model's size and computational requirements [326]. For example, coarse quantization drastically reduces the precision of an NN model's parameters to less than 8 bits, making the model suitable for an MCU. Moreover, the quantized NN model enjoys less EC and faster computation than its full-precision counterpart. However, these advantages come at the cost of lower NN accuracy [327]; ii) pruning, eliminate parts of the NN that contribute little to the output, such as weights that are close to zero [328]; iii) knowledge distillation, trains a smaller model (the student) to mimic the behavior of a larger model (the teacher). The smaller model is then used in place of the larger model, resulting in significant computational savings [329]; and iv) weight sharing, a tech-nique that involves using the same weights for multiple neurons, reducing the total number of weights that need to be stored [330]. Through the application of these and other techniques, it is possible to create small, efficient models that can run on low-power devices, thereby expanding the range of applications for machine learning. • Self-attention: The advent of Transformer architectures [266] has revolutionized the field of sequence data processing. Central to the operation of Transformer architectures is the self-attention mechanism. This mechanism allows each input in a sequence to weigh the importance of all other inputs, thereby enabling each element to 'attend to' or 'focus on' all other elements in the sequence [266]. Enabling high levels of parallelization and increased efficiency, this is a significant departure from traditional recurrent NNs or long short-term memory networks, which typically focus on preceding words in a sequence.
In the context of TinyML, the benefits of selfattention and Transformer architectures can be manifold. Firstly, the ability to process sequence data in a non-temporal manner allows for more efficient computation, which is crucial for low-power edge devices [331]. Secondly, the parallelizability of Transformer architectures makes them well-suited to the resourceconstrained environments typical of TinyML applications. By processing all elements of a sequence simultaneously rather than sequentially, Transformers can deliver faster inference times, which is crucial for real-time applications on edge devices [332]. However, it should be noted that the original Transformer models are often too large and computationally intensive for TinyML applications. Therefore, further research on model compression techniques and efficient Transformer variants, such as TinyBERT [333] and DistilBERT [334] is of paramount importance to fully harness the potential of Transformer architectures in TinyML. For example, AttendNets, a TinyML model, combines self-attention with a machine-driven design exploration, resulting in a compact deep NN with low-precision parameters. Using At-tendNets for pictorial recognition on a low-power device and testing it on the ImageNet 50 [335] dataset, it shows higher accuracy and EC and memory consumption in comparison to the benchmark framework MobileNet-V2, as shown in Table IX. • Memory Access: HW accelerators, consisting of DRAM (off-chip part), and SRAM and compute engine (on-chip parts), are required to implement complex NNs, such as CNN, on a device. The portion of the NN layer that an accelerator can process at a single time instance depends on the data storage capacity of SRAM. Also, the same input data could be used in multiple NN operations, which leads to multiple DRAM accesses for the same data. Notably, the energy required for DRAM access is relatively higher than that for a NN operation, which means that a significant amount of energy can be saved by reducing the average DRAM access energy in an inference. To achieve this, authors in [317] recommend using ROMANet [307], a memory access technique to cut down the average energy-per-DRAMaccess, on-chip buffer access energy, and the number of DRAM accesses. Specifically, RO-MANet partitions a NN layer into various portions and then schedules the processing of these portions to minimize the number of DRAM accesses. Based on the data storage capacity of SRAM, data is partitioned into various blocks. Then, ROMANet maps them to the available DRAM and SRAM resources to minimize the row buffer conflicts and maximize the banklevel parallelism, respectively. The advantages of using ROMANet with various NN architectures, such as AlexNet [336], VGG-16 [337] and MobileNet [338], are clearly visible in Table IX. The reader interested in TinyML can refer to the python package hls4ml [339], which allows designing ML algorithms for low-power FPGA or ASIC devices. Through hls4ml, it is possible to optimize the hardware implementation of TinyML models by leveraging the parallelism and pipelining capabilities of FPGA. This can lead to more efficient use of hardware resources and reduced EC.
3) Future Directions: Bio-inspired optimization [340] is a way to brush aside unnecessary computations in an ML algorithm, providing potential benefits also in TinyML. Furthermore, alternate NN models, such as spiking NN [341] and analog NN [342], should be investigated for designing a TinyML model. Also, there is a need to find new computing models for TinyML systems. Furthermore, the existing computing models, such as inmemory computing, require further research in the context of TinyML systems.
MobileNet is a baseline deep NN model for edge computing. However, no baseline TinyML model is currently available for the end devices. Finally, we thus propose that a baseline TinyML model should be designed to provide a reference point for future TinyML models.

VI. ENERGY-RELATED PERFORMANCE METRICS
The potential of a technique/technology, especially when compared to a competitor, can only be assessed via relevant metrics providing quantified performance information. In this section, we discuss several performance metrics related to energy that serve this purpose for the discussed techniques and technologies at a component/device and/or system/network level. They are classified as energy conversion and transfer metrics (Section VI-VI-A), energy storage and consumption metrics (Section VI-VI-B), EER metrics (Section VI-VI-C), and other metrics (Section VI-VI-D).
The key features of the discussed performance metrics in terms of related energy processes, application level (at the component/device or system/network level), and relevance/applications are summarized in Table X. Finally, Table XI lists some relevant performance targets (i.e., KPIs) for the current and next generation of wireless systems. The numeric values are indicative and were extracted from the vast literature consulted for this work, together with data extrapolation and trend analysis in some cases.

A. ENERGY CONVERSION AND TRANSFER METRICS
Every energy conversion/transfer process introduces losses and other potential non-linearities affecting device/network EE. Some related metrics are discussed below.

1) Component EE (CEE):
This metric refers to the ratio of the output energy to the total energy at the input of a certain electronic component. CEE considers losses due to heat, friction, and other inefficiencies, and describes the effectiveness of electronic components such as power supplies, power amplifiers, motors, filter circuits, phase shifters, etc. The ideal CEE value is 100% or 1.
2) EH Input/Output Relationship: The main parameters impacting the performance of EH circuits are sensitivity, saturation, and conversion efficiency. Sensitivity refers to the minimum magnitude or change in the input signal required for the EH circuit to produce a usable electric signal. Saturation, on the other hand, occurs when the circuit reaches maximum output and thus the harvested power is independent of any input power increase. Both sensitivity and saturation metrics correspond to absolute values of power or energy, while the EH conversion efficiency is defined as the ratio of the output electrical power to the total input power. Notably, the conversion efficiency is a non-linear function of the input power, and it is common in EH datasheets to include input-output power transfer curves under different settings and emphasize the maximum achievable conversion efficiency (MACE). Notice that in WET systems, one can exploit the fact that the conversion efficiency depends on the operating conditions, such as the operating frequency, incident energy, and distance to the source, to drive the EH circuits to their maximum conversion efficiency.
3) APD of EH transducers: This metric refers to the ratio of the EH transducer's peak output power to its size under specific operating conditions, thus it is given in W per unit of area/volume. Hence, it characterizes the level of miniaturization that an EH transducer can achieve. 4) PTE: This metric, given as a percentage or a dimensionless quantity, refers to the ratio of the power captured by EH receiver(s) to the power consumed by the corresponding transmitter(s). This metric captures the joint energy conversion efficiency of the transmitter(s), the medium, and the EH circuit(s). Several factors can affect PTE, including the distance between the ET and EH receiver and the specific WET technology.

5) EH Coverage:
This comprises a set of metrics characterizing the effectiveness of EH processes in a given network deployment. For example,  Notice that EOP and MDHE are relevant metrics for networks with battery-less devices. Meanwhile, AHE can suffice in scenarios where devices are equipped with batteries, especially when the AHE is measured per device. All in all, these metrics measure the extent to which a region or network has access to EH services and/or can maintain a reliable and affordable supply of energy over time. Therefore, they can be used to assess the level of energy availability/ubiquity.

B. ENERGY STORAGE AND CONSUMPTION METRICS
1) Battery capacity (BCAP) and energy density (ED): BCAP, given in ampere-hour (Ah), refers to the amount of potential energy (usually chemical energy) that batteries can store. Over time, batteries' electrodes degrade due to continuous chemical reactions causing capacity fades and unexpected voltage drops. Notice that even when batteries remain disconnected from an external circuit, the internal chemical reactions will cause a self-discharge thus reducing the amount of energy stored over time.
The ED of a battery measures the amount of power a battery can store per unit of volume (Wh/m 3 ) or weight (Wh/kg). High ED batteries can provide longer battery life while maintaining a compact and lightweight form factor. The factors that mainly impact batteries' ED are their chemistry and internal cell design.
2) Lifetime: In general terms, the lifetime of a device/component is a measure of its durability and refers to the period of time over which the device/component is expected to remain functional and perform its intended tasks without requiring significant repairs or maintenance. When the device/component is a battery, we refer to battery lifetime (BL), which is influenced by many factors such as the chemistry and design of the battery, the conditions under which it is used and stored, the level of usage and discharge, and the charging and maintenance practices over time. The BL may be given by the number of charging cycles the battery can endure before its capacity decreases significantly, or by the time the battery can hold a charge before it needs to be recharged.
Finally, in sensor networks, there is also the socalled network lifetime, which refers to the time period a network functions as intended. This metric is obviously related to the lifetime of the network devices, but the relationship depends on the specific network's applications and performance requirements.
3) Energy Consumption (EC): The EC of a component/device, specified in J, depends on the specific operation modes, i.e., active, idle, and sleep states, and the time spent in each. As briefly discussed in Section IV-IV-D, duty cycling takes care of properly scheduling these states to reduce EC subject to QoS requirements. Notably, tasks such as computation and communication, mostly executed in active modes, consume different amounts of energy. For instance, transmitting typically consumes more energy than receiving data wirelessly, and the EC from computational tasks increases with the computation complexity. Obviously, the EC of a system/network is given by the sum of the EC of its components/devices.
The EC scales linearly with the floating-point operations per second (FLOPS) and the million instructions per second (MIPS) in digital signal processors and computer systems, respectively [344]. The scaling factor depends on the amount of computational work, the number and type of functional units, the clock frequency, and the complexity of the instruction set architecture. Therefore, FLOPS and MIPS units are usually more useful than J units for EC performance comparisons in these systems. 14 Nevertheless, when assessing network EC performance figures, one may inevitably rely on J units and averages over different states, workloads, components/devices, etc.
Finally, we would like to mention three other metrics related to EC: • Net harvested energy (NHE) represents the amount of energy that is available for use by an EH device after accounting for losses, i.e., due to energy conversion and storage, and EC related to the EH protocol, e.g., for CSI acquisition in RF-WET systems [345]. • Relative power saving (RPS) quantifies the EC reduction driven by a certain approach compared to a specific benchmark [229]. • Peak demand (PKD) measures the highest level of energy demand during a given period. It can be used to assess the capacity of a renewable energy system to meet peak demand and identify opportunities for demand management and energy storage [346].

4)
Green EC Share (GECS): As discussed throughout the paper, the use of renewable sources is fundamental to support sustainability. Consequently, the exploitation of green energy sources will continue expanding in the next few years, thus, becoming increasingly relevant to quantify their relative contribution to the total energy budget of devices/networks/systems. For this, the GECS metric, which quantifies the portion of the energy consumed from renewable/green sources relative to the total EC, is undoubtedly attractive. Through the GECS, one can get a clear understanding of the extent to which a given device/network/system relies on green energy sources. The GECS can even evolve to quantify the energy contribution from specific renewable sources, thus, supporting more granular insights. Notice that this metric can be used as a benchmark to set targets for increasing the green energy share(s) of a solution, as part of a broader sustainability strategy.

5) Levelized cost of electricity (LCOE):
This metric evaluates the economic feasibility of electricity generation systems. It is defined as the ratio of the estimated costs of electricity divided by the estimated power plant's electricity generation during its lifetime. Thus, LCOE is given in monetary units per kiloWatt-hour (e.g., $/kWh). In the context of renewable energy, the estimation of the LCOE is heavily determined by the availability of ambient energy over the lifetime of the system. In fact, the number, size, and complexity of the EH transducer/circuits and the energy storage vary depending on the geographic location. Besides, the uncertainty in the amount of electricity generated by the system may cause unexpected expenses due to outages or energy trading with the grid to cope with sporadic high electricity demands. Fortunately, energy trading can also be a source of revenue if the network sells its surplus energy.
One important consideration is the market dynamics that may occur over the anticipated lifespan of EH systems, which is typically around 30 years. This timeframe allows for the potential increase in renewable energy market penetration, as well as the development of technological breakthroughs. These advancements could lead to reduced LCOE due to lower costs for spare parts and more efficient electricity generation.
In WET-enabled networks, we can extrapolate the LCOE to estimate the ratio between the cost of electricity to the harvested energy. In such cases, the PTE of the WET technology, the deployment of the ETs, and the conversion efficiency of the EH circuits may be the main impacting factors when estimating the LCOE.
6) Grid reliability/stability (GRS): This metric captures the ability of the grid (or microgrid) to maintain a stable energy supply at all times, despite fluctuations in demand and supply, especially when considering contributions from renewable energy sources. It can be given in percentage and used to assess the resilience of the grid and identify opportunities for improving grid management and infrastructure.

C. EER METRICS
An alternative (and popular) way to define EE is the ratio between the achievable QoS performance and the corresponding energy/power consumption that is required, or the ratio between the energy/power consumption and the corresponding achievable QoS performance. We define this category of EE metrics as EER, and notice that they differ from one another in the type of QoS performance metric that is used.
In any case, although EER metrics are not always material for optimization, they are certainly useful for comparison and drawing valuable performance insights. Next, we briefly discuss some of the most relevant QoS performance metrics and associated EER units.
1) Data transfer metrics: This is a set of highlyrelated measures used to evaluate the performance of data transmission over wireless networks, e.g., • spectral efficiency, in bps/Hz, measures the amount of data that can be transmitted per unit of time and frequency spectrum; • THP (capacity), in bps, measures the amount of data that can be transmitted over a wireless link or network (under ideal conditions) within a given period of time; • ϵ−capacity, in bps, constitutes the best upperbound for the attainable THP (capacity) supporting an outage probability that does not exceed ϵ; • goodput, in bps, measures the useful data rate (considering overhead and error correction) delivered to the end-user/application; • effective capacity, given in bits per channel use (bpcu), constitutes the highest arrival rate that can be served by a network under a particular latency constraint, thus capturing physical and link layers characteristics. The corresponding EER metrics are often given in bits/Hz/J, bits/J (or J/bit), or bpcu/J.
2) Range (RG) or coverage area (CA): The RG refers to the maximum distance over which a specific device, or a generic device representing a certain technology, can transmit and receive wireless signals effectively, according to, e.g., target QoS guarantees. Similarly, the CA refers to the geographical area within which wireless connections can be established and maintained with certain QoS guarantees. 15 The corresponding EER metrics are often given in W/m or W/m 2 , i.e., characterizing the required amount of power per distance/area unit.
Related to CA, but specifically for IRS-assisted networks, the authors in [350] proposed the socalled area of influence (AroI) metric. The AroI comprises the area of significantly improved wireless connectivity triggered by the IRS(s) when optimizing for the whole area under consideration instead of a single nominal receive position. Notably, the spatial resolution of IRSs in horizontal and vertical axes depends on the specific element array, thus, the AroI specified in m 2 might be insufficient. Similarly, in non-terrestrial networks, including those composed at least partially of UAVs, high-altitude platforms, and satellites, the network becomes inevitably three-dimensional. In such scenarios, the CA may evolve to coverage volume (CV) and the corresponding EER to be measured in W/m 3 .

3) Bandwidth:
This metric corresponds to the frequency bandwidth in which a technology or device operates, e.g., harvests sufficient energy in the case of RF-EH networks. A special case of interest is in IRS-assisted networks, where there is the socalled bandwidth of influence (BoI), which indicates the frequency range in which, any wave hitting the RIS, will be reflected [350]. The corresponding EER metric is given in Hz/W. 4) Accuracy: This metric somewhat characterizes the proximity of a measured/estimated value to the true value. Both relative units, such as percentages and relative errors (e.g., probability of error, miss-detection (PMD), false-alarm (PFA)), and absolute units, such as mean absolute error (MAE), root mean square error (RMSE), and standard deviation (STD), can be used depending on the scenario/application. Next, three specific examples are briefly discussed.
• ML accuracy (MLA) is a crucial KPI for any ML model, especially TinyML, which is subject to tight HW and SW-related constraints. In general, the definition of learning accuracy depends on the task for which the ML model has been employed. For example, for anomalous behavior detection [351] and object detection [25] tasks, MLA may be the percentage of correct predictions an ML model makes. Meanwhile, for tasks such as PV power prediction [304], learning accuracy can either be defined by the RMSE or MAE among the predictions and actual observations. Notice that the EER metric corresponding to an ML model may be expressed as the ratio between the relative/absolute accuracy that is achievable per EC at each inference step. Specifically, the corresponding EER metric can be expressed in %/J when using relative accuracy metrics. Finally, recall that FLOPs can be also used as a measure of energy as discussed in Section VI-VI-B-VI-B3, thus, resulting in %/FLOP EER units. • Localization accuracy (LA) is the precision with which a system can estimate the position of an object or feature in a given environment. LA can be given in angular or distance units, which may correspond to RMSE, MAE, or STD statistics. For instance, indoor localization systems supported by future 6G wireless communication systems may operate with sub-meter STD LA [352]. The corresponding EER metric may be given in rad/W or m/W. • Wake-up accuracy (WUA) refers to the ability of the WuR to reliably detect the WuS and avoid false alarms. The accuracy of WuR is typically characterized by two metrics: PFA and PMD [229]. PFA refers to the probability that the WuR will wake up erroneously in the absence of a WuS (false positive, FP). On the other hand, PMD refers to the probability that the WuR will fail to wake up in the presence of a WuS (false negative, FN). Specifically, PFA is given by FP/(FP+TN) while PMD is FN/(TP+FN), where TP (true positive) is the number of times the WuR correctly detects the WuS, and TN (true negative) is the number of times the WuR correctly determines that the WuS is not present. Moreover, WUA is calculated as (TP + TN)/(TP + TN + FP + FN). Both PFA and PMD depend on various factors such as the type of WuS used, the power level of the signal, the distance between the transmitter and receiver, and the RF conditions of the environment. The accuracy of WuR can be improved by optimizing these parameters and using more sophisticated WuS detection algorithms [230].
D. OTHER METRICS 1) Energy proportionality coefficient (EPC): This metric represents the power consumption of a device/system as a function of the offered load. In general, the observed power consumption increases non-linearly with the load and the power consumption in an idle state is often non-negligible, e.g., network switches consume up to 85% of their peak power consumption when idle [353]. EPC is defined in [−1, 1], where EPC = 1 (−1) means that each increase in load leads to an equal increase (decrease) in EC, while EPC = 0 describes the case when the system EC is constant and does not depend on the load.
Three other metrics related to energy proportionality are [353]: • Energy proportionality index (EPI), which captures the difference between the measured power and the ideal power, i.e., the power that the device should consume if it is fully energy proportional. EPI is expressed in the region between idle and peak power consumption only. EPC = 0 (1) indicates that the EC is agnostic (fully energy proportional) to the workload. • Idle-to-peak power ratio (IPR), which measures the ratio between the idle and the peak power consumption. IPR values tending to zero indicate energy-proportional designs. • Linear deviation ratio (LDR), which captures the deviation of the observed power consumption from the fully proportional case, i.e., a straight line connecting idle and peak power consumption values. LDR = 0 corresponds to a perfectly linear system. 2) WET exposure level: This metric describes how different WET energy-carrying signals disturb the surrounding environment.
A common metric to characterize RF transmissions is the power spectral density (in W/Hz) which describes the power distribution of a signal across different frequency components. In the spatial domain, one can resort to the effective isotropic radiated power (in W) which is the hypothetical power that an isotropic antenna must radiate to yield the same signal strength as the actual RF transmission in the direction of the antenna's maximum gain. Besides, the electromagnetic field (EMF) radiated by RF sources may cause disturbances in nearby equipment operating in the same frequency band. For such a case, international organizations have resorted to electromagnetic compatibility-a boolean metric that takes on pass or fail-which describes the ability of electronic equipment to successfully operate in a certain electromagnetic environment without being affected by (or affecting) other devices.
International EMF exposure limits are available, e.g., the ICNIRP guidelines [354], which have been set with substantial margins to protect against both short-and long-term health effects. Limits are provided for whole-body and local exposure scenarios and also for general public and occupational exposure. The exposure limit values relate to physical quantities that are closely related to RF-induced adverse health effects and are referred to as "basic restrictions" in the ICNIRP guidelines. For instance, the specific absorption rate (SAR, in W/kg), specific absorption (SA, in kJ/kg), absorbed power density (in W/m 2 ), and absorbed energy density (in kJ/m 2 ) are the basic restriction quantities measuring the absorption rate of electromagnetic energy in the human body. Other quantities that are more easily evaluated, termed "reference levels", e.g., incident power density (in W/m 2 ), incident energy density (kJ/m 2 ), electric field strength (in V/m), and magnetic field strength (in A/m) are also provided. These have been derived from the basic restrictions to provide a more practical way for demonstration of compliance. Reference levels provide an equivalent degree of protection as the basic restrictions, and thus an exposure is taken to be compliant with the guidelines if it is shown to be below either the relevant basic restrictions or relevant reference levels. There are also EMF exposure assessment standards available for wireless communication and wireless energy transfer technologies.
Ultrasound signals are commonly evaluated using a metric that measures the intensity of the pressure wave in dB, scaled to a frequency sensitivity response curve. However, it is important to note that the duration of exposure to the acoustic signal is also relevant, as prolonged exposure can cause hearing impairments in animals whose hearing response falls within the operating frequency of the WET technology. In humans, the impact of ultrasound signals is primarily characterized by two indexes: the thermal index and the mechanical index. The thermal index is a unitless metric that measures the ratio between the acoustic power penetrating the skin and the amount of power required to raise the body temperature by one degree Celsius. On the other hand, the mechanical index indicates the ability of the acoustic signal to create tissue mechanical stress and damage.
3) Wake-up time: or start-up time refers to the time it takes for a device to wake up from a lowpower sleep state and become fully operational, thus it is given in time units. This is an important factor in determining the EC and responsiveness of the network. Typically, wake-up time depends on several factors such as the type of device, the complexity of the wake-up process, and the power management scheme used. For instance, some devices may require a longer wake-up time to initialize, calibrate, or pre-heat sensors or to establish a wireless connection. Similarly, wake-up time can be longer if the device is in a deep sleep mode that requires more time to restore the device's state. To optimize EC and responsiveness, it is important to minimize the wake-up time. This can be achieved by using low-power HW components, optimizing the SW for fast wake-up time, and selecting appropriate power management schemes [355].

4) INFT:
INFT refers to the time it takes for an ML model to make a prediction (i.e, inference) on a new data sample. INFT can be affected by the complexity of the model, the size of the input data, the HW configuration, and the SW implementation. This metric is relevant for scenarios demanding realtime inference and/or low EC (since EC is proportional to active time) such as power management and sleep mode handling.

5) MDS:
This metric refers to the amount of memory required to store the model's parameters and configuration, thus it is given in bytes (kB, MB, etc.). A small MDS makes the ML model easier to deploy on resource-constrained devices and/or can lead to fast INFT since the number of computation parameters during the inference phase is limited [356]. However, reducing the MDS can also have a negative impact on model performance, since smaller models may have limited capacity to capture complex patterns in the data.
6) Peak SRAM for inference: SRAM is an onboard memory space that accepts both read and write operations. Thus, an ML model's mutable parameters during its runtime are stored in SRAM. Notably, the peak memory required by a TinyML model during its inference, also called peak SRAM, becomes noteworthy in MCUs because of the limitations on onboard available memory. The standard units to express peak SRAM are kB and MB. Also, [319] states that (a) peak SRAM depends on the memory scheduling procedure carried out by an inference library, while NAS decides MDS, (b) available SRAM on an MCU sets an upper bound on peak SRAM, while flash memory on an MCU constraints MDS.

VII. CONCLUSIONS & OUTLOOK
In this work, we provided valuable insights into energy-sustainable IoT and claimed it could only be supported by the harmonious coexistence of EP, ET, and EE processes. The latter refer respectively to the charging processes exploiting green energy sources, the intentional movement of energy from one device/system to another, and the ability of a device/system/process to perform its intended function with minimal energy. We overviewed the main technologies corresponding to these processes, together with use cases, recent advances, challenges, and research directions. Specifically, EH technologies based on light, heat, microbial fuel cells, vibration, flow, and RF were discussed within the EP processes. In the case of ET, the focus was on RF, inductive/capacitive coupling, laser, and acoustic-based technologies together with energy trading and microgrids. Meanwhile, in terms of EE support, we focused on technologies enabling lowpower communication, namely BC, metasurfaceaided communication, radio stripes, and WuR, and also ML approaches aiming to reduce the EC burden of application tasks at the device/edge/network side. In general, the appropriateness of a specific technology/technique depends on the characteristics and performance requirements of the target application, and increasingly on their sustainability support level. This may be assessed by considering proper performance metrics. Indeed, we discussed relevant performance metrics to assess energy-sustainability potential and listed some relevant target values for specific technologies in the next generation of wireless systems. Table XII compiles a summary of the main challenges and corresponding research directions of all the discussed technologies. Finally, notice that the focus of this work was on IoT energy-sustainability aspects during the operation phase. However, a truly self-sustainable ecosystem must consider sustainability aspects along the entire product lifecycle, i.e., planning, manufacturing, deployment, operation (including the target use case), maintenance, and disposal. Therefore, the natural progression of our work is to expand in this direction.