Nitrogen-Oxyanion-Doped HfO2 Resistive Random-Access Memory With Chemically Enhanced Forming

The high forming voltage of resistive random-access memory (RRAM) is one of the key bottlenecks for its integration in advanced silicon technology nodes. Here a nitrogen-oxyanion-doped (N-doped) hafnium oxide (HfO2) RRAM with overall improvement on forming voltage, on/off ratio and endurance is demonstrated. The critical electric field of N-doped RRAM for forming is 40% less than that of undoped RRAM. Therefore, a thicker resistive switching layer (RSL) can be used to obtain the same forming voltage, which benefits the on/off ratio and endurance performance. The N-doped RRAM achieves <inline-formula> <tex-math notation="LaTeX">$3\times $ </tex-math></inline-formula> improvement in on/off ratio and <inline-formula> <tex-math notation="LaTeX">$10\times $ </tex-math></inline-formula> improvement in endurance at the forming voltage of 2 V, a value applicable for integration with advanced silicon technology node. We propose the hypothesis that the nitrites ((NO2)−) in the RSL of N-doped devices promote forming through introducing extra chemical process as well as additional conductive paths. Physical characterization and first-principles calculations are further carried out to validate our hypothesis.


I. INTRODUCTION
R ESISTIVE random-access memory (RRAM) has been extensively studied due to its advantage in simple device structure, excellent compatibility with Si CMOS process for large-scale integration as well as high switching speed and low power [1], [2], [3]. The resistive switching characteristic of filamentary type RRAM attributes to the repeated growth and rupture of conductive filament (CF) inside the resistive switching layer (RSL). A large forming voltage, or critical electric field (∼5MV/cm for typical HfO 2 RRAM) is usually required to establish the CF at the beginning of operations [4], [5]. As CMOS keeps scaling, the required forming voltage also gradually increases due to narrow contact effect. Such a high forming voltage has become the bottleneck for integrating RRAM with advanced Si CMOS [6]. Various methods have been proposed to reduce the forming voltage of RRAM, including high temperature forming [7], energy band engineering [8], and material doping [9], [10], [11], [12], [13], [14], [15]. These methods, however, bring little change to the intrinsic electrochemical nature of forming. During forming, the oxygen atoms in the RSL get ionized, migrate under electric field and then combine with oxygen reservoir. This results in the formation of CF [16]. As a result, these methods often achieve optimization at the cost of increased process and operation complexity or degradation of other resistive switching characteristics such as on/off ratio and endurance.
In this work, we propose a novel nitrogen-oxyanion-doping (N-doping) method to improve the forming performance of HfO 2 RRAM. Different from previous doping methods, which use metallic element (Al [9], Mg [10], Ge [11], Ag [12]) dopants in the form of cations or non-metallic element (N [13], [14], F [15]) dopants in the form of anions, we use nitrogen oxyanions to dope HfO 2 RRAM. Magnetron sputtering is used to induce N-doping, where the high-energy plasma atmosphere promotes the chemical reactions between N, O and Hf. The N-dopants introduced extra chemical process as well as additional leakage paths to facilitate forming. The N-doped RRAM realizes a comprehensive improvement in forming voltage, on/off ratio and endurance. Physical experiments and first-principles calculations are performed to further investigate the mechanism of N-doping.

II. EXPERIMENTS
In this work, a 1kb one-transistor-one-resistor (1T1R) array was fabricated to statistically study the characteristics of HfO 2 RRAM. Fig. 1(a) shows the schematic of the 1kb array, in which the transistors and peripheral circuits were fabricated using a standard 110 nm Si CMOS process. The size of RRAM is 700 × 700 nm 2 . The length and width for the control transistor is 1 µm and 1.5 µm respectively. The RRAM stacks were deposited by magnetron sputtering (AJA ATC 2200-V). N-doping for HfO 2 was introduced by flowing N 2 together with O 2 to the chamber. The ratio of N 2 to O 2 is optimized as 9:1 to ensure a proper atomic concentration and chemical form of N-dopants. A capping layer of 0.7 m ·cm TaO x was deposited on top of the RSL, serving as oxygen reservoir. The transmission electron microscope (TEM, Titan Themis G2) image of the RRAM is shown in Fig. 1  composition of the HfO 2 layer was analyzed by X-ray photoelectron spectroscopy (XPS, AXIS Supra+). The electrical measurements were performed by Keysight B1500A for DC test and Speedcury ST2516 for pulse test. QuantumATK T-2022.03-SP1 [17] was used to perform molecular dynamics (MD) and density functional theory (DFT) calculations to further study the mechanism of N-doping. In practical applications, RRAM is operated by pulse rather than DC voltages. Therefore, we performed pulse test on the 1k RRAM array to obtain practical data of forming voltages. Here we utilized an incremental step pulse programming (ISPP) scheme [18] with a pulse width of 200 ns. The bit line (BL) voltage sweeps from 1 V to 6.4 V with a step of 0.1 V and the word line (WL) voltage sweeps from 0.7 V to 1.5 V with a step of 0.02 V. Since the thickness of RSL also affects the forming voltage [19], [20], RRAM devices with different RSL thicknesses were fabricated and tested for a comprehensive comparison. The results are shown in Fig. 2(c) for undoped devices and (d) for N-doped devices. The forming voltage is roughly in a linear relationship with the RSL thickness. From physics perspective, this is because the onset of forming process is determined by the electric field intensity in the RSL. The critical electric field of N-doped devices (2.9 MV/cm) is about 40% lower than that of undoped devices (4.7 MV/cm).
Although lower forming voltages can be easily obtained by decreasing the RSL thickness, it causes the degradation upon on/off ratio and endurance at the same time [7], [19], [20]. As the thickness of RSL decreases, fewer oxygen ions (O 2− ) are driven by electric field and stored at the oxygen reservoir during forming. During the reset process, the devices with thin RSL would have fewer O 2− to be driven back to rupture CF, which results in a lower on/off ratio. Furthermore, electric field and heat generated in the reset process would also drive some of the stored O 2− to migrate to regions outside the CF, which results in the loss of stored O 2− . During the repeated set and reset processes, RRAM with thinner RSL would run out of stored O 2− more quickly, resulting in degraded endurance. [21] Therefore, N-doping allows RRAM to have a thicker RSL under the same forming voltage without sacrificing on/off ratio and endurance, achieving an overall enhanced performance. Fig. 3 shows the comparison between undoped and N-doped devices on the tradeoff between forming voltage and on/off ratio ( Fig. 3(a)) as well as endurance ( Fig. 3(b)). The N-doped devices achieve comprehensive improvement on the trade-offs. At the forming voltage of 2 V, which is applicable for integration with 14 nm technology node according to [22], the N-dope devices achieve about 3× improvement on on/off ratio and 10× improvement on endurance. N-doped devices also show a relatively good retention. Fig. 4(a) compares the retention of N-doped and undoped devices. After being baked at 120 • C for 1 hour, 95% of N-doped devices maintained an on/off ratio larger than 5× between the high-and low-resistance states (HRS/LRS). This percentage is close to that of undoped devices (96.5%).
To further investigate the role N-dopants played in forming, XPS was performed to identify the compound N formed. The XPS results for the RSL of N-doped devices are shown in Fig. 4(b). A peak at 403.9 eV is observed, indicating N-dopants are in the form of nitrites ((NO 2 ) − ) rather than nitrogen ion (N 3− ). The atomic concentration of N is about 1.5%. It is speculated that (NO 2 ) − facilitate forming by absorbing oxygen and forming nitrates ((NO 3 ) − ) and oxygen vacancies (V O ) under electric field. The chemical equation is shown as follows: To verify this hypothesis, here we designed an annealing experiment, in which the RRAM devices were annealed at 300 • C for 15 minutes. The annealing promoted (NO 2 ) − in N-doped devices to be oxidized into (NO 3 ) − (408.3 eV peak), as shown in Fig. 4(b). After annealing, the forming voltage of N-doped devices increased to a level similar to undoped devices (Fig. 4(c)). We can see the enhancement effect on forming by N-doping diminished as (NO 2 ) − turned into (NO 3 ) − . These results prove that it is the (NO 2 ) − that played a vital role in forming. Although 300 • C is much higher than the maximum local temperature of RRAM array during device operation [23], the thermal stability of (NO 2 ) − to some extent still limit the subsequent manufacturing process. This is not a serious problem because RRAM is integrated through low-temperature back-end-of-line (BEOL) process.
On the other hand, N-doping increases the pre-forming leakage current of RRAM ( Fig. 4(d)), which also helps decrease the forming voltage. Larger leakage current indicates more leakage paths exist in the RSL of pre-forming N-doped devices. As a result, it is more likely to find a conductive path which is easier to form conductive filament during forming [25], [26]. The leakage current of N-doped devices decreased after annealing but was still higher than undoped devices (Fig. 4(d)), which in turn confirms that it is the combination of chemical process and leakage current that leads to the forming enhancement effect.
To further exploit the mechanism of N-doping, firstprinciples calculations were performed on a supercell of 146 atoms. The amorphous structure in Fig. 5(a) was generated by simulating melt-and-quench procedure in MD with a Morse potential set for HfO 2 [27]. The (NO 2 ) − dopant was established in accordance to the sp 2 hybridization of N. DFT simulation was carried out, with PseudoDojo pseudo-potentials [28] and generalized gradient approximation (GGA) based on Perdew-Burke-Ernzerhof (PBE) function [29], in order to extract the projected density of states (PDOS) for initial state (Fig. 5(b)) and final state (Fig. 5(c)) of chemical reaction (1). In the initial state, the (NO 2 ) − induced a deep impurity level in the energy bandgap. As the reaction generated a (NO 3 ) − and a V O , the deep impurity level transformed into a shallow level in the final state, which was contributed by Hf atoms  near the V O . The energy band diagram shown in Fig. 6 can be used to summarize the forming mechanism of N-doped devices. When the applied voltage is low, deep impurity level at the initial state induces multiple Poole-Frenkel emission and trap-assisted tunneling current paths, resulting in a large leakage current ( Fig. 6(a)). When the applied voltage is large enough, reaction (1) is triggered and conductive V O with a much lower barrier are formed, which then contribute to the growth of CF (Fig. 6(b)).

IV. CONCLUSION
In this work, we propose a novel N-doped HfO 2 RRAM with chemically enhanced forming. The N-doped RRAM has 40% lower critical forming electric field, which enables it to have a thicker RSL for the same forming voltage. It achieves 3× improvement of on/off ratio and 10× improvement of endurance at a forming voltage of 2 V that is applicable for integration with advanced nodes. The mechanism of N-doping is explained from both physical and chemical perspectives, which is validated by experiments and simulations. The N-doping method provides a new direction to improve the resistive switching characteristic of RRAM by inducing extra chemical process in forming.