AI & ML based Implementation of the Underwater Channel Model in the Tropical Littoral Waters of the Indian Ocean Region (IOR)

Acoustic waves are one of the predominant modes of signal propagation, in the underwater domain for numerous reasons. The Indian Ocean Region (IOR) is a strategic location and has been bustling with trade and military activities, particularly in the 21st century. However, the tropical littoral waters of the IOR have ensured sub-optimal performance of any sonar deployment in the region due to the high SOFAR channel depth, thus making the entire region acoustically shallow and ensuring multiple interactions of the acoustic propagations with the surface and the bottom boundaries. Additionally, random fluctuations in the surface roughness and bottom type further add to the complexity of propagation. Thus, underwater channel modeling is an interesting research problem with substantial applications across multiple stakeholders in the IOR. The rich biodiversity of the tropical littoral waters adds another dimension to the acoustic signal deterioration, as the random attenuation intensifies, even for the direct path propagation in the volume of the water body. It is in order that we map the underwater medium parameters like wind, speed, surface temperature, salinity, sound speed profile, bottom type, bottom profile, and more. The Modelling & Simulation (M&S) effort to map the underwater channel across all seasonal and diurnal variations, is a critical step to contain the impact on acoustic propagation. The real-time prediction of the underwater channel behavior through this M&S effort will facilitate mitigation of the underwater channel impact.The current methods of acoustical channel modeling have been mathematical, deriving differential equations from the fundamental equations of state, with different assumptions that lead to different models of use, limited to different regions. Due to the variability of the IOR waters, a range-dependent acoustic model is most suitable for the region, the downside being the high compute times and specialized hardware required, rendering these approaches useless for real-time calculations. We propose a data-centric approach based on the current state-of-the-art ML algorithms, to learn from simulation data. Recent Deep Learning algorithms have proven really successful in learning distributions conditioned on the input, making them universal function approximators. Considering the fact that the acoustic channel model can be derived from the fundamental equations of state, but the assumptions being made by the human solver render them not useful for certain regions. Hence, an ML algorithm, in theory, shall be able to learn the approximate function, and the assumptions to be made thereof. We let an Artificial Deep Neural Network learn from data generated from the range acoustic model, and compare the two approaches based on resource and error metrics.The result metrics we use compare the running times of the state-of-the-art model for range-dependent media in the low-frequency regime, indicating a significant decrease in computation runtimes. Furthermore, computing the errors of the proposed method indicates that just below 6% of the errors are greater than 10dB. Given such results, we conclude that a Deep Neural network can serve as a viable real-time replacement tool for the range-dependent acoustic channel model. This is currently limited to just the IOR but can be further adapted to worldwide data, thus forming a universal data-centric acoustic modeling technique for real-time applications.