Machine learning for point-to-point transmission loss estimates in ocean acoustic waveguides

Abstract

Underwater acoustic system performance depends on several complex and dynamic environmental parameters, and simulating such performance is vital to the success of development and implementation of these systems. Because of the complexity of the environment and governing physical equations, realistic simulations can become computationally prohibitive. This is especially true of for large environments with many active systems being assessed. By utilizing convolutional neural networks (CNNs) trained on data generated by well-established physics based models (such as BELLHOP’s ray tracing algorithm), network predictions can be used lieu of physics-based models to significantly reduce the computational burden in the loop for system performance simulations. In this paper, the usefulness and limitations of using CNNs to estimate transmission loss (TL), which is a key element in determining system performance, is explored. Using BELLHOP’s ray tracing algorithm as a baseline, CNN’s were able to produce TL results with significantly lower errors than those estimates made using other estimation methods such as spherical spreading and K-nearest neighbors. This indicates that the computational costs of large underwater acoustic simulations may be shifted from inside the simulation to network training, thus allowing for more efficient traditional and Monte Carlo style simulations.