Comparing analytical and machine-learning techniques for predicting acoustic fields in ideal single- and multi-path environments

Abstract

Neural-network-based machine learning provides a modern solution paradigm for a wide variety of problems. Trained neural networks can provide answers quickly because the primary computational effort is front-loaded into network training. This presentation provides the results of an investigation into the neural-network-development effort necessary to replace point-source Helmholtz-equation solutions in axisymmetric single-path, two-path, and multi-path environments having constant sound speed. Here, the acoustic and environmental input parameters are provided to the neural network as they would be to a propagation simulation. The single-path (free space) environment involves four parameters (sound speed, frequency, receiver range, and receiver depth). The two-path (Lloyd’s mirror) environment adds a reflecting surface and a fifth parameter, source depth. The multipath (ideal waveguide) environment further adds a second (deeper) reflecting surface and a sixth parameter, waveguide depth. In all cases, neural network training data and performance comparisons are developed from well-known analytical Helmholtz-equation solutions. Uniform and non-uniform sampling strategies for neural-netwok training are considered for frequencies in the 100s of Hz, depths up to 200 m, and ranges up to 2 km for sound speeds near 1500 m/s. Comparisons emphasize acoustic-field amplitude. Extensions of these results to acoustic-field phase and more than three propagation paths are considered.