Neural network predictions of acoustic transmission loss uncertainty

Abstract

Many applications rely on predictions of acoustic transmission loss (TL) computed with numerical tools given expected environmental properties and potential circumstances. Because environmental uncertainty limits the quality of these TL predictions, methods to quantify the corresponding TL uncertainty are critical for robust decision-making and uncertainty reduction. Given the variety of needs across relevant applications, a desirable TL-uncertainty quantification method should be compatible with realistic (database-precise) descriptions of environments and uncertainties, computationally non-intrusive for the many possible Helmholtz equation solvers, and fast enough for real-time applications. Supervised machine learning provides such a method where a neural network (NN) is trained to relate TL uncertainty to spatial patterns in the existing nominal TL-field solution across a dataset of examples generated from Monte Carlo simulations. While this method has been successfully implemented for long-range, underwater propagation simulated with a 2D Parabolic Equation solver, this study will assess its capability on different computational models, environmental scenarios, and sources and degrees of uncertainties. Examples will be given, which demonstrate how data-driven learning, physical understanding, and hyperparameter optimization can be used to design the key features of this method: the NN inputs, the NN architecture, and the training dataset(s). [Work supported by the NDSEG fellowship program]