Machine-learning models for the prediction of long-range outdoor sound propagation

Abstract

Long-range outdoor sound propagation is characterized by a large variance in sound pressure levels due to factors such as refractive gradients, turbulence, and topographic variations. While conventional numerical methods for long-range propagation address these phenomena, they are costly in computational memory and time. In contrast, machine-learning algorithms provide very fast predictions, which this study considers. Observations from either experimental data, or surrogate data from a numerical method, are required for the training of machine-learning models. In this study, a comprehensive training set for the machine learning was created from excess attenuation predictions made with a Crank-Nicholson parabolic equation (CNPE) model. Latin hypercube sampling of the parameter space (source frequency, meteorological factors, boundary conditions, and propagation geometries) generates a set of input for the CNPE model and machine-learning models. Consideration is given to ensemble decision trees, ensemble neural networks, and cluster-weighted models for nonlinear regression. The large variance in excess attenuation, from the CNPE model, presents a challenge for accurate machine-learning model predictions. For example, given 5000 samples the overall root-mean-square error for an ensemble decision tree model is 6.7 dB. Errors related to sample size, modeling approaches, and propagation ranges are quantified in this study.

mSOUND: An Open Source Toolbox for Modeling Acoustic Wave Propagation in Heterogeneous Media

“…mSOUND can be used as a stand-alone toolbox or in conjunction with other ultrasound toolboxes. We envision that this new toolbox can assist ultrasound researchers to tackle a variety of problems: to study the phase aberration in tissue [47], imaging reconstruction in PAT, ultrasound waveform tomography [48], data generation for machine learning [49], among others. In the future, we would also like to further improve mSOUND by leveraging the high computational capability of graphic processing units (GPUs) to deliver fast, accurate, and versatile ultrasound modeling solutions.…”