Infrasound propagation with deep neural operators

Wave propagation modeling has played a crucial role in various applications of infrasound. However, despite significant progress in recent years, solving the acoustic wave equation numerically still presents practical challenges, particularly for randomly layered media. On the other hand, while machine learning has emerged as a promising alternative, training deep neural networks requires a tremendous amount of data, which can be challenging and expensive to obtain. In this work, we combine a projection-based reduced-order model (ROM) of the wave equation with a Fourier neural operator (FNO) to learn mappings between atmospheric specifications and ground-based waveforms. The ROM is utilized to generate a comprehensive database of waveforms, taking the ECMWF ensemble reanalysis as input. Unlike traditional neural networks that are restricted to the prediction of solutions in a predefined configuration, it is shown that the FNO captures the leading order propagation operator as well as important properties of the dispersion relation. It is also demonstrated that the FNO predicts subtle effects in the waveforms that can be unambiguously associated with small-scale heterogeneities, such as turbulence and gravity waves. The feasibility of this approach is illustrated using repetitive infrasound events over the last 20 years.