Physics-informed learning for modeling infrasound propagation in extreme weather conditions

Extreme events in fluid flows are characterized by the coexistence of complex nonlinear dynamics, high intrinsic dimensionality and intermittency, which often results in spatially localized disturbances (turbulence spots, gravity wave breaking). Although many studies have shown that atmospheric ducting of infrasound is sensitive to these disturbances, yet the link between their statistical properties and that of the infrasound wavefield remains an open question, mainly because very little data are available for extreme events. The present work focuses on catastrophic events in climate systems where the amount of data available (typically a few decades) is not sufficient to extrapolate the PDFs. This class of problem involves geophysical fluid flows over climate scales where reanalysis data are a reliable source of information. In contrast to methods that rely on standard models to compute the PDFs from available data, the focus here is on data-driven methods that encode some information about the wave dynamics. The idea behind this approach is to combine two sources of information (reanalysis data and wave theory) using physics-informed neural networks to extrapolate the PDFs. The performance of this approach is illustrated around two types of events that affect infrasound propagation: sudden stratospheric warmings and mountain-induced extreme weathers.