Data-driven interpretable models of wave dynamics for infrasound monitoring

Abstract

Accurate and efficient models are essential to understand and predict acoustic signals propagating in the atmosphere. In many practical problems such as source localization and yield estimation, multiple models for describing the atmospheric fluctuations and infrasound are available, with varying approximation qualities. In standard practice, inferences are exercised as if the selected models had generated the observations. This approach ignores the model uncertainty, leading to biased inferences and to estimates that may be extremely sensitive to tunable parameters. This work explores a hierarchical Bayesian framework for producing interpretable models for both atmospheric gravity wave (GW) dynamics and acoustic propagation, from ground-based infrasound measurements. It is shown that there are only a few important terms that govern the GW dynamics and the interactions with infrasound. The resulting GW models can either be incorporated into global climate models to better describe the effects of GWs on the global circulation or used together with infrasound propagation models for improving inference accuracy and efficiency. This perspective, combining the resulting infrasound-driven models with sparse sensing and machine learning to monitor the atmosphere, is explored using recurring events such as the ammunition destruction explosions at Hukkakero, in northern Finland.