Deep learning methods for modeling infrasound transmission loss in the middle atmosphere

Accurate modeling of infrasound transmission losses (TLs) is essential to assess the detection thresholds of the global International Monitoring System (IMS) infrasound network, quantify their spatial and temporal variations, and refine interpretations of signals generated by events of interest. Among the existing tools, the method based on parabolic equations resolution (PEs) enables TLs to be modeled finely, but its computational cost does not currently allow exploration of a large parameter space for real-time prediction, making it inapplicable for operational monitoring applications in the framework of the Comprehensive Test Ban Treaty (CTBT).

To reduce computation times, Brissaud et al. (2022) explored the potential of convolutional neural networks (CNNs) trained on a large set of regionally simulated wavefields (>1000 km distance from the source) to predict TLs with an error of 5 dB compared to PE simulations with negligible computation times ( 0.05 s). However, this new method shows both larger errors in upwind conditions, especially at low frequencies, and causal issues with winds at large distances from the source affecting ground TLs close to the source.

To reduce prediction errors, we introduce a new Deep Learning method, seeking to predict TLs from globally simulated effective sound velocity fields (>4000 km distance), based on a Convolutional Recurrent Neural Network (CRNN) capable of accounting the sequentiality of the propagation phenomenon. This tool can be used to compute global detectability maps of infrasound events in an operational context.