Predicting infrasound transmission loss using deep learning

SUMMARY Modelling the spatial distribution of infrasound attenuation (or transmission loss, TL) is key to understanding and interpreting microbarometer data observations. Such predictions enable reliable assessment source characteristics such as ground pressure levels associated with earthquakes, man-made or volcanic explosion properties, ocean-generated microbarom wavefields. However, computational cost inherent in full-waveform modelling tools, parabolic equation (PE) codes, often prevents exploration a large parameter space, that variations wind models, frequency location, when deriving estimates atmospheric properties—in particular for real-time near-real-time applications. Therefore, many studies rely on analytical regression-based heuristic TL equations neglect complex vertical range-dependent variation properties. This introduces significant uncertainties predicted TL. In current contribution, we propose deep learning approach trained set simulated wavefields generated using PE simulations realistic winds predict ground-level amplitudes up 1000 km from ground-based source. Realistic range dependent are constructed by combining ERA5, NRLMSISE-00 HWM-14 small-scale gravity-wave perturbations computed Gardner model. Given profiles input, our new framework provides fast (0.05 s runtime) (?5 dB error average, compared simulations) estimate