Deep learning-based method for near-real time estimation of infrasound transmission losses in theatmosphere

Accurately modeling transmission loss is essential for a variety of applications, such as
improving atmospheric data assimilation for numerical weather prediction, assessing attenuation
maps of sources of interest, or estimating detection capabilities of the International Monitoring
System infrasound network. However, the high computational cost of numerical modeling solvers
makes them impractical for a near-real-time analysis. To address this, a previous study trained a
Convolutional Neural Network on regional wind fields, predicting transmission losses in less than 0,05
seconds with a mean-squared error of 5 dB. However, this approach uses interpolated atmospheric
specifications and focused only on winds, limiting its applicability for long-range modeling. In this
work, we develop a convolutional recurrent network to predict transmission losses leveraging
realistic, range-dependent atmospheric specifications combining horizontal winds and temperatures,
including small-scale perturbations. The resulting model reaches an error of 4 dB while extending
propagation range up to 4,000 km and providing epistemic and data uncertainty estimates. First
studies of such an algorithm on regional scaled events (Tonga-Hunga eruption, Hukkakero explosions,
etc.) were performed to further evaluate the model. Predicted attenuation are compared with
results from an alternative regionally fine-tuned neural network. The model also demonstrated its
ability to adapt to new frequencies.