We introduce a machine learning (ML) model that reconstructs atmospheric conditions at the location of an explosive source using regionally recorded synthetic infrasound. A convolutional neural network (CNN) processes full waveforms and source-receiver geometries to estimate the vertical profile of effective sound speed. The model is trained on synthetic data generated by propagating a source time function through modeled atmospheres out to random station locations. Prediction accuracy is high with an average root mean squared error (RMSE) of 8 m/s. We find that some poorly predicted profiles nonetheless reproduce the original waveforms with high accuracy, indicating a nonuniqueness between atmospheric conditions and recorded waveforms. Unlike traditional physics-based inversions that rely on array deployments, our ML model can analyze single-channel data and, once trained, makes predictions within milliseconds. Given the performance on synthetic data, future tests should include real-world data to further assess the benefits of ML-based frameworks for infrasound analysis.
