Speaker
Description
The International Monitoring System (IMS) infrasound network is well positioned to record atmospheric nuclear explosions, but algorithmically classifying and characterizing these events is challenging. Difficulties stem from the variable and dynamic atmosphere that modulates acoustic transients at regional to global distances. Machine learning (ML) is well suited to classify infrasound activity but generally requires large training datasets. However, data from the relatively few large chemical explosions and sparse global infrasound network are insufficient to train a ML model given how complex and dynamic the atmosphere is at global scales. Instead, we propose a physics-based data augmentation method to produce an entirely synthetic training dataset. Realistic source time functions are generated and propagated through modeled atmospheres out to several hundred kilometers, thus producing a catalog of synthetic events. These data are then used to train a time convolutional neural network (TCN) that classifies explosions and background noise. We show the TCN not only identifies modeled events but is also effective at detecting and characterizing real world explosive transients, including those from the Humming Roadrunner experiments.
This work was supported by the Nuclear Arms Control Technology (NACT) Program at Defense Threat Reduction Agency (DTRA). Approved for public release; Distribution is unlimited.
Promotional text
We present a new infrasound based method to detect and classify nuclear blasts using machine learning. This approach will help elevate the usefulness of global infrasound deployments as a tool to monitor for atmospheric nuclear activity.
