Nov. 19, 2019,
3:30 p.m. - 4:30 p.m.
Earthquakes don’t happen on a schedule, so seismologists record continuous ground motion data, typically at 100 samples per second, 365 days a year. Increasingly capable and affordable sensors have led to a dramatic increase in the volume of continuous seismic data being recorded. Seismologists have to sift through those massive data sets to find not only the large, infrequent earthquakes, but also the much more numerous, but still important, small earthquakes. This need to extract as much information as we can from large data sets motivates a new generation of more efficient, robust, and scalable earthquake monitoring approaches based on machine learning techniques. I will present details of deep learning tools we have developed for earthquake signal detection, denoising, discrimination, and earthquake characterization. These algorithms employ different training strategies from unsupervised to supervised approaches using real or semi-synthetic data, but they are all implementations of deep neural networks. We use different network architectures based on the character of the problem to be solved. These include convolutional networks, convolutional-recurrent networks, and autoencoders. I will present examples that demonstrate the power of these approaches for developing dramatically improved earthquake catalogs and for providing new insight into earthquake processes.