Unsupervised Learning and Data-Driven Discovery: Detection, Classification, Forecasting, Separation

GMIG makes use of supervised and unsupervised machine learning to analyze seismic data, which is currently collected around the world at a rate that exceeds the capacity for human analysis. GMIG’s machine-learning approach not only analyzes greater quantities of data more quickly, but also addresses the biases of standard seismological models. The group’s seismology programs focus on polyphonic detection, segmentation, classification and separation (denoising).

So far, the program has built a learnable scattering network, a recurrent scattering neural network, and a variational recurrent scattering autoencoder, all of which use attention mechanisms. With data from multiple sensors, association in an unknown environment is realized through deep consensus. GMIG’s models use GMMs, fGMMs and hierarchical clustering to classify seismic data. GMIG has worked on blind detection and identification of precursors, glitches and noise, with the goal of denoising seismic data.

GMIG’s models for causal inference identify and characterize tremors by combining seismic data with GPS data. The group’s seismic monitoring is driven implicitly by features, while its forecasting uses occurrence rates of selected clusters, drift in epistemic uncertainty, and other indicators.

The deep learning approaches in this program apply not only to traditional seismic sensors and general land acquisition, but also to DAS. The learning of noise characteristics can also be broadly applied to GPS data, creating opportunities for corporate partnerships and collaborations in the growing field of geographic information systems.

Further reading: Clustering earthquake signals and background noises in continuous seismic data with unsupervised deep learning