At the 10th International Summer School on AI and Big Data, Dr. Christopher William Johnson will talk about Applications of AI to Earthquake Physics: Learning Fault Slip and Precursors to failure to Advance Earthquake Prediction.

Talk: Applications of AI to Earthquake Physics: Learning Fault Slip and Precursors to failure to Advance Earthquake Prediction

Establishing models to predict when, where, and how a large earthquake may rupture has been considered an unfathomable goal for several reasons. Damaging earthquakes occur on existing fault systems with important spatial scales in the range 10^-3-10³ kilometers and repeat times of order 10^-3-10³ years. The physics of seismogenesis includes stress accumulation on a fault, mechanical weakening/strengthening of faulted materials under different loading conditions, and the release of energy during an earthquake.

The primary observations are continuous time series of seismic waveforms and ground-based measurements from global navigation satellite systems (GNSS). The spatial coverage of monitoring networks is exceedingly good in some regions, e.g., Japan, Alaska, and southern California, but sparse in many other regions around the world. Since earthquake repeat times are long, these data are usually temporally limited to a small fraction of the interevent loading cycle. A systematic pattern of precursory signals has yet to be discovered, i.e., large earthquakes occur with no advanced warning.

Considerable insight has been gleaned from applying machine learning to laboratory earthquake experimental data, where multiple earthquake cycles at the laboratory scale can be generated. This work shows that it is possible to predict the time of future events (i.e., time to failure), the fault zone stress state, and many other earthquake characteristics using lab fault zone microseismicity and the evolution of fault zone elastic properties. Recently in the laboratory, better results have been obtained by leveraging machine learning techniques such as transformer networks and other encoder-decoder techniques, in combination with transfer learning and physics informed modeling.

When applied to Earth data containing multiple earthquake cycles, the models scale well for the cases that have been studied, including slowly-slipping non-damaging earthquakes. Laboratory data studies point to the importance of faint deformation signals and complex patterns, emitted from slipping and/or locked faults preceding large earthquakes. Knowledge from the laboratory experiments and the exponential availability of seismic data recordings from anywhere on Earth is motivating us towards seismic foundation models and a novel stage of earthquake understanding.

Bio

Dr. Christopher W. Johnson is an early career Los Alamos National Laboratory staff scientist and former Distinguished Feynman Postdoc fellow. He completed a B.S. at Georgia Institute of Technology and a Ph.D at University of California Berkeley. Prior to joining LANL he was a NSF Postdoc fellow at University of California San Diego. Johnson is a trained geophysicist that specializes in seismology and geodesy measurements with a focus on surface deformation, seismic signal detection, and crustal stress changes. His work develops machine learning applications for physics constrained models applied to continuous time series of geophysical data. Johnson is regularly invited to present research at national and international venues (AGU, JPGU, Earthscope) and has published more than 40 peer reviewed manuscripts.