ERL FISH: Paul Johnson (LANL)

Friday, October 14, 2022 at 12:00pm

Building 54, Room 209
21 AMES ST, Cambridge, MA 02139


Learning fault-slip behaviors

We analyze large amounts of continuous seismic data applying machine learning, with the goal of identifying hidden signals connected to earthquake cycles. Our goal is to learn more about earthquake fault processes, and advance earthquake hazards assessment. In the laboratory, we find that continuous seismic waves are imprinted with fundamental information regarding the state of faults. Statistics of low-amplitude, noise-like signals can be used to estimate earthquake fault friction, displacement, and predict upcoming failure with great accuracy. These results hold true for a broad spectrum of laboratory slip behaviors. Our first application to Earth data in the Cascadia subduction zone showed that a similar approach can be used to infer slow slip fault displacement (estimated by surface GPS) from continuous seismic waves. We then applied a similar approach to the Parkfield segment of the San Andreas Fault to determine of the approach could be effective in quantifying seismogenic faults. In this work we attempted to determine if there was information contained in low frequency earthquakes (LFE’s) broadcast from the lower portion of the fault, regarding timing of small earthquakes located in the seismogenic portion of the fault. This effort was not successful. We speculate this was due to the very low signal amplitudes emitted and the significant noise levels in the recordings. We are currently developing alternative approaches to address seismogenic faults in Earth. The first approach is based on applying transfer learning using fault simulations generated with a combined finite-element and discrete element fault model, where synthetic ‘acoustic emissions’ (AE) are used as input for deep learning model training. The model label is predicting frictional behaviors such as friction and earthquake timing on the actual fault. This approach was successful using fault simulation AE as model input and laboratory friction data as label. A second approach is based on including a physical model of friction in the deep learning model, in a procedure known as the ‘physics of informed learning’. These efforts are in progress. 


About this Series

About this series: The Earth Resources Laboratory (ERL) Friday Informal Seminar Hour presents guest speakers on geophysics, seismology, inversion, imaging, ML, and the energy industry. Contact for more information and Zoom password.


Event Type


Events By Interest


Events By Audience

MIT Community

Events By School

School of Science



Department of Earth, Atmospheric and Planetary Sciences
Contact Email

Add to my calendar

Recent Activity