Exciting news! We're transitioning to the Statewide California Earthquake Center. Our new website is under construction, but we'll continue using this website for SCEC business in the meantime. We're also archiving the Southern Center site to preserve its rich history. A new and improved platform is coming soon!

2018 USEIT: Using Machine Learning to Forecast Earthquakes

Anthony A. Guerra, Brandon T. Ho, Varduhi Kababjyan, Ramon Mei, Tomoe Mizutani, Tiffany Streitenberger, Shalani Weerasooriya, Jordan Wolz, Guillermo Beas, Shengyu Wang, Abhijit Kashyap, & Jacquelyn J. Gilchrist

Published August 21, 2018, SCEC Contribution #8843, 2018 SCEC Annual Meeting Poster #311

As part of the 2018 Undergraduate Studies in Earthquake Information Technology (UseIT) internship program, the Machine Learning (ML) team was challenged to investigate how ML might be used to derive statistical earthquake forecasts from deterministic earthquake simulations. The Rate-State Earthquake Simulator (RSQSim) is a physics-based faults system model that generates earthquake catalogs. The team filtered a one million year California statewide catalog produced by RSQSim using the Uniform California Earthquake Rupture Forecast version 3 (UCERF3) fault model for M≥7 events, which were then binned into in ten year intervals on four selected fault sections: Cholame, Carrizo, San Bernardino North, and Mojave South. Various ML algorithms were tested on this data to forecast ruptures, including a recurrent neural network, classification algorithms, and regression. Using a recurrent neural network we took into account the temporal and cyclic nature of earthquake ruptures, considering a series of time internal snapshots as one input. For the classification method, we took the rupture time intervals and classified them based on whether a rupture occurred, applying several algorithms including: logistic regression, decision trees, linear discriminant analysis, K-nearest neighbors, naive Bayes classifier, and then compared their accuracies and reliabilities in predicting the occurrence of ruptures. Further, we compared the fault rupture probabilities between RSQSim and the (UCERF3) transition probabilities that were obtained through regression on 4 different fault sections. We created a risk gradient fault labeling system based on recurrence intervals, which improved the R-squared scores for our classification algorithms. Future work should investigate adjustments to algorithms and rupture data processing, continuing to build on the potential for machine learning in the realm of earthquake rupture forecast in the future. With this increase we conclude that with more data, machine learning techniques are efficient tools that can help forecast earthquake activity by detecting patterns in data that humans are unable to recognize.

Guerra, A. A., Ho, B. T., Kababjyan, V., Mei, R., Mizutani, T., Streitenberger, T., Weerasooriya, S., Wolz, J., Beas, G., Wang, S., Kashyap, A., & Gilchrist, J. J. (2018, 08). 2018 USEIT: Using Machine Learning to Forecast Earthquakes. Poster Presentation at 2018 SCEC Annual Meeting.

Related Projects & Working Groups
Communication, Education, and Outreach (CEO)