[M-IS09] Interdisciplinary studies on pre-earthquake processes

[MIS09-01] Overview of Nowcasting Earthquakes with QuakeGPT: Methods, First Results, Future Directions

★Invited Papers

*John B Rundle¹ (1.University of California Davis)

キーワード：Earthquakes, Nowcasting, Machine Learning, Artificial Intelligence, Forecasting

On August 6, 2024 a magnitude M5.2 earthquake occurred 24 km southwest of Lamont, CA, north of the intersection of the Garlock fault with the San Andreas fault in the area of the big bend and the transverse ranges. Then, on August 12, 2024, a magnitude M4.4 earthquake occurred 4 km south-southeast of Highland Park, CA, within the Los Angeles basin. Aftershocks continue from both events. These events underscore the urgency of evaluating whether events such as these may signal the onset of a major earthquake on the San Andreas or Garlock faults. Earthquake nowcasting is a relatively new method that employs a simple 2-parameter filter on the observed monthly seismic rate of small earthquakes, and uses machine learning to improve the nowcast filter, and these methods are being actively developed in a variety of new ongoing research projects, which we call "QuakeGPT". This new idea uses stochastic earthquake simulations to train a science transformer model, an attention-based technology similar to that which underlies ChatGPT. As the previous earthquake nowcasting idea generates earthquake probability by optimization using Receiver Operating Characteristic (ROC) methods, the goal of QuakeGPT is to determine with greater accuracy when and where a major earthquake would occur. With respect to the existing earthquake nowcasting technology using ROC machine learning, we have computed examples of videos that illustrate the advantages of following the change in earthquake risk through time. In the southern California video, the recent earthquakes in southern California are coincident with an enhanced spatial and temporal probability of a significant earthquake (M³6.75) in southern California. In the Southern California area, the spatial probability density is in the vicinity of the Lamont earthquake, and generally lies along the Garlock fault, stretching from the epicentral region of the 2019 Ridgecrest earthquake, towards the junction of the Garlock fault with the San Andreas. In the California video, the enhanced probability is not as obvious. We plan to closely monitor the incoming data with these methods, and continue to evaluate the results for their importance and significance.
Videos: https://rundle.physics.ucdavis.edu/Movies/SouthernCalifornia.mp4
https://rundle.physics.ucdavis.edu/Movies/California.mp4
https://rundle.physics.ucdavis.edu/LA_One_Hot_M6.0.mp4