Machine Learning Predicts the Slip Duration and Friction Drop of Laboratory Earthquakes in Sheared Granular Fault

• Published in: Journal of geophysical research. Machine learning and computation Vol. 1; no. 4
• Main Authors: Wei, Mengfan, Gao, Ke
• Format: Journal Article
• Language: English
• Published: Wiley 01.12.2024
• ISSN: 2993-5210; 2993-5210
• DOI: 10.1029/2024JH000398

Predicting laboratory earthquakes using machine learning has progressed markedly recently. Previous related studies mainly focus on predicting the occurrence time and shear stress of laboratory earthquakes using acoustic emission signals. Here, based on numerical simulations, we use machine learning to show that statistical features of plate motion signals contain information about the slip duration and friction drop of laboratory earthquakes. We find that the plate motion signals during the initial slip stage contain the precursor information about the slip duration of laboratory earthquakes. However, to accurately predict the friction drop, we need to incorporate the plate motion signals during the entire slip stage. The results demonstrate that the high‐order moment and variance of plate motion signals are respectively among the best predictors for the slip duration and friction drop of laboratory earthquakes. Our work provides new insights for future investigations into natural earthquake prediction through machine learning. Plain Language Summary Earthquake prediction is an important but challenging task. The booming of machine learning brings new hope for earthquake prediction. However, the direct utilization of machine learning for earthquake prediction is still tricky, simply because the available data our human beings currently have can only cover a limited number of earthquake cycles. In this study, we use numerical simulations to represent earthquakes as frictional slips and record hundreds of simulated laboratory earthquake cycles alongside their corresponding fault motion data. We then train the motion data using machine learning to predict the slip duration and friction drop of upcoming laboratory earthquakes. The results show that the slow pre‐seismic slip stage, similar to that observed in natural earthquakes, has the potential to indicate seismic nucleation. Particularly, we find that the fault motion signals during the initial slip stage contain the precursor information about the slip duration of laboratory earthquakes. Our findings may contribute to predicting the slip duration and friction drop of natural earthquakes. Key Points The prediction of slip duration and friction drop of laboratory earthquakes are explored using ensemble learning The fourth‐order moment of the plate motion signals during the initial slip stage is among the best estimators of slip duration The second moment of the plate motion signals during the entire slip stage is the best predictive feature for friction drop