Predicting earthquakes has been a long-standing challenge. Recently, machine learning (ML) approaches have been employed to predict laboratory earthquakes using stick-slip dynamics data obtained from shear experiments. However, the data utilized are often acquired from only a few sensor points, thus insufficient in feature dimension and may limit the predictive power of ML. To address this issue, we adopt the combined finite-discrete element method (FDEM) to simulate a two-dimensional sheared granular fault system, from which abundant fault dynamics data (i.e., displacement and velocity) during stick-slip cycles are collected at 2203 “sensor” points densely placed in the numerical model. We then use the simulated data to train the LightGBM (Light Gradient Boosting Machine) models and predict the normalized gouge-plate shear stress (an indicator of stick-slips). Meanwhile, to optimize features, we build the importance ranking of input features and select those with top importance for prediction. We iteratively optimize and adjust the feature data, and finally reach a LightGBM model with an acceptable prediction accuracy (R² = 0.91). The SHAP (SHapley Additive exPlanations) values of input features are also calculated to quantify their contributions to prediction. We show that when sufficient fault dynamics data are available, LightGBM, together with the SHAP value approach, is capable of accurately predicting the occurrence time and magnitude of laboratory earthquakes, and also has the potential to uncover the relationship between microscopic fault dynamics and macroscopic stick-slip behaviors. This work may shed light on natural earthquake prediction and open new possibilities to explore useful earthquake precursors using ML.

断层,剪切黏滑,机器学习,预测