Predicting earthquakes has long been an unceasing exploration in geoscience. Recently, machine learning (ML) has been tried to predict laboratory slip events based on the stick-slip dynamics data obtained from laboratory shear experiments, with the ultimate goal of seeking appropriate approaches and procedures for natural earthquake prediction. However, the data utilized in existing work are generally small, i.e., acquired from only single or a few sensor points. Here, by employing the combined finite-discrete element method (FDEM), we explicitly simulate a two-dimensional sheared granular fault system, and place 2203 densely distributed “sensor” points inside the model to collect abundant fault dynamics data such as displacement and velocity during the stick-slip cycles. We use LightGBM to train the data and predict the normalized gouge-plate shear stress (i.e., the indicator of stick-slips). During the training, we build the importance ranking of input features, and select the ones with top importance to prediction as optimized features. We gradually optimize and adjust the input feature data, and finally reach a LightGBM model with acceptable prediction accuracy (R² = 0.91). The SHAP (SHapley Additive exPlanations) values of input features are also calculated to quantify their contributions to the prediction results. The ML analyses demonstrate that the large amounts of fault dynamics data contain the necessary information for predicting upcoming slip events; however, they may be redundant and thus should be optimized to improve prediction performance. The LightGBM together with the SHAP value approach could not only accurately predict the occurrence time and magnitude of laboratory earthquakes, but also have 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 also provides a possible way to explore useful precursors for earthquake prediction using ML approaches.
 

实验室地震,机器学习