Estimating maximum aftershock magnitudes using machine learning methods

Accurate forecasting of the maximum aftershock magnitude following a strong earthquake is essential for effective emergency response planning and risk assessment. Conventional forecasting approaches mainly emphasize the spatial and temporal patterns of aftershock sequences but overlook the physical properties of the mainshock, including the apparent stress and focal mechanism. In this study, we used earthquake catalogs with source parameters (e.g., the Global Centroid-Moment-Tensor and National Earthquake Information Center catalogs) to construct mainshock-aftershock sequences. The seismic data were initially cleaned to ensure that the seismic events had a uniform magnitude type. We then established maximum aftershock magnitude forecasting models using machine learning (ML) techniques such as Light Gradient Boosting Machine, Support Vector Machine, XGBoost, and Random Forest. To determine how the mainshock magnitude affects the forecasting of the largest aftershock magnitude, we used two training strategies: training with the entire mixed-magnitude dataset and training with the data stratified into magnitude intervals. The Bayesian-optimized XGBoost model produced the best results in the test set that used the entire mixed-magnitude dataset, with an accuracy of 46.45% and R² of 0.65. When applying the stratified training strategy, the relative advantage of the ML models increased with a decreasing prediction threshold error, reaching a maximum at ±0.2 magnitude units. The results indicate that ML techniques can capture the complex nonlinear correlations between maximum aftershock magnitude and the properties of the mainshock. Consequently, this study demonstrates the potential for using ML methods in aftershock forecasting, thereby offering valuable insights for future aftershock risk assessments.