A Hybrid Framework Combining Physical and Data‐Driven Models for Predicting the Timing of Fault Instability

Reliable prediction of earthquakes and rock mass failure has long been hampered by a fundamental trade‐off: achieving both a long warning window and high prediction accuracy. Physical models, such as the failure forecast method (FFM) based on the time‐reversed Omori law, can provide longer warning windows but are sensitive to data noise and exhibit limited accuracy. In contrast, data‐driven models like the long short‐term memory (LSTM) network can achieve higher accuracy but typically offer short warning times and suffer from limited generalization. To overcome this bottleneck, this study develops a hybrid prediction framework that integrates physical mechanisms with data‐driven methods. Specifically, we propose a “half‐length subset” algorithm to improve FFM (MFFM), which dynamically screens precursor data to stabilize the power‐law fitting process. Concurrently, we introduce the Grey Wolf Optimizer (GWO) to automatically tune the hyperparameters of LSTM (GWO‐LSTM), enhancing its robustness and generalization capability. The proposed MFFM and GWO‐LSTM models were validated using two typical fault failure patterns simulated in laboratory uniaxial loading experiments, corresponding to the “foreshock‐mainshock” and “mainshock‐aftershock” sequences observed in natural earthquakes. Results demonstrate that MFFM and GWO‐LSTM significantly outperform their traditional counterparts. MFFM reduces the maximum root mean square error (RMSE) of FFM by up to 85.6%, while GWO‐LSTM further reduces the RMSE of LSTM by up to 98.7%, while also effectively extending the warning time. The proposed hybrid framework offers a reliable new approach that combines physical robustness with data adaptability, enabling high‐precision, long‐term failure time prediction.