Earthquake Forecasting and Major Event Detection Using Multi-Task Deep Learning in Central Asia

This manuscript presents a deep learning approach for earthquake magnitude forecasting and major event detection (M\((\geq)\)5) in Central Asia, with a focus on Kyrgyzstan. A Multi-Task Learning (MTL) model with an attention-based Long-Short-Term Memory (LSTM) architecture is used to perform two tasks: (i) regression to forecast earthquake magnitude, and (ii) binary classification to determine if a big event will happen in the next 30 days. Input features are spatiotemporal, fault proximity, rolling seismic statistics and time-based indicators. Class imbalance in the binary task was handled by oversampling. Hyperparameter tuning was done using Optuna and the model was evaluated using Mean Absolute Error (MAE), F1-score, ROC AUC and PR AUC. The final model achieved MAE of 0.154 for magnitude forecasting and F1-score of 0.989 for big event detection, with ROC AUC and PR AUC of 0.993 and 0.9996 respectively. The approach was validated on historical earthquake data and tested on a real case study. The model is reliable and applicable. The proposed method is good for regional seismic risk forecasting and can be used for early warning systems and public safety in seismically active regions like Kyrgyzstan.