Search

Forecasting earthquake magnitude and epicenter is of great significance for disaster management and hazard mitigation. Existing machine learning (ML)-based earthquake forecasting has faced two shortcomings: limited historical earthquake samples for training and lack of explicit consideration of seismic prior knowledge. We propose a novel spatiotemporal prior-informed deep network (STPiDN) that incorporates seismic spatiotemporal prior knowledge into deep neural networks using limited historical earthquake samples. In our method, a physics-informed recurrent graph network (PRGN) is developed to extract representations of observed earthquake precursor data in a physics-informed manner. In order to make prior-guided earthquake predictions, we develop seismic prior knowledge activation layers that combine earthquake event representations with prior knowledge (e.g., fault distribution) through activation gates. An adaptive multitask loss function is proposed to achieve joint magnitude and epicenter forecasting with the consideration of alleviating the magnitude and epicenter imbalance problem. Our empirical evaluation results show that the proposed forecasting method outperforms several competing methods on a real-world earthquake dataset, proving that physics-informed prediction methods have the potential to capture complex earthquake patterns using limited training samples.