Deep Quake Dynamics: A Multimodal Fault-Aware Approach to Earthquake Magnitude and Occurrence Time Forecasting

Abstract

Earthquake (EQ) forecasting is critical for early warning systems. We propose an EQ Magnitude-Time (EQMT) Forecasting Model that integrates real-world inter-fault relations, multiple data sources, and advanced modelling techniques to forecast EQ magnitudes and classify occurrence times. The model converts fault lines into graph representations and extracts node embeddings to capture spatial inter-fault dependencies. We evaluated EQMT on seismic events in the northern Aegean region, where the North Anatolian Fault intersects the Aegean extensional system. The proposed model attains a mean absolute error of 0.33 for magnitude forecasting and an F1 score of 81.39% for occurrence-time classification. Furthermore, to enhance the interpretability of both magnitude and occurrence time forecasting tasks, we applied SHapley Additive exPlanations (SHAP)-based explainable Artificial Intelligence (XAI) analyses, which provide insights into the contribution of individual seismic and fault-related features to the model's forecast. SHAP-based explainability analysis reveals that observation-duration features and graph embeddings are the primary contributors to model predictions. These results demonstrate the potential of fault-aware, graph-based approaches for operational EQ forecasting.