SEISHYBRIDNET: A PHYSICS-INFORMED HYBRID DEEP LEARNING FRAMEWORK FOR ENHANCED SEISMIC EVENT PREDICTION WITH INTERPRETABLE FEATURE ATTRIBUTION

Earthquake prediction remains a formidable challenge in geoscience, with limited accuracy exhibited by traditional machine learning models. Although convolutional neural networks (CNNs) have shown promise, inadequate temporal modeling and limited real-time deployment capabilities restrict their performance.. This study introduces SeisHybridNet, a novel hybrid DL framework that integrates vision transformers (ViTs) for global spatial pattern recognition, temporal convolutional networks (TCNs) for long-range temporal dependency modeling, and physics-informed neural networks (PINNs) to enforce seismic wave propagation constraints. Evaluation of the STEAD dataset demonstrates state-of-the-art performance: 93.7% accuracy, 0.891 F1-score, and 0.947 AUC-ROC, surpassing EQTransformer (91.8% accuracy, 0.869 F1-score) and representing a 41.2% relative improvement in F1 over the CNN baseline. Ablation studies have confirmed the complementary contributions of each component. The model achieves an expected calibration error of 0.043 and an inference latency of less than 120 ms, making it suitable for real-time EW systems.