Physics-Informed Cross-Modal Attention (PI-CMA) Reveals Physically Consistent Geomagnetic–Seismic–Hydrologic Precursory Coupling
Abstract
Abstract We present PI-CMA, a physics-informed crossmodal Transformer that aligns multimodal geophysical time series—geomagnetic total-field (F), seismicity (b-value and moment release), and groundwater level—to test whether learned attention weights are physically consistent with known earthquake precursor mechanisms. Using a strictly retrospective blind test (train cutoff 2023-06-30; test 2023-07-01–2024-12-31) over 513 M≥6.0 event windows and 5 INTERMAGNET/ USGS geomagnetic observatories (2020–2024), we find a mean Physical Consistency Index (PCI) of 0.170±0.099 for the tri-modal model (geomagnetic ↔ seismic + groundwater), with median per-window p =7.57e-02; 0/513 windows survive Bonferroni correction. Adding groundwater does not change PCI (0.170→0.170), and quiescent controls show a higher PCI (0.51–0.52), indicating that attention–b-value consistency mainly tracks baseline structure rather than eventspecific anomalies. Against classical baselines (b-value alarm, Coulomb-stress proxy, logistic regression, random forest, 1D-CNN) the method is competitive on ROC-AUC (0.676) while providing a physically interpretable consistency metric rather than a black-box forecast. We release the 5-year multimodal dataset and a reproducible pipeline, and caution against overclaiming AI-based precursory skill.
Read the paper
Related studies
- Current state and future directions for deep learning based automatic seismic fault interpretation: A systematic review
- Physics of Partially Saturated Porous Media: Residual Saturation and Seismic-Wave Propagation
- Modern Imaging Using Seismic Reflection Data
- Effects of marine seismic surveys on free-ranging fauna: a systematic literature review
- A review of the recent seismic activity in the southern part of egypt (upper egypt)
- A review of research on mechanism of seismic attenuation and energy compensation
Go deeper in Evidence Hub
Evidence Hub adds a study data sheet (design, sample size, evidence level, key numbers), an AI summary and a chat grounded in the paper's text.