Physics-informed TKAN for earthquake forecasting in the Western Tien Shan

Introduction: Earthquake forecasting remains a challenging problem, as most existing approaches rely primarily on the statistical analysis of seismic catalogs and often lack a clear physical basis. Materials and methods: In this study, we propose a physics-informed framework that integrates numerical modeling of the stress–strain state of the Earth’s crust with a hybrid deep learning architecture, combining Kolmogorov–Arnold Networks (KAN) and Long Short-Term Memory (LSTM). The feature space includes both seismological parameters and geodynamic variables derived from numerical simulations. Results: The proposed model demonstrates improved performance compared to traditional machine learning approaches. The inclusion of physics-based features enhances both classification and regression accuracy, and the model successfully identifies zones of increased seismic hazard in the Western Tien Shan, showing good agreement with observed seismicity. Conclusions: The results confirm that integrating physically meaningful geodynamic features with interpretable deep learning architectures improves the reliability and consistency of earthquake forecasting, providing a promising direction for future research.