Numerical stress fields meet machine learning: A new approach to earthquake forecasting in Western Tien Shan

Earthquake prediction in regions of complex tectonics remains a fundamental challenge in geophysics. Central Asia, and particularly Uzbekistan, is characterized by intense crustal deformation due to the interaction of the Indian, Arabian, and Eurasian plates. In this study, we integrate numerical modeling of the stress–strain state of the lithosphere with advanced machine learning methods to improve seismic hazard assessment. A regional stress model at seismogenic depths (15–20 km) was developed, incorporating tectonic zoning and boundary conditions that reflect both welded and frictional fault contacts. The model reproduces the first-order distribution of stresses and velocities and shows strong agreement with independent GPS observations and focal mechanism solutions. To extend these results toward predictive applications, the outputs of the numerical model were combined with seismological and morphological features and analyzed using a hybrid Kolmogorov–Arnold Network coupled with an LSTM framework. The machine learning system successfully identified zones of elevated seismic hazard in the Western Tien Shan, which coincided with the locations of strong earthquakes that occurred in 2023. These results confirm that combining geodynamic modeling with artificial intelligence provides a powerful approach for linking stress accumulation to earthquake occurrence. The methodology offers both a robust geophysical foundation and a practical predictive tool for regions of high seismic risk.