Determining locations of possible earthquakes in the western Tien Shan using artificial neural network and a mathematical model of tectonic processes

In this paper, we developed a numerical model of the stress state of the earth’s crust of the Western Tien Shan microplate to use as additional features for machine learning. An alternative to the deep learning models could be a neural network based on the Kolmogorov‒Arnold (KAN) general approximation theorem. What distinguishes KAN from existing machine learning networks is its interpretability, i.e. the ability to explain the “logic” of the model’s operation and high accuracy in complex physical processes. KANs differs from existing machine learning networks in its high interpretation and accuracy in complex physical processes. Unlike conventional networks, KAN neural network requires only one or two layers to obtain a solution to the problem, which significantly reduces the computing power. Using the KANs algorithm, we have built for the first time a neural network for classification and regression applied to the medium-term earthquake prediction in the Western Tien Shan microplate. The results obtained allowed us to predict the locations of possible earthquakes with a magnitude of 5 > M < 6 in environs of the city Tashkent (the Capital of Republic of Uzbekistan). The performed retrospective analysis of strong earthquakes that occurred in 2024 within the West Tien Shan microplate showed that the developed model predicts the locations of earthquakes with a magnitude of M < 6 with an accuracy of geographic coordinates of ±0.1° N, ±0.1° E and a magnitude of ΔM = ±0.4.