Year: 2025
Type: article
Abstract: Machine learning models based on regression analysis methods have been selected to identify areas of maximum risk for major seismic events. The performance of nine linear and nonlinear models was evaluated, resulting in the selection of the Random Forest model. The quality of training for the Random Forest model was improved through hyperparameter tuning as well as the use of clustering and polar coordinates. It allowed the improvement of quality of model training, increasing the value of the coefficient of determination to 0.86. An analysis was conducted on the applicability of two neural networks with deep learning: Multi-layer Perceptron (MLP) and Long Short-Term Memory (LSTM), using training parameters that were selected for the Random Forest model. Using this model and selected neural networks with deep learning, areas of maximum risk for seismic events were predicted for the entire globe, with a detailed analysis of predicted areas for the territory of the Russian Federation. As a result of the conducted research, the use of neural networks with deep learning made it possible to identify a greater (~40%) number of zones of maximum seismicity (with M>6) compared to the improved Random Forest model.
DOI: 10.2205/2025es001055
Authors: Valeriy Bondur, Olga Voronova
Institution: Aerocosmos
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4414058947