American Journal for Young Scientists ยท ISSN 3068-9015
Toward Earthquake Prediction: Combining Physics-Based Models with Machine Learning
Authors: Swagatika Pindiyath Menon
Published: July 14, 2026
Category: Environmental Science
Abstract
One of the biggest goals in seismology is to accurately predict when large earthquakes will occur. This is very difficult because it is hard to model how stress builds up, moves, and is released beneath the Earth's crust. In this study, the authors investigate whether machine learning can improve physics-based simulations to better predict the timing and location of major earthquakes. The team used a physics-based model called the Olami-Feder-Christensen (OFC) model. This model simulates how stress builds up, spreads through rock layers in the Earth's crust, and is eventually released as earthquake-like events. To make the simulations more realistic, the model was adjusted using two well-known seismic laws: the Gutenberg-Richter Law, which describes earthquake magnitudes, and Omori's Law, which explains how aftershocks decrease over time. The stress-field data produced by the OFC model was then used to train a Convolutional Neural Network (CNN). Since stress fields have spatial patterns similar to images, CNNs are well suited for finding patterns that often appear before large earthquakes. The network learned to recognize stress patterns linked to a high chance of major seismic events. The trained CNN showed promising results on test data, achieving a ROC-AUC of 0.67 and a PR-AUC of 0.90. Although it did not predict every large earthquake perfectly, it was able to identify many seismic events while reducing false predictions. Visualizing the stress fields also showed that the CNN was learning real spatial stress patterns instead of relying on random patterns in the training data. Overall, these results show that combining machine learning with physical models is a promising way to improve earthquake forecasting. Although accurately predicting earthquakes remains a difficult challenge, these models can detect subtle, meaningful warning signs and may help create more reliable and effective early warning systems.