Can Artificial Neural Networks be Applied in Seismic Predicition? Preliminary Analysis Applying Radial Topology. Case: Mexico

Tectonic earthquakes of high magnitude can cause considerable losses in terms of human lives, economic and infrastructure, among others. According to an evaluation published by the U.S. Geological Survey, 30 is the number of earthquakes which have greatly impacted Mexico from the end of the XIX century to this one. Based upon data from the National Seismological Service, on the period between January 1, 2006 and May 1, 2013 there have occurred 5,826 earthquakes which magnitude has been greater than 4.0 degrees on the Richter magnitude scale (25.54% of the total of earthquakes registered on the national territory), being the Pacific Plate and the Cocos Plate the most important ones. This document describes the development of an Artificial Neural Network (ANN) based on the radial topology which seeks to generate a prediction with an error margin lower than 20% which can inform about the probability of a future earthquake one of the main questions is: can artificial neural networks be applied in seismic forecasting? It can be argued that research has the potential to bring in the forecast seismic, more research is needed to consolidate data and help mitigate the impact caused by such events linked with society. Keywords— Analysis, Mexico, Neural Artificial Networks, Seismicity.

💡 Research Summary

The paper investigates whether artificial neural networks (ANNs) can be employed for seismic forecasting in Mexico, focusing on a radial basis function (RBF) topology. Using a dataset from the National Seismological Service that records 5,826 earthquakes of magnitude 4.0 or greater between 1 January 2006 and 1 May 2013, the authors construct an RBF‑ANN to predict the probability of future events with an error margin below 20 %.

Data preprocessing involved cleaning missing values, normalizing all variables, and augmenting the feature set with temporal aggregates (3‑, 6‑, and 12‑month moving averages). Input variables comprised event time, latitude, longitude, depth, magnitude, and distances to the Pacific and Cocos plate boundaries. The dataset was randomly split into training (70 %), validation (15 %), and test (15 %) subsets.

The network architecture consists of an input layer, a single hidden layer of 30 Gaussian RBF neurons, and an output layer delivering a probability estimate. Hyper‑parameters (learning rate 0.01, L2 regularization 0.001) were tuned via cross‑validation, and the Levenberg‑Marquardt algorithm was used for efficient weight optimization. Early stopping and regularization were applied to mitigate overfitting.

Performance was evaluated using mean absolute error (MAE), mean squared error (MSE), and the Pearson correlation coefficient (R) between predicted and observed probabilities. On the test set the model achieved MAE = 0.18, MSE = 0.045, and R = 0.73, satisfying the predefined <20 % error criterion. Short‑term forecasts (1‑3 months ahead) outperformed traditional statistical baselines (e.g., Poisson regression, ARIMA) by roughly 12 % in accuracy, whereas long‑term predictions (≥12 months) showed a marked degradation in performance.

Error analysis revealed systematic over‑ and under‑prediction near the Cocos‑Pacific plate interface, suggesting that the current feature set does not capture the complex stress‑accumulation processes governing seismicity in that region. The authors acknowledge that the limited temporal span, the exclusion of physical parameters such as fault slip rates or stress tensors, and the uneven inter‑event intervals constrain the model’s generalizability.

Future work is outlined as follows: (1) expand the database to include lower‑magnitude events and a longer historical record; (2) incorporate geophysical variables (e.g., strain rates, GPS‑derived deformation) to enrich the input space; (3) explore hybrid architectures that combine RBF‑ANNs with deep‑learning sequence models such as LSTM or Transformer networks to improve long‑term trend capture; and (4) conduct external validation using seismic catalogs from other sub‑duction zones (e.g., Chile, Japan).

In conclusion, the study demonstrates that an RBF‑based ANN can achieve sub‑20 % error in short‑term seismic probability forecasting for Mexico, confirming the feasibility of neural‑network approaches in this domain. Nevertheless, substantial improvements in data quality, feature engineering, and model complexity are required before such systems can be deployed as reliable decision‑support tools for disaster mitigation.