Earthquake Magnitude Prediction Using AI in Light of Environmental Factors

Abstract

Earthquake magnitude-category modelling is difficult because damaging events are uncommon, catalogue observations are temporally dependent, and environmental associations may be weak. This study integrates daily United States Geological Survey (USGS) earthquake records with mean temperature and precipitation from seven Alaska Climate Research Center (ACRC) stations for 1960–2023. A reproducibility track retained the submitted random 80/20 split, whereas a reviewer-driven chronological protocol used five expanding-window validation years (2017–2021) and an untouched 2022–2023 holdout. On the submitted split, the original FFNN achieved 75.30% accuracy, macro precision of 73.13%, macro recall of 71.01%, and macro F1 of 68.54%; its Major-class precision, recall, and F1 were 71.32%, 45.83%, and 55.80%, respectively. Relative to the strongest non-FFNN baseline, this was +0.41 percentage points in accuracy and +1.63 points in macro F1, although stacking gave a slightly higher Major F1 (56.65%). “Under chronological validation, the compact FFNN averaged 65.72% accuracy and 40.57% macro F1. On the untouched holdout, the decision tree produced the highest Major F1 (14.67%). A paired environmental ablation showed modest lagged-feature gains: holdout Major F1 increased from 9.86% to 14.57% for FFNN and from 15.22% to 21.40% for LightGBM. However, the confidence intervals crossed zero and Holm-adjusted tests were not significant. SHAP analysis ranked depth and year above temperature and precipitation. The evidence therefore supports environmental variables as secondary, model-dependent signals rather than dominant or causal predictors.