Performance Evaluation of LSTM Networks for Earthquake Magnitude Prediction in Iran

Iran, located along an active seismic belt, frequently experiences destructive earthquakes, making accurate forecasting crucial for disaster preparedness and risk mitigation. This study employs Long Short-Term Memory (LSTM) networks to predict earthquake magnitudes using a dataset of 6,916 seismic events recorded in Iran from 1900 to the present, with magnitudes ranging from 4.0 to 7.7. Various loss functions and resampling methods were applied to optimize predictive accuracy, and the performance of four models was compared. Results indicate that LSTM networks achieved a high correlation across the full magnitude range, with yearly resampling yielding the most accurate predictions. For large earthquakes (6.0 ≤ M < 7.7), the Pseudo-Huber loss function improved model stability, though predictions were constrained by data scarcity. While daily and monthly predictions exhibited higher variance, yearly forecasting provided more reliable long-term trends. This study underscores the importance of selecting appropriate time intervals and loss functions in earthquake prediction models. The findings contribute to seismic hazard assessment efforts and can aid in developing early warning systems for earthquake-prone regions.