<![CDATA[Peak ground acceleration (PGA) is an important parameter in seismic hazard assessment because it represents the maximum ground acceleration generated by an earthquake and informs structural design and disaster mitigation. Predicting PGA remains challenging because seismic-wave propagation is nonlinear and affected by geological heterogeneity. This study developed, evaluated, and spatially mapped recurrent neural network (RNN) and long short-term memory (LSTM) models for predicting PGA values calculated using the McGuire empirical equation in BMKG Regional II, Indonesia, which extends from South Sumatra to West Java. A historical earthquake catalog covering 1971–2025 was used to generate a spatial dataset comprising 3,727 grid points based on surface-wave magnitude and hypocentral distance. The dataset was divided sequentially into 70% training data and 30% testing data. On the test set, the LSTM produced an RMSE of 55.1559, an MAE of 31.0443, and an R² of 0.7112, whereas the RNN produced an RMSE of 56.7773, an MAE of 36.5842, and an R² of 0.6940. The spatial results also showed that the LSTM reproduced high PGA values more closely than the RNN, which produced smoother estimates in areas with abrupt PGA variations. Within the dataset and model configuration used in this study, the LSTM therefore provided better predictive performance than the RNN for reconstructing the spatial distribution of McGuire-based PGA. The resulting model may support regional seismic hazard assessment, but validation against observed ground-motion records and the inclusion of local site parameters remain necessary before its use in engineering or regulatory applications.]]>
Copyrights © 2026
