Work

Physically Consistent Peak Ground Acceleration Prediction in Bali Using Ensemble Machine Learning

Year: 2026
Type: article
Abstract: Bali faced significant seismic risks from the Java Trench and back-arc thrust faults, making accurate Peak Ground Acceleration (PGA) prediction crucial. While traditional empirical models struggled with complex, non-linear site effects, machine learning offered a robust alternative. This study aimed to develop and compare Random Forest and XGBoost algorithms, utilizing moment magnitude (Mw), hypocentral distance (Rhypo), and local site conditions (Vs30) as primary inputs, to predict PGA in the Bali region. The research utilized 1,000 strong-motion records from 287 earthquakes (Mw > 4.0) recorded between 2018 and 2025. Following rigorous quality control, we applied logarithmic transformations and hyperparameter optimization via Randomized Search with 3-fold cross-validation. Expanding upon our preliminary model evaluations, this comprehensive analysis demonstrated that the Random Forest architecture outperformed XGBoost by maintaining superior predictive stability, yielding a Coefficient of Determination of 0.741, a Mean Absolute Error of 0.340, and a highly constrained residual standard deviation of 0.434. Furthermore, feature importance analysis revealed that Random Forest was physically consistent with seismological principles. It correctly identified hypocentral distance as the dominant ground-motion attenuation predictor, whereas XGBoost exhibited an overfitted bias toward local site conditions. Ultimately, the Random Forest algorithm provided a highly reliable model for PGA prediction, effectively supporting the development of Probabilistic Seismic Hazard Analysis maps and rapid real-time ShakeMaps in Indonesia.
Language: English
Open Access status: diamond
OpenAlex ID: W7223449049
FWCI:
Cites:
Cited by:
Related to: 9