A machine learning-based ground motion model for the Chilean subduction zone and its application to probabilistic seismic hazard analysis

Accurate prediction of ground motion intensity measures (IMs) is critical for seismic hazard assessment, particularly in zones where complex tectonic interactions govern earthquake dynamics. Traditional Ground Motion Models (GMMs) rely on common regression-based frameworks, often overlooking nonlinear source-path-site interactions and limiting their IM scope to peak parameters (e.g., PGA, PGV). This study presents a novel neural network-based GMM (GMM-NN) tailored for Chilean subduction earthquakes, leveraging machine learning (ML) to predict a comprehensive suite of 14 IMs—including directional response spectra, cumulative metrics, and peak parameters. The model employs a two-level ensemble architecture: a first-level ensemble generates distributions for non-spectral IMs, such as significant durations and Arias Intensity, while a second-level ensemble predicts spectral accelerations, explicitly incorporating the natural period of vibration (T) as an input to enable continuous period-dependent modeling. Trained on a comprehensive Chilean database, the GMM-NN demonstrates superior accuracy with R² scores up to 30% above conventional models, resolving magnitude-dependent attenuation, and, notably, capturing regional geological variability through the inclusion of latitude as an input parameter, something that no other GMM has ever done. A mixed-effects framework quantifies heteroscedastic uncertainty, with ensemble-derived standard deviations that are 25% lower than those of traditional models, ensuring robust probabilistic seismic hazard analysis (PSHA). Model integration into PSHA workflows yields refined uniform hazard spectra (UHS), validated for cities all throughout Chile, and advances performance-based design by supporting multidirectional structural analysis. While the current dataset limits extrapolation to megathrust earthquakes (Mw > 8.8), the GMM-NN establishes a scalable template for data-driven seismic hazard models, emphasizing the transformative potential of ML in capturing ground motion complexity.