Machine Learning for the Spatiotemporal Prediction of Seismic Ground Motion Fields: A State‐of‐the‐Art Review

ABSTRACT

Spatiotemporal prediction of seismic ground motion represents a critical scientific challenge in earthquake engineering and seismology, with important implications for seismic design, rapid post‐earthquake damage assessment, and emergency response. Traditional methods, which primarily rely on physics‐based simulations and ground motion prediction equations, face inherent limitations in balancing prediction accuracy and computational efficiency. In recent years, the continuous expansion of strong motion records and numerical simulation databases has enabled machine‐learning (ML) approaches to demonstrate significant advantages in ground motion prediction. This study systematically reviews advances in ML–based spatiotemporal ground motion prediction, focusing on two key aspects: spatial ground motion parameter prediction and time series prediction. The reviewed studies are further categorized according to input data type, including feature parameters, sparse spatial fields, and station time series data, while recent developments in physics‐informed learning are also summarized. Existing research indicates that ML approaches show strong potential for modeling complex nonlinear spatiotemporal characteristics of seismic motions and for improving computational efficiency under specific data conditions. However, important challenges remain regarding data imbalance, cross‐region generalization, physical consistency, interpretability, and uncertainty quantification. Current evidence suggests that ML‐based approaches should presently be regarded as complementary tools rather than replacements for traditional ground motion prediction frameworks in engineering‐critical applications. This review aims to provide a critical reference for future development of ML‐based spatiotemporal ground motion prediction methods.