A long short-term memory framework for surface intensity prediction using source and borehole parameters for improved seismic hazard assessment

The complex refraction and reflection of seismic waves through heterogeneous soil layers introduce significant stochasticity, necessitating advanced techniques for accurate seismic hazard analysis. Predicting surface-level ground motion intensity measures (IMs) is crucial for geotechnical, structural, and earthquake engineering applications, including probabilistic seismic hazard analysis (PSHA). This study presents a novel end-to-end framework using long short-term memory (LSTM) recurrent neural networks (RNNs) to predict 40 surface IMs including amplitude, frequency, duration, and energy, using earthquake source and site parameters. The framework comprises two sequential models: (i) E2B, predicting borehole-level IMs from source parameters, and (ii) EB2S, predicting surface-level IMs using both source parameters and E2B outputs. This composite approach is benchmarked against a direct baseline (E2S) model which predicts surface IMs directly from source and site parameters. A robust Japanese dataset of over 2,600 surface–borehole IM pairs is used for training and validation. Two feature selection strategies are evaluated: a physics-derived (PD) set and a data-driven (DD) set based on random forest importance. The DD EB2S model shows up to 30% coefficient of determination ((Formula presented.)) for short-period IMs (e.g. (Formula presented.)) and maintains high accuracy (test (Formula presented.)) for long-period IMs. The PD EB2S model also improves upon PD E2S, though with smaller gains (typically 2–6%). The DD EB2S model also better preserves inter-IM correlations, offering a scalable and interpretable tool for IM prediction.