Hybrid broadband ground-motion simulation using neural networks with spatial, inter-period, and cross-component correlations

Simulated ground motions are increasingly used in earthquake engineering, particularly in regions with sparse strong-motion recordings where constraining non-ergodic ground-motion models (GMMs) remains challenging. Physics-based simulations (PBS) can reproduce key source and wave-propagation effects but are often limited to low frequencies, whereas stochastic methods are computationally efficient but typically lack physically coherent three-component behaviour and realistic near-fault features. Here we develop a hybrid broadband framework with three main innovations (i) an enhanced artificial neural network (ANN) that predicts short-period spectral accelerations (SAs) from long-period PBS SAs and scalar source-site predictor variables, while jointly modelling the three ground-motion components, (ii) a transfer learning strategy that enables regional calibration in data-limited settings, and (iii) an explicit multivariate correlation model for within-event residuals that restores realistic broadband dependence jointly across space, periods, and components. The approach is validated using regional 3D PBS of the Mw 6.5 and Mw 6.4 June 2000 South Iceland earthquakes. The resulting broadband fields reproduce observed short-period attenuation and near-fault saturation, match the local Icelandic GMM trends, and preserve physically plausible directionality and component ratios, including spatially varying FN/FP patterns and V/H ratios consistent with empirical models. Inter-period correlations of within-event residuals follow established empirical trends and, importantly, remain continuous across the stochastic–PBS transition, while spatial correlation ranges are comparable to published observations. Owing to its computational efficiency, the framework is well suited for seismic hazard and risk applications that require spatially and spectrally coherent broadband ground motions in dense spatial grids.