QuakeF2G: A region-specific transformer-based ground motion prediction model considering fault segments modeling

Rupture directivity remains a major challenge for near-fault ground motion prediction, particularly for urban regions located close to active faults. Empirical ground motion prediction equations (GMPEs), even with directivity correction factors, often oversimplify rupture geometries, while physics-based wave propagation simulations are computationally prohibitive for rapid and dense analysis. To bridge this gap, we present QuakeF2G, a Transformer-based architecture for probabilistic ground motion prediction that fuses heterogeneous geometric and observational inputs through a unified cross-attention framework. To avoid explicit rupture parameterization, the model tokenizes the rupture as a set of discrete fault segments, encodes each with location and orientation attributes, and fuses the resulting fault sequence with station tokens through cross-attention augmented by a geometry-aware relative-position bias. A flexible masking strategy enables the same model to perform pure prediction from fault geometry and site metadata alone, or conditional interpolation when PGV observations are available at some stations. After training on approximately 626,000 synthetic scenarios for peak ground velocity (PGV) from the CyberShake dataset in the Southern California, QuakeF2G achieves a speedup of more than three orders of magnitude in prediction time compared to physics-based simulations while maintaining high predictive fidelity. Ablation experiments demonstrate that explicitly incorporating finite-fault geometry reduces root-mean-square error (RMSE) by 36% compared to a point-source baseline. In spatial interpolation, the model outperforms Ordinary Kriging across all sampling densities, achieving a lower RMSE using only 50% of training stations than Ordinary Kriging achieves with all training stations. QuakeF2G demonstrates a general methodology for fusing irregular geometric inputs with sparse observations in engineering prediction tasks.