Minisymposium Presentation

This study integrates the Multiple-Input Fourier Neural Operator (MIFNO) with the diffusion model by Gabrielidis et al. (2024) to address challenges in capturing mid-frequency details in synthetic earthquake ground motion. MIFNO, a computationally efficient surrogate model for seismic wave propagation, processes 3D heterogeneous geological data along with earthquake source characteristics. It is trained to reproduce the three-component (3C) earthquake wavefield at the surface. The HEMEWS-3D database (Lehmann et al., 2024) is used, comprising 30,000 earthquake simulations across varying geologies with random source positions and orientations. These reference simulations were conducted using the high-performance SEM3D software (CEA et al., 2017), which excels in simulating fault-to-structure scenarios at a regional scale. While SEM3D provides accurate results at lower frequencies, its performance degrades with increasing frequency due to complex physical phenomena and a known bias in neural networks, which struggle with small-scale features. This limitation restricts MIFNO's applicability in earthquake nuclear engineering. The proposed combination with the diffusion model aims to mitigate this issue and improve the accuracy of mid-frequency predictions in synthetic ground motion generation.