Oral Presentation

Realistic seismic wavefield simulation is essential for numerous applications, including acquisition design, imaging, and inversion. However, traditional numerical simulators are computationally intensive for large-scale 3D models and often yield discrepancies between simulated and observed data due to uncertainties in the chosen wave equation and inaccuracies in input parameters such as subsurface elastic properties and source characteristics. To overcome these limitations, we introduce a data-driven artificial intelligence (AI) framework Conditional Generative Modeling for efficient and realistic seismic wave simulation. CGM learns complex 3D wave physics and subsurface heterogeneities directly from observed seismic data, eliminating the need for explicit physical modeling. Once trained, CGM-based models act as stochastic wave propagation operators conditioned on local subsurface properties captured in the training data.

These models can simulate multicomponent seismic responses for arbitrary acquisition configurations within the area of observation, using source and receiver geometries, as well as source parameters, as input conditions. We present four CGM variants, CGM-GM-1D, CGM-GM-3D, CGM-Wave, and CGM-FAS, and evaluate them on two datasets: (1) a low-density natural earthquake dataset from the San Francisco Bay Area, and (2) a high-density induced seismicity dataset from the Geysers geothermal field. Our results show that CGM effectively reproduces observed waveforms, spectra, and kinematic features, demonstrating its ability to generalize across different spatial and source configurations. Moreover, the ground motions estimated by CGM can be used for ground motion modeling with magnitude extrapolation, further enhancing its utility in seismic hazard assessment.