Simulating realistic seismic wavefields is crucial for a range of seismic tasks, including acquisition designing, imaging, and inversion. Conventional numerical seismic wave simulators are computationally expensive for large 3D models, and discrepancies between simulated and observed waveforms arise from wave equation selection and input physical parameters such as the subsurface elastic models and the source parameters. To address these challenges, we adopt a data-driven artificial intelligence approach and propose a conditional generative modeling (CGM) framework for seismic wave simulation. The novel CGM framework learns complex 3D wave physics and subsurface heterogeneities from the observed data without relying on explicit physics constraints. As a result, trained CGM-based models act as stochastic wave-propagation operators encoded with a local subsurface model and a local moment tensor solution defined by training data sets. Given these models, we can simulate multicomponent seismic data for arbitrary acquisition settings within the area of the observation, using source and receiver geometries and source parameters as input conditional variables. In this study, we develop four models within the CGM framework-CGM-GM-1D/3D, CGM-Wave, and CGM-FAS-and demonstrate their performance using two seismic data sets: a small low-density data set of natural earthquake waveforms from the San Francisco Bay Area, a region with high seismic risks, and a large high-density data set from induced seismicity records of the Geysers geothermal field. The CGM framework reproduces the waveforms, the spectra, and the kinematic features of the real observations, demonstrating the ability to generate waveforms for arbitrary source locations, receiver locations, and source parameters. We address key challenges, including data density, acquisition geometry, scaling, and generation variability, and we outline future directions for advancing the CGM framework in seismic applications and beyond.
