Towards Fourier Neural Operators for Source and Structural Inversion of the San Francisco Bay Area

Claire Doody, Jiun-Ting Lin, Qingkai Kong, Arthur J. Rodgers, Luis Vazquez, Caifeng Zou, Youngsoo Choi, Zachary E. Ross, Kamyar Azizzadenesheli, & Robert W. Clayton

Submitted August 30, 2026, SCEC Contribution #15407, 2026 SCEC Annual Meeting Poster #TBD

Fourier Neural Operators (FNOs) are machine learning methods that approximate solution operators for partial differential equations, enabling efficient calculations of wave propagation. These methods have substantially reduced the computational costs in forward modelling applications (e.g., Kong et al., 2025; Yang et al., 2021) and have been applied to ambient noise tomography modelling in the Los Angeles Basin (Zou et al., 2024). Here, we present a three-dimensional FNO model developed for the San Francisco Bay Area. We trained the model on nearly 20,000 3D wavefield simulations generated from randomly distributed sources within a 160 km x 160 km domain. In addition to predicting wavefield, we apply the trained model to estimate source parameters (strike, dip, and rake) for the earthquakes that occurred across our domain. Finally, we discuss a future framework for imaging 3D structure across the Bay Area, combining FNO-based forward simulations with an auto-differentiation-based approach for structure inversion.

Citation
Doody, C., Lin, J., Kong, Q., Rodgers, A. J., Vazquez, L., Zou, C., Choi, Y., Ross, Z. E., Azizzadenesheli, K., & Clayton, R. W. (2026, 08). Towards Fourier Neural Operators for Source and Structural Inversion of the San Francisco Bay Area. Poster Presentation at 2026 SCEC Annual Meeting.

Related Projects & Working Groups
Seismology