Ambient Noise Full Waveform Inversion with Physics-Informed Generative Adversarial Networks

Extreme climate events and increasing geohazard risks require high-resolution near-surface seismic imaging to better characterize subsurface structures. Ambient seismic noise provides a cost-effective alternative to active-source surveys and has been widely used for S-wave velocity imaging through dispersion-based ambient-noise tomography. However, these approaches rely on accurate Green’s function retrieval, which assumes isotropic and uncorrelated noise sources—conditions rarely satisfied in real field environments. As a result, waveform distortions and resolution loss are common, limiting the quantitative interpretability of conventional ambient-noise imaging.

Ambient-noise full waveform inversion (FWI) offers a promising pathway to overcome these limitations by directly fitting cross-correlation waveforms and fully exploiting waveform information. Nevertheless, its application remains challenging due to strong trade-offs between subsurface structure and unknown noise source characteristics, severe nonlinearity and cycle-skipping, and the lack of reliable constraints on noise source distributions. These issues have so far hindered the practical implementation of ambient-noise FWI in complex near-surface settings.

To address these challenges, we develop a physics-informed generative adversarial network (PIGAN) framework for ambient-noise waveform inversion to accurately estimate physically consistent velocity models in a distributional sense. The wave-equation-based cross-correlation operator is embedded into the generator to ensure physical consistency, while a neural-network discriminator evaluates the mismatch between observed and simulated data. A one-dimensional Wasserstein distance is adopted to enhance robustness to noise and phase uncertainties. The proposed method organically integrates wave-equation constraints, deep learning, optimal transport metrics, and a minimax game formulation, combining the strengths of physics-informed modeling and data-driven representation. This framework enables joint inversion for subsurface velocity structure and ambient noise source characteristics, effectively mitigating source–structure trade-offs. Moreover, it does not require labeled datasets or network pretraining; therefore, the framework is flexible and enables inversion with minimal user interaction. Synthetic tests and field applications in the Qinghai–Tibet Engineering Corridor demonstrate improved resolution and deeper illumination, providing new constraints on fault zone structures and implications for geohazard assessment.