Robust Physics-Informed Reparameterized Full-Waveform Inversion via CNN-Enhanced Vision Transformer

Abstract Full-waveform inversion (FWI) provides high-resolution subsurface characterization but remains vulnerable to ill-posedness, cycle skipping, and local minima when the starting model is inaccurate or low-frequency information is missing. We introduce a physics-informed reparameterized FWI framework that leverages a hybrid architecture combining convolutional neural network (CNN) and a vision transformer (ViT) enhanced with spatial-reduction attention (SRA), which reduces the computational cost while preserving global dependencies, to enhance robustness under challenging acquisition conditions. In the proposed scheme, the CNN extracts multi-shot local seismic attributes, whereas the ViT models long-range correlations and enforces structural coherence. The untrained nature of the hybrid network acts as an implicit regularization, enabling smooth and geologically plausible model updates while reducing the non-uniqueness of the inversion. Numerical tests on representative synthetic models demonstrate that the method reliably reconstructs velocity structures from low-quality initial models and outperforms conventional FWI and CNN-based FWI approaches, particularly under noise contamination and low-frequency-deficient data. The field data example further demonstrates that the recovered velocity models lead to improved seismic imaging quality and provide a more reliable foundation for subsequent imaging workflows.