Physics-Guided VAE Reduces Non-Uniqueness in Rayleigh-Wave Inversion on Devine Field Data

The arXiv preprint (v1, Oct 2026) trains the PG-VAE on elastic synthetic gathers, then applies a differentiable surrogate forward model as a consistency loss to penalize dispersion violations. A residual decoder biases outputs toward stratified geology while Grad-CAM confirms the network attends to physically meaningful phase-velocity features rather than noise. On held-out synthetics the method cuts velocity RMSE by roughly 30 percent versus plain CNN baselines; field results at Devine show sharper shallow interfaces consistent with known logs. Traditional dispersion picking scales poorly with modern dense arrays and ignores solution non-uniqueness, often yielding overly smooth models that misrepresent hazard-prone layers. The PG-VAE's physics regularizer directly addresses this, yet the single-site field test leaves open questions about performance across varying near-surface regimes and noise conditions. A multi-site blind benchmark with independent full-waveform inversion reference models would strengthen claims of operational readiness. Follow-on work could embed the VAE latent space inside elastic FWI workflows to supply geologically plausible starting models, potentially shortening convergence time for high-resolution hazard studies.