Implicit Neural Representation for Elastic Full-Waveform Inversion

Full-waveform inversion (FWI) has become a cornerstone for high-resolution seismic imaging, yet it remains computationally demanding and sensitive to initial model assumptions and noise. Recent advances have shown that representing the subsurface model, using implicit neural representations (INRs), can provide compact, continuous and differentiable parameterizations that improve convergence and reduce overfitting. In this study, we extend the INR-based FWI framework to the elastic regime, with a focus on near-surface applications and the inversion of surface waves. In particular, we performed the inversion of both synthetic and field surface wave datasets. Our method leverages Deepwave for elastic wave simulation and gradient computation via automatic differentiation. In the synthetic test, we compare the performance obtained using different INR architectures to find the optimal configuration. For the field dataset inversion instead, we compare our results with those obtained using a standard deterministic FWI approach, highlighting its superior robustness with respect to initialization.