AutoDiff-Based Full Waveform Inversion for Simultaneous Estimation of the Source Wavelet and Velocity Model

Full Waveform Inversion (FWI) is a powerful technique for high-resolution subsurface imaging, yet it remains highly sensitive to inaccuracies in input parameters, especially the source wavelet. Traditional FWI methods typically assume the wavelet is known and fixed, which can lead to convergence toward incorrect models. In this study, we present a novel FWI framework that leverages Automatic Differentiation (AD) to simultaneously estimate both the subsurface velocity model and the per-shot source wavelets. By implementing a custom AD-compatible finite-difference wave propagation algorithm, we compute gradients with respect to all model parameters, including wavelets, without requiring analytical adjoint derivations. Synthetic experiments using a modified Marmousi model demonstrate that joint inversion leads to significantly improved velocity reconstructions and better waveform matching compared to fixed-wavelet FWI. Our results highlight the potential of AD-based inversion for tackling complex and ill-posed seismic problems.

کلیدواژه ها:

Automatic Differentiation ، Full Waveform inversion ، deep learning ، simultaneous inversion ، source wavelet estimation