Deep Learning Matching Filtering for Elastic Full Waveform Inversion

Abstract

Elastic full waveform inversion (EFWI) aims to reconstruct high‐resolution elastic properties by minimizing the waveform misfit between observed and simulated multicomponent seismic data, yet its practical performance is often limited by cycle skipping and modelling errors that conventional misfit‐based methods cannot fully resolve. We propose a deep‐learning‐based matching‐filtering framework, in which a lightweight neural network is trained to learn adaptive, data‐driven matching filters that map synthetic elastic wavefields towards the observed data. Instead of directly enforcing waveform agreement, the learned filters absorb phase, amplitude and dispersion discrepancies arising from inaccurate initial models, elastic‐parameter trade‐offs and imperfect physics. The matching filters are optimized jointly with elastic parameters within an AD framework, enabling seamless gradient propagation through both the wave‐equation solver and the neural components. Numerical experiments on synthetic elastic models and field data demonstrate that the proposed approach significantly reduces cycle skipping and improves the recovery of P‐ and S‐wave velocities compared to conventional EFWI, particularly when the initial model is strongly biased. The results suggest that deep learning–based matching filtering provides a physically interpretable and computationally efficient pathway for robust elastic waveform inversion under realistic modelling uncertainties.