Explaining Deep Learning Models in Full Waveform Inversion: Enhancing Transparency in Seismic Data Interpretation

Abstract We present a physics-informed neural network approach for full-waveform inversion (FWI) that embeds a Deep Image Prior (DIP) U-Net within the inversion loop and applies explainable AI (XAI) methods to interpret the learned model. Starting from a decimated Marmousi velocity model, we generate synthetic seismic data via 2D acoustic wave propagation. A smoothed version of the true model initializes a randomly initialized U-Net (the DIP), which is pretrained to replicate the smoothed model. Thereafter, the DIP's weights replace the cell-wise velocity parameters in FWI, and an iterative optimizer adjusts these weights to minimize discrepancies between observed and simulated waveforms. The resulting high-resolution velocity model arises from physics-constrained learning. To reveal the network's decision process, we compute attribution maps using SHapley Additive exPlanations (SHAP), Integrated Gradients, and occlusion sensitivity. These maps highlight the spatial features; primarily high-contrast layers and anomalies that drive the DIP's output. Comparative analyses of true vs. predicted models, residuals, and combined depth/lateral sensitivity profiles confirm that the deep model captures essential subsurface structures. This framework enhances transparency in seismic inversion by linking data-driven models to geophysical knowledge, thus bridging the gap between "black-box" deep learning and interpretable imaging.