Bridging the Acoustic-Elastic Gap in Seismic Inversion via Robust Summary Statistics

Simulation-based seismic inversion provides a principled framework for uncertainty quantification, but its performance degrades when the physics used in simulation differ from those of observed data. In practice, seismic inversion is often misspecified because the true physics are only partially known or too costly to model at scale, making field observations out of distribution relative to the simulations used for training; here, we study this issue through the acoustic-elastic gap. We introduce ElasNet, a misspecification-robust probabilistic inversion framework that bridges acoustic simulations and elastic observations through physics-informed summary statistics. We use acoustically imaged common-image gathers (CIGs) as structured summaries of seismic data, compress them with an attention-based summary network, and use maximum mean discrepancy (MMD) to align acoustic and elastic feature distributions during training while learning a conditional normalizing flow for posterior estimation with only acoustic data. Experiments on 2D Compass models show that the proposed approach reduces acoustic-elastic discrepancies in inferred velocity models and improves structural consistency and uncertainty calibration. These results demonstrate that robust summary statistics and distribution matching provide a scalable pathway for simulation-based seismic inversion under realistic physics mismatch.

https://slim.gatech.edu/Publications/Public/Submitted/2026/deng2026IMAGE...