Generative inverse modeling for improved geological CO2 storage prediction via conditional diffusion models

Geological CO2 storage is expected to play a pivotal role in achieving climate-neutrality targets by 2050. Accurate prediction of long-term CO2 storage performance relies on inverse modeling procedures that precisely characterize spatially varying geological properties using practically available observed data. Traditional inversion methods necessitate extensive forward simulations to iteratively calibrate uncertain geological parameters, which can impose a significant computational burden. In this work, an end-to-end generative inversion framework based on the conditional diffusion model is proposed for efficiently characterizing heterogeneous geological properties and accelerating the inversion process. By employing an improved U-net to learn the conditional denoising diffusion process, the proposed framework enables the direct generation of high-dimensional property fields that closely match the observed data. Additionally, the probabilistic nature inherent in the diffusion approach allows for producing an ensemble of plausible geological realizations, facilitating effective quantification of parametric and predictive uncertainties. The performance of the proposed framework is validated by estimating stochastic permeability fields for both two-dimensional and three-dimensional carbon storage models. Comprehensive comparisons with the conditional generative adversarial network-based method demonstrate that the proposed framework yields more accurate inversion results and better quantifies the uncertainty in the predicted flow responses. This work offers a promising tool for subsurface inverse modeling and uncertainty quantification, potentially paving the way for broader adoption and exploration of generative diffusion models in the realm of energy system management.