AI Surrogate Model Predicts CO2 Storage Operations in Faulted Aquifers

Researchers have developed a multimodal auto-regressive transformer surrogate to model variable operations and quantify uncertainty in geological carbon storage. The study, available on arXiv (2608.02629), addresses the challenge of optimizing well perforation and injection strategies under geological uncertainty. The surrogate processes three input modalities: a 3D geomodel, scalar parameters for relative permeability functions, and control variables (stage durations and injection rates). These are fused via self-attention in a transformer encoder, with a temporal decoder generating predictions auto-regressively. The model was trained on 4000 GEOS flow simulations using a modified SEAM CO2 geomodel, which includes a faulted system with three stacked aquifers. Two injection wells are perforated in stages from bottom to top. This approach aims to improve the efficiency of carbon storage operations by enabling rapid predictions and uncertainty quantification, potentially aiding in the design of more effective carbon capture and storage (CCS) projects.