Multimodal Transformer Models Carbon Storage Operations and Uncertainty

Optimizing geological carbon storage (GCS) operations, such as variable well perforation and injection strategies, is crucial for improving efficiency. However, modeling these operations accurately under geological uncertainty presents a significant challenge. High-fidelity simulations are computationally intensive, making it difficult to explore a wide range of scenarios and quantify uncertainties. Researchers have developed a novel multimodal auto-regressive transformer surrogate model to address this. This model is designed to process three distinct input modalities: 3D geomodel data, scalar parameters characterizing relative permeability functions, and control variables like injection rates and perforation stage durations. Each modality is handled by separate encoders, with their outputs fused via self-attention in a transformer encoder. A temporal decoder then generates predictions auto-regressively using encoder-decoder cross-attention. The surrogate was trained using 4000 GEOS flow simulations based on a modified SEAM CO2 geomodel, which includes a faulted system with stacked aquifers. It accurately predicts saturation and pressure at monitoring locations, total injected and mobile CO2 mass, and saturation footprints. On a new test set, the model achieved a median saturation MAE of 0.028 and low relative errors for other quantities, notably capturing the switch from rate to bottom-hole-pressure control. When integrated into a hierarchical Markov chain Monte Carlo data assimilation procedure, the surrogate substantially reduced uncertainty for key metaparameters, especially fault permeabilities, demonstrating its utility for robust GCS management.