Stanford's multimodal transformer surrogate nails CO2 storage with 0.028 error
A new AI surrogate predicts CO2 injection dynamics with just 0.2–5% relative error, slashing simulation costs.
Yifu Han and Louis J. Durlofsky from Stanford University have introduced a multimodal auto-regressive transformer surrogate designed to model variable well operations and quantify geological uncertainty in carbon storage. Traditional full-physics simulations using GEOS are accurate but computationally expensive, limiting their use in uncertainty quantification and real-time decision-making. The new surrogate processes three input modalities—a 3D geomodel, scalar relative permeability parameters, and control variables like stage durations and injection rates—through separate encoders, then fuses them via self-attention in a transformer encoder. A temporal decoder generates predictions auto-regressively using encoder-decoder cross-attention, capturing both saturation and pressure at monitoring points.
Trained on 4,000 GEOS simulations of a modified SEAM CO2 geomodel (with a faulted system and three stacked aquifers), the surrogate achieves a median saturation MAE of 0.028 and median relative errors of 0.2–5% for total injected and mobile CO2 mass and saturation footprints. It also learns to switch between rate and bottom-hole-pressure control, a key operational behavior. When integrated into a hierarchical Markov chain Monte Carlo data assimilation workflow, the model substantially reduced uncertainty for critical metaparameters, especially fault permeabilities, while keeping posterior predictions consistent with true model results. This makes the surrogate a practical tool for CO2 storage risk assessment and adaptive injection management.
- Fuses three modalities — 3D geomodel, relative permeability scalars, and well control variables — via transformer self-attention, enabling accurate joint modeling.
- Trained on 4,000 GEOS flow simulations, it achieves median saturation MAE of 0.028 and median relative errors of 0.2–5% for key CO2 storage quantities.
- Enables hierarchical MCMC data assimilation, cutting uncertainty for fault permeabilities and predicting storage behavior with significantly lower computational cost.
Why It Matters
This surrogate makes uncertainty-aware carbon storage modeling practical, enabling faster, safer CO2 sequestration decisions with fewer costly simulations.