Multimodal Auto-regressive Transformer Surrogate for Modeling Variable Operations and Quantifying Uncertainty in Geological Carbon Storage

Storing carbon dioxide underground is a key strategy for reducing atmospheric emissions, but the success of each project hinges on how well injection wells are managed. Variable perforation schedules and changing injection rates can improve sweep efficiency, yet they also introduce complexity that makes traditional simulators expensive to run for many scenarios. A new surrogate model addresses this challenge by learning the relationship between well‑operation controls, geological uncertainty, and the resulting CO₂ plume behavior.

What You Need to Know

The researchers built a multimodal auto‑regressive transformer that takes three distinct types of input: a 3‑D geological model of the storage formation, scalar parameters that describe relative permeability functions, and a set of control variables defining well perforation stage durations and injection rates for each well. Each modality is processed by its own encoder, and the resulting representations are fused through self‑attention layers within the transformer. The model then predicts, step by step, the evolution of pressure and CO₂ saturation fields, providing both a mean forecast and an estimate of uncertainty.

The surrogate was trained and tested on a modified SEAM CO₂ geomodel that features a faulted system with three stacked aquifers. Two injection wells are perforated sequentially from bottom to top; the timing of each perforation stage and the injection rate during that stage serve as the controllable inputs. By treating these as variables, the surrogate can explore a wide operational space without re‑running a full‑physics simulator for every combination.

Because the transformer is auto‑regressive, it generates predictions sequentially in time, allowing the model to capture temporal dependencies between earlier injection decisions and later plume migration. Uncertainty quantification is achieved by propagating the variability in the geological model and permeability parameters through the network, yielding predictive distributions rather than single deterministic outputs.

Why It Matters

Accurate, fast forecasts enable operators to test many perforation and rate schedules in the time it would take a single high‑fidelity simulation to run. This capability supports real‑time decision‑making, such as adjusting injection rates to avoid pressure buildup near faults or to improve sweep of poorly connected aquifers. Consequently, projects can achieve higher storage efficiency while reducing the risk of unintended leakage or induced seismicity.

Beyond operational gains, the surrogate’s uncertainty estimates give stakeholders a clearer picture of confidence in predicted outcomes. When evaluating regulatory compliance or economic viability, knowing the range of possible plume extents helps in designing monitoring plans and setting appropriate safety margins. The approach also demonstrates how modern machine‑learning architectures can be tailored to the multimodal nature of subsurface problems without sacrificing physical interpretability.

Key Details

• Modified SEAM CO₂ geomodel: faulted architecture with three vertically stacked aquifers.
• Two injection wells perforated in stages from bottom to top; stage duration and well‑specific injection rate are control inputs.
• Three input modalities processed by separate encoders: 3‑D geomodel (voxel or mesh), scalar permeability parameters, and control variables.
• Encoders fused via self‑attention layers in a transformer architecture.
• Auto‑regressive output predicts pressure and CO₂ saturation fields sequentially in time.
• Uncertainty quantified by propagating geological and permeability variability through the model, yielding predictive distributions.
• Training dataset generated from a limited set of high‑fidelity simulator runs covering the control‑variable space.
• Validation shows the surrogate reproduces simulator statistics with errors within a few percent for key metrics such as plume extent and pressure rise.

What’s Next

The team plans to extend the surrogate to more complex well configurations, including multiple injection and monitoring wells, and to incorporate additional data sources such as time‑lapse seismic or pressure measurements for online model updating. Another direction is to couple the surrogate with optimization algorithms that directly seek perforation schedules maximizing stored mass while respecting operational constraints. Finally, field‑scale testing on pilot sites will assess how well the model’s uncertainty estimates align with observed behavior, paving the way for broader adoption in carbon‑storage project workflows.

📌 Source: Arxiv Ml