Physics-Guided Machine Learning Using Dimensionless Groups to Predict CO2 Trapping Indices in Deep Saline Aquifers
Year: 2026
Type: dissertation
Abstract: Estimating the amount of CO2 trapped is essential for evaluating the long-term feasibility and security of geological carbon storage (GCS). This study presents a physics-guided machine-learning framework that uses dimensionless groups derived from reservoir and operational parameters to predict the temporal evolution of CO2 trapping mechanisms in heterogeneous saline aquifers. A conventional reservoir simulator was used to generate 200-year GCS simulation data, and these outputs were transformed into physically meaningful dimensionless inputs to improve model transferability and interpretability. Several machine-learning and deep-learning models were trained and compared for predicting supercritical, dissolved, residually trapped, and mineralized CO2 fractions. Among the evaluated models, TabM showed the best overall performance, achieving test-set R2 values of 0.957, 0.964, 0.942, and 0.930 for supercritical, dissolved, residually trapped, and mineralized CO2, respectively, with similarly strong 5-fold cross-validation R2 values of 0.943, 0.942, 0.933, and 0.938. Symbolic regression was further used to derive explicit closed-form equations for the trapping indices, which provides interpretable predictive equation. Shapley Additive Explanations (SHAP) were applied to quantify feature importance and identify the dominant controls for each trapping mechanism. Dissolved CO2 was governed mainly by brine salinity and injected pore volume, residual trapping was controlled primarily by residual gas saturation and storage time, and mineral trapping was influenced most strongly by storage duration and temperature. For supercritical CO2, injection period and brine salinity were found to be the main controls, while buoyancy-related effects were a secondary role. Beyond prediction, the TabM model was applied to Monte Carlo analysis, sensitivity analysis, and inverse-design optimization. High-performing cases were associated with intermediate-to-high PV injected, moderate dimensionless drawdown, moderate heterogeneity, limited residual gas immobilization, and relatively high gravity number, temperature, and salinity. These trends indicate that optimal mineral trapping is controlled by a balanced combination of flow, storage, and reservoir-condition parameters. Overall, the proposed framework provides a computationally efficient alternative to repeated full-physics simulations and offers a practical basis for rapid screening, sensitivity assessment, and recommended operating ranges for CO2 storage in heterogeneous saline aquifers.
Author: Bayanov, Ilya
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
SDG: Climate action
Open Access status: green
OpenAlex ID: W7198391248