Physics-guided machine learning using engineered dimensionless parameters to predict CO2 trapping indices in heterogeneous carbonate saline aquifers

Ilya Bayanov, Irina Spivakova, Peyman Pourafshary, Masoud Riazi

2026

Full text Open access: Hybrid Metadata (CSL-JSON)
DOI 10.1016/j.jgsce.2026.205985
Data confidence 85 high
How this score is calculated

It is built from checkable properties of the record itself: whether the article has a DOI, how many independent sources confirmed its data, whether the required fields are filled in, and how recent the latest confirmation is. Disagreement between sources and a missing identifier lower it. A value of 80 or above means the record was published without further checking; 60 to 79 means it is published at reduced confidence, with a manual check advisable. A record verified by an editor by hand is never pushed below 90 by a recalculation. This scores our data about the article, not the article itself.

Abstract

Estimating the amount of CO 2 trapped is critical for assessing the feasibility of long-term geological carbon storage (GCS). This study introduces a novel approach that uses engineered dimensionless features to develop a machine learning (ML) framework. A traditional reservoir simulator was used to generate 200-year GCS simulation data. Reservoir and operational parameters were used to derive engineered dimensionless parameters, which were then used as inputs. Based on these dimensionless parameters, the ML model predicts CO 2 trapping indices. Six ML models were trained and evaluated - Random Forest (RF), CatBoost, XGBoost, Light Gradient Boosting Machine (LightGBM), Gradient Boosting Regressor (GBR), and symbolic regression (SR). CatBoost outperformed other models, with R 2 values of 0.94 for supercritical and dissolved CO 2 , and 0.92 for residually trapped and mineralized CO 2 . In addition, the SR model was used to produce closed-form equations for the trapping indices. The SR-derived formulas also demonstrated reasonable accuracy, with test-set R 2 values ranging from approximately 0.87 to 0.92. Shapley Additive Explanations (SHAP) were also used to measure feature importance, and the main predictors for each trapping mechanism were identified. For solubility trapping, brine salinity and injected pore volume were the dominant factors, together contributing around 60% of the impact on the output, whereas residual trapping was governed primarily by residual gas saturation and storage time, with a combined contribution of about two-thirds. For mineral trapping, storage duration and temperature were the dominant features. The analysis also indicated that the supercritical CO 2 fraction was controlled mainly by elapsed storage time and brine salinity, while gravitational effects played a secondary role. In summary, the proposed framework indicates a computationally efficient alternative to full-physics simulations and is suitable for rapid screening and sensitivity analysis within the modeled class of heterogeneous carbonate saline-aquifer systems.