Real-time prediction and inverse design of multiphase CO2 trapping in deep saline aquifers using machine learning enhanced by SHAP analysis

Abstract

Accurate forecasting of multiphase CO₂ behavior is critical for safe large-scale deployment of carbon capture and storage projects. A total of 9835 high fidelity CMG GEM simulations, covering geological and operational uncertainties in deep saline aquifers, were generated and subsequently used to benchmark thirteen c algorithms for the concurrent prediction of CO₂ trapping mechanisms. Attention based tabular networks TabTransformerLite and SAINTLite together with the decision ensemble NODELite outperformed linear, support vector, tree ensemble, and fully connected neural network baselines. TabTransformerLite secured the best performance with nRMSE of 0.033, R² of 0.92, and minimal fold to fold variability. Shapley value analysis placed depth and permeability as primary controls on dissolved, residual, and mineral precipitated CO₂ compared to the injection rate and porosity. An inverse design workflow used the trained model to identify operating windows that lift mineral trapping above the ninetieth percentile, indicating optimal depth near 2.25 km, permeability above 1.6 Darcy, and injection rates between 7.5 and 14.7 × 10³ m³/day. The resulting framework connects model predictions to geographic information systems for site screening and to real time digital twin optimization, providing a scalable millisecond speed alternative to physics-based simulation for CO₂ storage planning and monitoring.