AdvancedOptimized Deep-Learning Model for PreciseEvaluation of Subsurface Carbon Dioxide Trapping Efficiency

As global warming intensifies, geological carbon storage (GCS) in saline aquifers could play a vital role in mitigating CO₂ emissions. CO₂ trapping occurs mainly through solubility and residual trapping, requiring an accurate prediction using solubility trapping (STI) and residual trapping (RTI) indices. Machine learning shows promise for estimating CO₂ trapping in saline aquifers, but current models often lack effective feature selection, parameter optimization, and advanced deep learning techniques, limiting their performance. This study develops predictive models for RTI and STI using CNN, LSTM, and hybrid algorithms by combining them with growth optimization (GO) and cuckoo optimization (COA). An extensive data set of 6,811 global data points was analyzed, with feature selection using the nondominated sorting genetic algorithm and random forest analysis. Model performance was based on independent testing data, and Shapley additive explanation (SHAP) analysis identified key features. For RTI, residual gas saturation (RGS), injection rate (IR), permeability (Perm), elapsed time (Te), porosity (Por), and salinity (Sal) were the most influential. Conversely, RGS, thickness (Th), Te, Perm, Sal, and Por were most critical for STI. The results confirm that hybrid DL models outperformed standard DL models, with metaheuristic optimization enhancing accuracy and generalization. The CNN-COA model achieved the lowest root-mean-square error (RMSE) for RTI (0.0011 for training; 0.0035 for testing) and STI (0.0005 for training; 0.0028 for testing) predictions. SHAP analysis revealed RGS and Perm as the most and least influential features for RTI predictions and Th and Perm as the most and least influential features, respectively, for STI predictions. This study is innovative in its integration of advanced feature selection methods and hybrid deep learning algorithms with effective optimization and feature selection. This integration leads to improved GCS model prediction performance, robustness, and adaptability to diverse geological conditions.