Characterizing interfacial tension for hydrogen storage: An ensemble and deep learning approach

Le résumé fourni par la source

Underground hydrogen storage (UHS) in depleted reservoirs and saline aquifers represents a pivotal solution for large-scale energy storage and grid stabilization. A key parameter controlling multiphase flow, capillary trapping, and sealing efficiency in UHS systems is the interfacial tension (IFT) between injected gas mixtures and formation fluids. Given the high cost and complexity of experimental IFT measurements under reservoir conditions, predictive modeling offers a practical alternative. This study introduces a comprehensive framework for estimating IFT as a function of thermodynamic and compositional variables, including temperature, pressure, salinity, and the molar fractions of CO₂, CH₄, and H₂. Data preprocessing incorporated a robust outlier detection scheme using the ±2σ criterion to ensure dataset integrity. Eight predictive models spanning traditional, ensemble, and deep learning architectures were developed and validated via 5-fold cross-validation. Performance was assessed using R², MSE, and AARE%. Results highlight the superiority of tree-based ensemble methods, with AdaBoost and Random Forest achieving test R² values of 0.943 and 0.942, respectively, alongside low error rates (MSE=6.65 and 7.37; AARE%=2.72 and 2.83). While the Decision Tree achieved the highest training accuracy (R²=0.979), its generalization was weaker (MSE=9.06), underscoring the ensembles’ robustness. Deep learning models such as CNN demonstrated moderate predictive strength (test R²=0.730), whereas KNN suffered severe overfitting (test R²=0.410). SVR and MLP-ANN also showed limited generalization compared to ensemble methods. To enhance interpretability, SHapley Additive exPlanations (SHAP) were applied, revealing the relative importance of salinity, temperature, pressure, and gas composition in shaping IFT. The proposed framework, particularly the AdaBoost and Random Forest pipelines, provides an accurate, interpretable, and computationally efficient tool for predicting IFT under diverse reservoir conditions, thereby supporting the secure design and optimization of underground hydrogen storage projects.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.