Tabular Deep Learning for Hydrogen–Brine Interfacial Tension Prediction with Uncertainty Quantification
Energy & Fuels · 30.07.2026 · 40 (30) · 16671-16691
DOI: 10.1021/acs.energyfuels.6c01887 ↗Kurz gefasst
Fachartikel in der Zeitschrift Energy & Fuels, online erschienen am 16.07.2026. Verfasst von Peng-Cheng Jiao und 3 weiteren Autor:innen. Die Arbeit nennt 48 Literaturangaben.
Abstract (Originalsprache)
Underground hydrogen storage (UHS) in geological formations is pivotal for large-scale energy buffering. However, its feasibility and operational safety rely heavily on the accurate characterization of interfacial tension (IFT) between hydrogen–cushion gas mixtures and formation brines. While experimental measurements provide essential baseline data, existing historical data sets often struggle to represent the complex geochemistry of subsurface fluids, particularly multicomponent mixed brines, and are frequently affected by interlaboratory systematic noise arising from diverse measurement protocols. To overcome these limitations, this study presents a robust feature engineering and tabular deep learning framework developed from a rigorously curated database of 450 experimental IFT measurements. The unified feature space incorporates thermodynamic conditions (p, T), mole fractions of cushion gases (CO2, CH4), specific molalities of individual salts (NaCl, KCl, CaCl2, MgCl2), and a categorical experimental method indicator to explicitly account for interlaboratory systematic bias. Two state-of-the-art tabular deep learning architectures, Tabular Model (TabM) and Tabular Prior-data Fitted Network (TabPFN), were systematically evaluated. Among them, TabPFN delivered the highest predictive accuracy on the testing data sets, achieving an R2 of 0.9941, RMSE of 0.92 mN/m, and AARD of 1.15%. The model further demonstrated good transferability on completely independent external data sets, yielding R2 values of 0.9923 and 0.9841 for H2–pure water and H2–NaCl systems, respectively, while reproducing the macroscopic IFT trends of an unseen quaternary brine system (NaCl+KCl+CaCl2+MgCl2). Shapley additive exPlanations (SHAP) and partial dependence analyses confirmed the physicochemical consistency of the framework, identifying temperature and CO2 mole fraction as the dominant IFT-reducing factors while highlighting the critical role of the methodological feature in mitigating systematic measurement noise. Furthermore, quantile regression was employed to generate calibrated 95% prediction intervals, reducing uncertainty widths by 33% after incorporating the methodological feature. Finally, the probabilistic IFT predictions were integrated into a theoretical assessment workflow for a Saudi Arabian geological formation, enabling conservative estimation of capillary entry pressure, maximum hydrogen column height, and storage capacity under uncertainty. This work provides an interpretable and risk-aware predictive framework capable of capturing complex geochemical interactions and methodological variability, thereby offering a rigorous scientific basis for the safe design and optimization of UHS operations.
Zitieren
APA
Jiao, P., Bian, X., Wang, L., & Chen, J. (2026). Tabular Deep Learning for Hydrogen–Brine Interfacial Tension Prediction with Uncertainty Quantification. Energy & Fuels, 40(30), 16671-16691. https://doi.org/10.1021/acs.energyfuels.6c01887
Vancouver
Jiao P, Bian X, Wang L, Chen J. Tabular Deep Learning for Hydrogen–Brine Interfacial Tension Prediction with Uncertainty Quantification. Energy & Fuels. 2026;40(30):16671-16691. doi:10.1021/acs.energyfuels.6c01887
BibTeX
@article{Jiao2026,
author = {Jiao, Peng-Cheng and Bian, Xiao-Qiang and Wang, Lian-Guo and Chen, Jing},
title = {Tabular Deep Learning for Hydrogen–Brine Interfacial Tension Prediction with Uncertainty Quantification},
journal = {Energy & Fuels},
year = {2026},
volume = {40},
number = {30},
pages = {16671--16691},
doi = {10.1021/acs.energyfuels.6c01887}
}
EUVIMED