Accurate prediction of gas–brine Interfacial Tension (IFT) is essential for multiphase flow simulations in porous media because IFT governs capillary pressure scaling, trapping, and capillary number effects that control injectivity and plume migration [1]. This is central to CO2 sequestration and underground hydrogen storage, as well as pressure maintenance and enhanced oil recovery [2]. Many existing IFT correlations are calibrated on narrow datasets with limited brine chemistry, reducing transferability across gases and operating conditions [3]. We present a thermodynamics guided, data driven IFT closure that couples a harmonised literature database with an interpretable sparse regression model. The database contains 9,886 IFT records, of which 9,461 are retained as quality controlled “OK” entries and 425 are flagged by an explicit taxonomy, including incomplete brine chemistry, duplicates, and declared erroneous entries. The database spans approximately 271 to 573 K and 0.6 to 2800 bar and includes CO2, H2, CH4 and N2 systems, including mixtures. The workflow is designed for simulation use and requires only four user inputs: pressure, temperature, gas composition, and brine composition. Additional predictors are generated in a backend pipeline that maps heterogeneous salt and ion reporting into consistent concentration bases and chemistry descriptors, while gas composition is mapped to thermophysical descriptors and reduced variables at the specified state. To avoid dependence on expensive laboratory measurements, a density contrast descriptor is constructed from calculated phase properties rather than measured phase densities. A Lasso regressor is adopted as the primary predictor to maximise interpretability and deployment simplicity. In the standard configuration, Lasso retains 19 nonzero coefficients out of 31 engineered predictors, yielding a compact closure over physically meaningful descriptors. Group aware evaluation on a held-out test split yields RMSE 8.80 mN/m with approximately 1,800 samples, with improved robustness under a 99th percentile trimmed error analysis. Benchmarking against published correlations shows that accuracy alone can be misleading when correlations apply to small subsets, and that applicability coverage should be reported alongside errors for transferable IFT closures in porous media modelling. interfacial tension; porous media; multiphase flow; CCUS; hydrogen storage; brine chemistry; thermodynamics guided machine learn