Leakage-aware machine learning for data-driven performance prediction of metal–organic framework systems
Abstract
Metal–organic frameworks (MOFs) are highly tunable porous materials whose performance is governed by complex interactions among structural, chemical, material, and operating variables. This study develops a leakage-aware, data-driven framework for predicting two distinct MOF performance endpoints: loading capacity and cell viability. The curated datasets comprised 161 loading-capacity observations described by 110 metal, ligand, functionality, and molecular-fingerprint variables, and 444 cell-viability observations described by 25 MOF, cell-category, and exposure variables. Histogram-based gradient boosting regression (HGBR) and partial least squares regression (PLSR) were optimized using differential evolution within five-fold grouped cross-validation. Exact duplicate records were constrained to the same data partition to prevent information leakage, while an additional source-held-out test was employed to assess transferability to publications not represented during model training. On the duplicate-safe 20% holdout, HGBR achieved the strongest predictive performance for cell viability ( R 2 = 0.774, RMSE = 11.655 percentage points, MAE = 8.153, and AARD = 16.69%), whereas PLSR performed best for loading capacity ( R 2 = 0.557, RMSE = 0.345 g/g, and MAE = 0.210 g/g). Permutation analysis identified concentration, Zn and Fe indicators, zeta potential, cell category, and particle size as the most influential variables for viability prediction, while PLSR loading predictions were most sensitive to 2,5-dioxidoterephthalate and Mg-related descriptors. However, negative R 2 values obtained when entire source publications were held out revealed substantial distribution shift and limited cross-study generalization. Overall, the results demonstrate the value of leakage-aware machine learning for rigorous, data-driven performance prediction and screening of MOF systems, while showing that broader standardized datasets and independent-study validation are essential before reliable predictive deployment.