Machine learning-based optimization and interpretation of the adsorption capacities of metal-organic frameworks for endocrine-disrupting compounds

Drinking-water contamination by endocrine-disrupting compounds (EDCs) demands process designs that are both effective and interpretable. We compiled a condition–response dataset ( n = 390) for five representative EDCs (carbamazepine, bisphenol A, 17α-ethinylestradiol, perfluorooctanoic acid, and ibuprofen) on aluminum-based MOFs, focusing on the optimization of adsorption conditions using machine learning (ML) models. Across 19 algorithms, tree-based ensembles (CatBoost, XGBoost, HistGradientBoosting) best predicted adsorption from experimental conditions alone (held-out R² typically > 0.95; up to 0.97). To couple prediction with mechanism, we integrated post-adsorption FT-IR spectra with the structured variables; a 1D-CNN on the FT-IR modality improved over unimodal baselines (test R² = 0.870). Grad-CAM highlighted chemically meaningful bands (i.e., 1580–1600, 1100–1250, and 870–880 cm⁻¹), consistent with π–π interactions, C-O stretching, and aromatic C-H bending. Observed adsorption capacities spanned 250.4 mg∙g⁻¹ across the five compounds, with CBM showing the highest uptake. Practically, the framework recommends MOF dosage and contact time under variable pH/ionic strength, supports re-use planning, and reduces trial-and-error in pilot design. To our knowledge, this is the first systematic comparison of 19 algorithms together with an FT-IR–augmented multimodal model for EDC adsorption on Al-MOFs, providing state-of-the-art accuracy and mechanism-aware interpretability for real-world water treatment.