Hybrid (mixed-matrix) polymer membranes offer powerful opportunities for wastewater treatment by embedding, for example, catalytically active inorganic particles within a continuous polymer matrix. Such materials can degrade trace micropollutants while the polymer phase governs transport and mechanical integrity. However, tailoring performance is difficult because formulation–processing–property relationships are high-dimensional, nonlinear, and thus complex.

By employing an ML-guided workflow based on Gaussian Process Regression (GPR), hybrid membranes with customized filtration performance and mechanical robustness can be engineered. GPR is a Bayesian, kernel-based, nonparametric model that (i) learns complex input–output mappings without assuming a fixed functional form, (ii) provides uncertainties alongside predictions, and (iii) remains data-efficient.

A small dataset of 36 membrane formulations described by five input variables was used to predict multiple targets, including filtration performance and mechanical properties (e.g., strength, modulus). The trained GPR models achieved predictive fidelity sufficient to guide materials selection across competing objectives. Feature relevance was interpreted via length scales. Using a simple, uncertainty-aware optimization scheme, candidate formulations were identified that, when synthesized, delivered filtration and mechanical properties comparable to commercial membranes.