Systematic Comparison of SrTiO3 Loading Methods Combined with Machine Learning for High-Performance Photocatalytic Membrane Reactors in Cephalexin Removal

• Photocatalytic membranes fabricated by 3 different SrTiO 3 catalyst loading Methods. • SrTiO 3 loading by reducing MWCO and creating surface charge, boost separation. • Removal increased from 20% for raw membrane to 86% for SA-STO 3 under UV irradiation. • The experimental removal value was 86% SA-STO 3 , while ML predicted value was 84%. • After 390 min, SA-STO₃ achieved flux, removal and FRR of 160 LMH, 91%, and 81%. Despite considerable progress in photocatalytic-membrane-reactors (PMs) for antibiotic removal, the influence of catalyst immobilization strategies and surface properties on membrane performance remains insufficiently understood. In this study, SrTiO₃-based PMs were developed for cephalexin removal through combined molecular-sieving, Donnan-repulsion, and photocatalytic-degradation. SrTiO₃ was immobilized using three approaches: poly(acrylic-acid) polymerization (PAA-STO i ), polyamide polymerization (PA-STO i ), and sodium-alginate-deposition (SA-STO i ). The novelty of this study lies in identifying the most effective method for maximize PMs performance. Structural and elemental analyses confirmed the uniform distribution of SrTiO₃ on the membrane surface. Thermogravimetric-analysis showed improved thermal stability, while leaching tests confirmed negligible catalyst release. For PA-STO 3 , the BJH-pore-size, zeta-potential, and reaction-rate-constant were 1.217 nm, −28.9 mV, and 0.028 min⁻¹, reflecting the synergistic effects of membrane-separation and photocatalysis-degradation. Further, feature-importance-analysis obtained by machine-learning, revealed that MWCO (key indicator of membrane-separation), and lamp-distance from membrane surface (key factor in photocatalytic-degradation) significantly affect removal, with a contribution shares of 33 and 22%. As a result, removal increased from 20% for the raw-membrane under dark-conditions to 78% under UV-irradiation for PA-STO 3 . Machine-learning models also accurately predicted membrane performance, with predicted removal of 79%, 78%, and 85% for PAA-STO 3 , PA-STO 3 , and SA-STO 3 , closely matching experimental-results (78, 80, and 85%). The PMs also exhibited self-cleaning behavior, maintaining stable flux and fouling resistance; SA-STO₃ achieved 160 LMH flux, 91% removal, and 81% FRR after 390 min of operation. This work demonstrates a novel approach by presenting a systematic strategy for optimizing PMs through identifying the most effective features for antibiotic removal, supported by machine-learning.