Data-driven design of TiO2–zeolite photocatalysts for sustainable vanillin production

Abstract

We integrate machine learning (ML) with TiO₂–zeolite photocatalysis to convert biomass-derived 4-vinylguaiacol (4-VG) to vanillin via partial oxidation, thereby elucidating the dominant descriptor governing vanillin selectivity. Exploratory screening of a set of 16 zeolites and 12 TiO₂ photocatalysts revealed that composite systems enhance vanillin selectivity up to 36%, compared with 15% for TiO₂ alone. ML models such as random forest (RF), trained on 20 zeolite datasets (16 training and 4 test samples) incorporating six physicochemical descriptors, identified the Si/Al ratio of zeolite as the dominant descriptor. Validation experiments guided by ML predictions attributed its effect to pH buffering at the TiO₂ liquid interface, sustaining an alkaline microenvironment and stabilizing selectivity against pH drift. The synergy between ML analysis and ML-guided validation experiments proved effective even when applied to chemically curated datasets of limited size, providing transferable design principles for composite photocatalysts.