Fine-Tuning Foundation Models to Unveil Inhibitory Dynamics of Organic Cosolutes on TiO₂ Photocatalysts

The efficient utilization of solar energy for water treatment through photocatalytic processes has been limited by the challenge of understanding and optimizing interactions at the photocatalyst surface, particularly in the presence of non-target cosolutes. This study employs first-principles methods to predict the inhibitory effects of a series of small organic molecules during TiO₂ photocatalytic degradation of <i>para</i>-chlorobenzoic acid (pCBA). Tryptophan, coniferyl alcohol, succinic acid, gallic acid, and trimesic acid were chosen as interfering agents against <i>pCBA</i> to examine the resulting competitive reaction kinetics through bulk and surface phase reactions, following Langmuir-Hinshelwood adsorption dynamics. Fine-tuning models pretrained on millions of configurations with a relatively small dataset yielded machine learning interatomic potentials (MLIPs) achieving chemical accuracy. These models enabled extended timescale simulations that offered deeper insights into adsorption behaviors and inhibitory effects. Our simulations indicated that trimesic and gallic acids exhibited the highest inhibition, confirmed by experimental results. Further, a strong correlation between solute-surface interaction energies and experimental observations indicates that adsorption site interactions overshadow the role of general reactivity with OH radicals. ML-accelerated explicit solvation simulations showed that water molecules saturate the anatase active sites, leading to primary competitive adsorption between inhibitory cosolutes and water. These findings demonstrate the potential of ML-driven approaches to optimize photocatalytic systems for more effective water treatment applications by enhancing our understanding of surface inhibitory dynamics.