Predicting photodegradation rate constants of water pollutants on TiO2 using graph neural network and combined experimental-graph features

Photocatalytic materials, which are known for their ability to harness solar energy to facilitate chemical reactions, are widely used in environmental remediation, including wastewater treatment. The efficiency of photocatalytic reactions in degrading pollutants is influenced by several factors, making parameter optimization time-consuming. In this context, machine learning techniques provide an appropriate solution for designing optimal photocatalysts. In addition to experimental variables such as solution pH and temperature, the molecular structure of the contaminant significantly affects the reaction efficiency. Molecular graphs are powerful tools for representing molecular structures for machine learning methods. This study integrates structural and five common experimental features in photocatalysis to develop graph neural networks (GNNs) for predicting the degradation rate constants of the organic water contaminants using TiO2. Three GNN models were developed: Graph Convolutional Network (GCN), Graph Attention Network (GAT), and a combined GAT-GCN model. Among these, the GAT performed better, achieving RMSE of 0.17, MAE of 0.13, and R2 of 0.90. The results demonstrate the effectiveness of integrating structural and experimental features, which can be further applied to predict other properties of the materials.