Investigating the effect of structural variations on the band gap of titanium dioxide using machine learning

Abstract Predicting the band gap of titanium dioxide (TiO 2 ) is crucial for its photocatalytic applications, yet first-principles calculations of band gap properties are often computationally expensive. Machine learning offers an efficient alternative, but predictive accuracy strongly depends on how structural features are represented. In this work, we investigate the structure–property relationships governing the TiO 2 band gap by comparing different feature construction strategies. Traditional structure descriptors based on local atomic environments show limitations in effectively encoding complex structural correlations. In contrast, graph neural network (GNN) models more effectively learn structure–property relationships and achieve higher prediction accuracy. Further deconstruction of the model reveals that its predictive performance relies on a synergistic integration of chemical identity, local geometry, and non-local structural information. Feature visualization further shows that the model learns task-adaptive representations. These findings highlight the necessity of incorporating non-local structural information for predicting global properties and provide new insights into the machine learning driven materials design.