Attention-based composition learning for thermodynamic and electronic property prediction in inorganic materials

Consistent prediction of material properties such as formation energy and electronic bandgap is very important for accelerating materials discovery in the data-driven era. Although recent machine learning and deep learning models have shown promising results, their performance mainly depends on the quality of the dataset and the effective capture of complex composition–property relationships. In this study, we present an integrated framework that addresses these difficulties through three main contributions. First, a curated dataset was built by collecting formation energies and band gaps obtained from density functional theory (DFT) calculations for various inorganic compositions. Special care was taken during the dataset curation process to preserve physical plausibility and statistical diversity. Second, we developed a lightweight, composition-sensitive multi-head attention-based network model to capture element-level interactions directly from compositional information without requiring explicit crystal structure data for material property prediction. It also performs consistently well on the proposed dataset. Third, we evaluate the proposed model on the AFLOW benchmark dataset. The results demonstrate that CompAttNet achieves competitive predictive performance on both the curated dataset and the AFLOW benchmark while maintaining a lightweight composition-based architecture that requires only chemical composition as input.