Crystal Fractional Graph Neural Network Predicts High-Entropy Alloy Energy

A new machine learning model, the crystal fractional graph neural network (CFGNN), has been developed to predict the energy of high-entropy alloys (HEAs). HEAs are materials composed of multiple principal elements in near-equimolar ratios, known for their exceptional mechanical and thermal properties. The CFGNN integrates local atomic environments and global compositional information through three components: a crystal graph neural network using graph attention layers to learn local interactions among 16 on-site atoms; a fractional neural network that embeds global fractions of constituent elements; and a feature fusion network that combines outputs to predict total crystal energy. The model was trained on 1,049 crystal structures and validated on 198 quaternary structures, with hyperparameters optimized via Optuna. The study is published on arXiv.