Element-wise Representations with ECNet for Material Property Prediction and Applications in High-Entropy Alloys

When developing deep learning models for accurate property prediction, it is sometimes overlooked that some material physical properties are insensitive to the local atomic environment. Here, we propose the elemental convolution neural networks (ECNet) to obtain more general and global element-wise representations to accurately model material properties. It shows better prediction in properties like band gaps, refractive index, and elastic moduli of crystals. To explore its application on high-entropy alloys (HEAs), we focus on the FeNiCoCrMn/Pd systems based on the data of DFT calculation. The knowledge from less-principal element alloys can enhance performance in HEAs by transfer learning technique. Besides, the element-wise features from the parent model as universal descriptors retain good accuracy at small data limits. Using this framework, we obtain the concentration-dependent formation energy, magnetic moment and local displacement in some sub-ternary and quinary systems. The results enriched the physics of those high-entropy alloys.