Cluster expansion augmented transfer learning for property prediction of high-entropy alloys

Abstract

We propose a deep learning model that combines transfer learning with a cluster expansion method for accurately predicting the physical properties of high-entropy alloys (HEAs). We utilize a small amount of first-principles data and generate a large amount of quasi-first-principles background data through cluster expansion, and then transfer the physical knowledge therein to the prediction of HEAs through transfer learning, thus overcoming the challenge of insufficient first-principles data. When applied to the FeNiCoCrMn/Pd high-entropy alloy systems, the model shows enhancement in the prediction accuracy of formation energy and average atomic magnetic moment. By transferring the physical information from low-component alloy data to the prediction of multi-component alloys, we improve the predictive capabilities of the model for multi-component alloys, reducing the prediction RMSE from 0.011 to 0.008 eV per atom for formation energy and from 0.133 to 0.090μ_B per atom for atomic magnetic moment. The transfer learning model maintains high accuracy even at a small dataset limit where only 20% of the multi-component alloy data is retained. Moreover, we interpret the prediction results of the transfer learning model for the physical properties of HEAs using the effective cluster interactions (ECIs) of cluster expansion, demonstrating that the cluster expansion model contains transferable physical knowledge. This validates the reliability of the transfer learning model from a physical perspective. The proposed framework combines the efficient data utilization of transfer learning with the clear physical insights of cluster expansion, augmenting the predictive capability of machine learning on small datasets.