This study investigates transfer learning strategies for adapting machine-learned interatomic potentials (MLIPs) from medium-entropy to high-entropy alloy systems, focusing on the CrCoNi-CrCoNiFe-CrCoNiFeMn progression. Three transfer strategies were evaluated: feature freezing, progressive unfreezing, and multi-task learning. Progressive unfreezing showed the best performance, achieving a 66.6% improvement in energy prediction accuracy with only 50% of the target training data for the CrCoNi → CrCoNiFe transfer.
Key findings
High-entropy alloys (HEAs) are promising materials with exceptional mechanical properties.
MLIPs offer an alternative to first-principles calculations for computational modeling of HEAs.
Transfer learning strategies can accelerate MLIP development in HEA systems.
Progressive unfreezing achieves the best overall performance among the tested strategies.
Transfer effectiveness degrades with compositional distance, with 1-hop transfers outperforming 2-hop transfers.
Limitations & open questions
The study focuses on face-centered cubic transition metal alloys, limiting generalizability.
Real DFT data was not used for training, which could affect the transferability of findings.