Transfer Learning for Predicting Local Atomic Order in Multi-Principal Element Alloys
PI: Mia Jin (Nuclear Engineering)
Team Members: Yang Yang, Assistant Professor, Engineering Science and Mechanics. The PI will serve as mentor for the Junior Researcher.
This project aims to develop a machine learning framework that leverages transfer learning from binary alloy datasets to predict chemical short-range order (CSRO) in multi-principal element alloys (MPEAs), such as high-entropy alloys (HEAs), where data are scarce.
CSRO, the tendency for specific atomic pairs to prefer or avoid each other locally, strongly influences many macroscopic properties of materials, including strength, thermal stability, and corrosion resistance. Understanding and controlling SRO can be exploited to design advanced materials in energy, aerospace, and nuclear applications. However, direct experimental observation of SRO is extremely challenging due to the subtle nature of local atomic arrangements and limitations in spatial or spectral resolution of existing characterization tools.
Numerically, SRO can be predicted using DFT-based cluster expansions or Monte Carlo molecular dynamics simulations. However, such predictions require expensive computational cost or accurate interatomic potentials, which become increasingly difficult to obtain as the number of elements increases. Note that few robust potentials exist beyond 3-component systems. As a result, predictive modeling of SRO in MPEAs remains challenging with conventional methods.
To address this challenge, we propose to pretrain graph neural networks (e.g., CGCNN, SchNet) on large datasets of binary alloy configurations with known or easily computed SRO metrics, and then fine-tune these models using limited data from more complex MPEAs. By leveraging knowledge from well-studied binary systems, where reliable interatomic potentials and highquality data are more readily available, the model can learn to generalize atomic environments and make SRO predictions (e.g., Warren–Cowley parameters) in higher-order systems with sparse data. Finally, the model’s predictions will be validated against high-fidelity simulations or available experimental references.
If successful, this approach will enable efficient screening of vast alloy compositional spaces for targeted local ordering behaviors, which in turn are closely tied to improving key properties such as mechanical strength.
Planned Activities (Stage 1):
• Assemble datasets of SRO-labeled binary alloy configurations from DFT/cluster expansion archives
• Pretrain graph neural network models on binary systems
• Fine-tune models using small ternary/quaternary datasets (e.g., CoNiCr, NiCrFe)
• Evaluate model performance in predicting known SRO trends and comparing them with existing data
• Host biweekly research meetings with the ICDS Junior Researcher to co-develop the modeling pipeline
Desired Computational/Data Science Skills:
• Experience with graph neural networks (e.g., CGCNN, SchNet, DimeNet)
• Background Knowledge of transfer learning or fine-tuning ML models on small datasets
• Background knowledge with atomic structure data, SRO parameterization, or cluster expansion
• Python/NumPy/PyTorch, materials informatics tools (e.g., pymatgen, ASE)
Other Expectations of ICDS Junior Researcher:
• Regular availability for meetings (weekly or biweekly, times flexible)
Project Objectives:
• Build and validate a ML framework that predicts SRO from atomic configuration input
• Generate preliminary results to support an upcoming DOE proposal
• Submit a peer-reviewed paper
Medium to Long-Term Goal:
• Submit a full DOE proposal leveraging the transfer learning techniques for alloy discovery and/or optimization.
Connection to ICDS Mission:
This project directly supports the ICDS mission by advancing data- and AI-driven discovery in computational materials science and training junior researchers in interdisciplinary research combining ML and materials modeling.
Team Engagement with ICDS:
The PI has been engaging with ICDS (e.g., serving on the Cyberinfrastructure Faculty Advisory Committee) and sees this project as a gateway to broader collaboration. This effort will also contribute to establishing interdisciplinary research in future initiatives.