About this Abstract |
| Meeting |
MS&T25: Materials Science & Technology
|
| Symposium
|
Materials Processing and Fundamental Understanding Based on Machine Learning and Data Informatics
|
| Presentation Title |
Machine Learning Disordered Materials Properties |
| Author(s) |
Hengrui Zhang, Jie Chen, James M. Rondinelli, Wei Chen |
| On-Site Speaker (Planned) |
Wei Chen |
| Abstract Scope |
Machine learning (ML) significantly advances materials discovery, yet the vast combinatorial space of potential materials – arising from diverse constituents and flexible atomic-scale configurations – poses a major hurdle. This challenge is particularly acute in disordered systems like molecular mixtures (e.g., battery electrolytes) and high-entropy alloys (HEAs), where conventional ML methods often falter. To address this, we present MolSets, an ML model tailored for these complexities. MolSets leverages graph neural networks (GNNs) for detailed molecular-level or local environment representations and a deep sets architecture for mixture-level aggregation, thus capturing local chemical intricacies while retaining global configurational flexibility.
We demonstrate MolSets by effectively predicting lithium battery electrolyte conductivity, validated experimentally. Applied to HEAs, it yields property predictions with good performance and interpretability. This framework extends GNN applicability to disordered materials, showing potential for autonomous synthesis and contributing to accelerated discovery within complex combinatorial spaces. |