| About this Abstract |
| Meeting |
MS&T22: Materials Science & Technology
|
| Symposium
|
AI for Big Data Problems in Advanced Imaging, Materials Modeling and Automated Synthesis
|
| Presentation Title |
Multi-property Graph Networks for Novel Materials Discovery |
| Author(s) |
Alexander New, Nam Le, Michael Pekala, Kyle McElroy, Janna Domenico, Christine Piatko, Elizabeth Pogue, Tyrel McQueen, Christopher Stiles |
| On-Site Speaker (Planned) |
Alexander New |
| Abstract Scope |
Machine learning (ML) approaches have the potential to accelerate material property prediction. Conventional approaches to materials discovery are expensive, but supervised ML models can rapidly screen large materials databases and identify candidates to test. When searching for a candidate, multiple properties will determine its relevance. Ideally, a single machine learning model could predict all desired properties for a given material. This is analogous to the ML concept of multi-task learning – by leveraging similarity between different prediction tasks, this single model will be better at prediction than a suite of property-specific models. We compare some state-of-the-art multi-task ML models to single-property models for predicting elastic material properties. These models often do not surpass single-property models, in contrast to existing findings in supervised learning. We present loss-surface curvature metrics that have the potential to explain this performance disparity, further suggesting research directions for multi-task ML to perform well for property prediction. |