About this Abstract |
Meeting |
2023 TMS Annual Meeting & Exhibition
|
Symposium
|
AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification
|
Presentation Title |
M-9: Loss Curvature-informed Multi-property Prediction for Materials and Chemicals via Graph Neural Networks |
Author(s) |
Alex New, Michael J Pekala, Nam Q Le, Janna Domenico, Christine D Piatko, Christopher D Stiles |
On-Site Speaker (Planned) |
Alex New |
Abstract Scope |
Properties of interest for crystals and molecules, such as band gap, elasticity, and solubility, are related to each other: they are determined by the same underlying physics. However, when state-of-the-art graph neural networks predict multiple properties simultaneously via multi-task learning, they frequently underperform a suite of single property predictors. This suggests graph networks may not be fully leveraging these underlying property similarities. Here we investigate a potential explanation for this phenomenon – the curvature of each property’s loss surface significantly varies, leading to ineffective learning. This difference in curvature can be assessed by looking at spectral properties of the Hessians of each property’s loss function, which is done in a matrix-free manner via randomized numerical linear algebra. We evaluate our hypothesis on two benchmark datasets (Materials Project for crystals and QM8 for molecules) and consider how these findings can inform the training of novel multi-task learning models. |
Proceedings Inclusion? |
Planned: |
Keywords |
Machine Learning, Computational Materials Science & Engineering, Mechanical Properties |