| About this Abstract |
| Meeting |
2023 TMS Annual Meeting & Exhibition
|
| Symposium
|
Accelerated Discovery and Insertion of Next Generation Structural Materials
|
| Presentation Title |
How Should You Select an Algorithm for a Materials Discovery Campaign with Multiple Objectives, Complex and High-dimensional Structure-processing-property Relationships, and a Small Adaptive Design Budget? |
| Author(s) |
Sterling G. Baird, Jeet Parikh, Trupti Mohanti, Taylor D. Sparks |
| On-Site Speaker (Planned) |
Sterling G. Baird |
| Abstract Scope |
Industry-relevant materials discovery tasks are often hierarchical, noisy, multi-fidelity, multi-objective, non-linearly correlated, and exhibit mixed numerical and categorical variables subject to linear and non-linear constraints. Examples include formulation optimization, compositional design of high entropy alloys, and multi-step synthesis. Choosing an algorithm that can expertly navigate such complex design spaces is a non-trivial task, and no single algorithm is supreme. So, how do you pair an algorithm to a design task? Here, we introduce PseudoCrab: a high-dimensional property predictor framed as a pseudo-materials discovery benchmark with fake compositional (linear) and "no-more-than-X-components" (non-linear) constraints. We apply a state-of-the-art high-dimensional Bayesian optimization algorithm (SAASBO) in conjunction with a multi-objective parallel Noisy Expected Hypervolume Improvement (qNEHVI) acquisition function and compare it against other high-performing models. Because PseudoCrab is customizable, researchers can adjust the PseudoCrab benchmark to more closely match their applications of interest during the algorithm downselection process prior to expensive materials discovery campaigns. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Machine Learning, Computational Materials Science & Engineering, |