About this Abstract |
| Meeting |
2027 TMS Annual Meeting & Exhibition
|
| Symposium
|
Algorithms Development in Materials Science and Engineering
|
| Presentation Title |
Implementation of Bayesian Optimization over Function Networks with Partial Evaluations for Cost-Aware AI Guided Material Discovery of Critical Mineral Lean Aluminum Alloys |
| Author(s) |
Xavier Zaitzeff, Sterling Baird |
| On-Site Speaker (Planned) |
Xavier Zaitzeff |
| Abstract Scope |
Discovering substitutes to standard additive manufacturing aluminum alloys is a high-dimensional, high-fidelity, multi-objective problem with several steps from simulation to testing material properties. To optimize the element composition parameters and printing parameters for laser powder bed fusion systems, we leverage a directed, acyclic network structure to implement Bayesian Optimization over a function network. Leveraging the structure enables smaller Gaussian Processes (GP) to more accurately model each subprocess. We incorporate SAASBO, a surrogate model better attuned to higher dimensional problems. We use Partial Knowledge Gradient over Function Networks (pKGFN) as the acquisition function to select the next optimal candidate to test. pKGFN implements cost-benefit analysis of running each subprocess so it will suggest both a new input and subprocesses to test. We adapt these algorithms to a multi-objective setting to optimize tensile strength and hardness. We demonstrate results on preliminary synthetic functions and early campaign results. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, Machine Learning, Modeling and Simulation |