| About this Abstract |
| Meeting |
MS&T24: Materials Science & Technology
|
| Symposium
|
Additive Manufacturing Modeling, Simulation, and Machine Learning: Microstructure, Mechanics, and Process
|
| Presentation Title |
Self-Supervised Feature Distillation and Design of Experiments for Efficient Training of Micromechanical Deep Learning Surrogates |
| Author(s) |
Patxi Fernandez-Zelaia, Jiahao Cheng, Jason Mayeur, Yousub Lee, Kevin Knipe, Kai Kadau |
| On-Site Speaker (Planned) |
Patxi Fernandez-Zelaia |
| Abstract Scope |
Computationally efficient reduced order models (ROMs) are needed in engineering design and optimization tasks. In micromechanics it is the spatial full-field response which is desired and a great deal of work has thus far focused on establishing machine learning (ML) based ROMs. In comparison relatively little work has focused on establishing microstructural experimental design strategies. Here we show that strategic selection of microstructural volume elements for evaluating physics models enables the establishment of more accurate ROMs. The two key challenges towards constructing efficient experimental designs are: (1) microstructural feature quantification and (2) establishment of a criteria that selects diverse training data. For a specific problem considered an improvement of up to 8\% in surrogate performance may be achieved. For larger problems trends indicate this approach may be even more beneficial. These results illustrate that selection of an efficient experimental design is essential when establishing ML-based micromechanical ROMs. |