| About this Abstract |
| Meeting |
2025 TMS Annual Meeting & Exhibition
|
| Symposium
|
Verification, Calibration, and Validation Approaches in Modeling the Mechanical Performance of Metallic Materials
|
| Presentation Title |
Quantifying Error in Machine Learning Predictions of Macroscopic Yield Surfaces of Polycrystalline Materials |
| Author(s) |
Matt Kasemer, Lloyd van Wees, Karthik Shankar, Mark Obstalecki, Paul Shade |
| On-Site Speaker (Planned) |
Matt Kasemer |
| Abstract Scope |
In this study, we discuss the use of partial input convex neural networks (pICNN) to train constitutive models directly relating the crystallographic texture of polycrystalline samples to the resulting macroscopic yield surface. The pICNN models generate predictions at significantly reduced computational cost when compared to typical microstructurally sensitive modeling frameworks such as crystal plasticity finite element modeling. The use of reduced-order models, however, have the potential to introduce error at multiple points in the model development workflow. In addition to the error inherent in the predictions made by the trained model, such frameworks often rely on simplified material descriptions which may additionally introduce errors. Here, we will discuss both errors, and in particular the degree to which a reduced-order description of crystallographic texture (necessary to facilitate rapid data generation and pICNN training) introduces error in micromechanical predictions. Additionally, we will discuss both interpolative and extrapolative errors in the trained model. |
| Proceedings Inclusion? |
Planned: |