| About this Abstract |
| Meeting |
2025 TMS Annual Meeting & Exhibition
|
| Symposium
|
Microstructural Evolution and Material Properties Due to Manufacturing Processes: A Symposium in Honor of Anthony Rollett
|
| Presentation Title |
Modeling Microstructure Fatigue Indicator Parameters Using Symbolic Regression with Graph Neural Networks |
| Author(s) |
Jacob D. Hochhalter, Jonas Merrell, Krzysztof Stopka, Michael Sangid |
| On-Site Speaker (Planned) |
Jacob D. Hochhalter |
| Abstract Scope |
Data-driven (i.e., empirical) methods for model development have recently been a focus of the materials research community due to the many promising developments within the machine learning (ML) community. However, it is also known that empirical models often lack the ability to generalize across materials or processing methods, due in part to lacking rationale for the model form. Physics-informed ML methods can help regularize developed models but for some modeling scenarios, like fatigue indicator parameters, the physics regularization to be used is unknown. Additionally, microstructure-dependence inherently forms a high-dimensional ML modeling scenario. To help understand the generalizability of the ML model, we combine symbolic regression (SR) with graph neural networks (GNN). This hybrid approach enables an interpretable model for low-order microstructure features while the high-order, coupled features are captured with the GNN, thereby balancing accuracy and interpretability within the high-dimensional space. We present results of this ML approach for microstructure-dependent FIPs for use in additive manufacturing material qualification with uncertainty quantification. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Additive Manufacturing, Computational Materials Science & Engineering, Machine Learning |