About this Abstract |
| Meeting |
2026 TMS Annual Meeting & Exhibition
|
| Symposium
|
Artificial Intelligence Applications in Integrated Computational Materials Engineering (AI-ICME)
|
| Presentation Title |
Physics-Constrained Neural Network for Increased Generalizability in Predicting Material Microstructure Evolution |
| Author(s) |
Benjamin Rhoads, Lars Kotthoff, Samrat Choudhury |
| On-Site Speaker (Planned) |
Benjamin Rhoads |
| Abstract Scope |
Neural networks are powerful surrogates for materials science models, most effective for interpolating between processing parameters, material properties, and simulation time. However, because of their lack of physics awareness, they often fail to extrapolate beyond the training data. While physics-informed neural networks (PINNs) address this through soft constraints in the loss function based on known equations, an alternative is physics-constrained neural network (PCNN), which enforces physical laws directly in the model’s architecture. In this study, we apply custom physics constraints to train a network that predicts the evolution of nickel-aluminum microstructures generated by a phase field model. By embedding mass conservation, periodic boundary conditions, and a power law equation for precipitate growth, we improve the model’s ability to generalize beyond the training window while also reducing computational cost. This demonstrates that incorporating physical principles into neural network design can lead to more efficient and generalized predictions of microstructure evolution. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Computational Materials Science & Engineering, Machine Learning, Modeling and Simulation |