About this Abstract |
| Meeting |
2024 TMS Annual Meeting & Exhibition
|
| Symposium
|
AI/Data Informatics: Computational Model Development, Verification, Validation, and Uncertainty Quantification
|
| Presentation Title |
Physics-constrained Bayesian Neural Networks to Predict Grain Evolution |
| Author(s) |
Luka Malashkhia, Dehao Liu, Anh Tran, Yan Wang |
| On-Site Speaker (Planned) |
Yan Wang |
| Abstract Scope |
Lack of training data is the major obstacle to applying machine learning tools to construct the surrogate models of process-structure-property relationships. Physics-informed neural networks were developed to tackle the data sparsity challenge by applying the physics-based models as the constraints to guide the training. In this work, we propose a physics-constrained Bayesian neural network to quantify the model-form and parameter uncertainties in neural networks. By taking advantage of the adaptive weight scheme and a new minimax architecture, the training convergence can be significantly improved in solving complex problems. The physical models of stochastic differential equations are utilized as the constraint. The new physics-informed neural network framework is used to predict rapid solidification and grain coarsening in metal additive manufacturing. |
| Proceedings Inclusion? |
Planned: |
| Keywords |
Additive Manufacturing, |