About this Abstract |
| Meeting |
TMS Specialty Congress 2026
|
| Symposium
|
4th World Congress on Artificial Intelligence in Materials & Manufacturing (AIM 2026)
|
| Presentation Title |
Physics-Guided 1D Convolutional Neural Networks for Scalable Prediction of Thermo-Mechanical Response in Laser Powder Bed Fusion |
| Author(s) |
Ali Behbahani, David Fieser, Farzana Nasrin, Reza Abedi |
| On-Site Speaker (Planned) |
Ali Behbahani |
| Abstract Scope |
Laser Powder Bed Fusion (LPBF) involves highly transient thermal cycles that generate complex residual stresses and plastic deformation, limiting part qualification and real-time optimization. While high-fidelity thermo-mechanical finite element simulations provide accurate predictions, their computational cost restricts practical use. This work presents a physics-guided surrogate modeling framework that maps full nodal thermal histories directly to mechanical response quantities using a one-dimensional convolutional neural network (1D-CNN). Tree-based models and recurrent neural networks (LSTM and GRU) are evaluated as baselines. Although recurrent models achieve strong accuracy, they require substantially longer training times due to sequential evaluation. In contrast, the proposed 1D-CNN exploits localized temporal convolution to encode physically meaningful thermal gradients, enabling parallel training and improved generalization. Physics-based regularization enforcing equilibrium, yield consistency, and traction-free boundaries is incorporated using a curriculum learning strategy, resulting in superior accuracy with significantly reduced computational cost. |
| Proceedings Inclusion? |
Undecided |