About this Abstract |
| Meeting |
2026 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2026)
|
| Symposium
|
2026 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2026)
|
| Presentation Title |
Hierarchical Graph Neural Networks with Local Attention for Accelerated Part-Scale Scanwise Thermal Simulation in Laser Powder Bed Fusion |
| Author(s) |
Berkay Bostan, Albert To |
| On-Site Speaker (Planned) |
Albert To |
| Abstract Scope |
Scanwise thermal simulation in laser powder bed fusion (LPBF) is critical for defect prediction and process optimization but remains computationally prohibitive due to extreme spatial (micron-scale laser vs. millimeter–centimeter parts) and temporal (microsecond time steps) disparities. This work presents an accelerated framework for part-scale, scanwise thermal simulation of 3D components. The approach integrates two key modules: (1) a hierarchical graph encoder that captures both local and global geometric influences on thermal behavior and boundary conditions, and (2) a transformer-based decoder with local temporal attention that predicts nodal thermal histories using encoded geometry and time-dependent inputs such as laser trajectories and time-step sizes. The framework accurately handles complex geometries, including overhangs and in-plane variations, across simulations involving millions of nodes and time steps. It achieves approximately two orders of magnitude computational speedup while maintaining low prediction error, enabling efficient and scalable thermal modeling for realistic LPBF applications. |
| Proceedings Inclusion? |
Undecided |