About this Abstract |
| Meeting |
2025 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2025)
|
| Symposium
|
2025 Annual International Solid Freeform Fabrication Symposium (SFF Symp 2025)
|
| Presentation Title |
Train Small, Predict Big: Scalable Thermal Surrogate Modeling for LPBF Using MODAL |
| Author(s) |
Berkay Bostan, Praveen Vulimiri, Dhruba Aryal, Albert To |
| On-Site Speaker (Planned) |
Berkay Bostan |
| Abstract Scope |
Complex metal parts can be printed using Laser Powder Bed Fusion (LPBF), but the process is prone to defects due to rapid, localized thermal fluctuations. While in-situ monitoring and detailed simulations provide valuable thermal insights, they are often impractical due to hardware constraints and high computational cost. This work presents MODAL (MODular Assembly Learning), a scalable surrogate model for scanwise thermal simulations in LPBF. MODAL is trained on a small set of diverse geometrical blocks and uses randomized padding of nodes and heat sources to extend predictions to much larger, more complex parts. On average, test blocks are 14 times larger than the training blocks in spatiotemporal scale and include thermal histories across three consecutive melting events per node. MODAL achieves high accuracy—10% MAPE, 46.7 °C MAE, and 0.92 R²—while delivering substantial acceleration, with up to 357× speed-up on a single GPU and 1421× using four GPUs. |
| Proceedings Inclusion? |
Planned: Post-meeting proceedings |