About this Abstract |
| Meeting |
MS&T26: Materials Science & Technology
|
| Symposium
|
Additive Manufacturing Modeling, Simulation, and Machine Learning: Microstructure, Mechanics, and Process
|
| Presentation Title |
Fast Prediction of Thermal History in Powder Bed Fusion Through Autoregressive Transformer-Based Diffusion |
| Author(s) |
David Guirguis |
| On-Site Speaker (Planned) |
David Guirguis |
| Abstract Scope |
Accurate prediction of thermal history in powder bed fusion is critical for understanding defect formation, residual stress development, and part-scale variability. However, high-fidelity thermal simulations remain computationally expensive, especially for complex geometries and scan paths. This computational burden limits the use of thermal modeling during process planning.
This work presents a fast surrogate modeling framework for predicting thermal history using an autoregressive transformer-based diffusion approach. The proposed method treats thermal evolution as a sequential spatiotemporal problem, where the thermal state of each layer depends on the accumulated history of previous layers, local geometry, and scan strategy. A transformer module captures long-range dependencies across layers and spatial regions, while the diffusion component learns to reconstruct high-resolution thermal fields from compact latent representations. By autoregressively propagating thermal information from one layer to the next, the model predicts transient thermal accumulation and localized overheating patterns without requiring full numerical simulation at every step. |