STEAM: A scalable data-driven surrogate modeling framework for part-scale scanwise thermal process simulation of laser powder bed fusion

Simulating scanwise thermal behavior at part scale in laser powder bed fusion (LPBF) is essential for understanding defects and optimizing process parameters, but it remains prohibitively expensive for real engineering parts. While finite element method (FEM) based process simulations provide valuable insights, the steep computational cost limits their routine application. This work introduces a scalable data-driven surrogate modeling framework for scanwise LPBF thermal simulation through deep learning called STEAM (Scanwise Thermal Emulator for Additive Manufacturing) to predict the transient thermal history of complex parts. The key features behind the proposed framework include 1) a compact dataset constructed from geometrically distinct modular blocks capturing representative thermal patterns, and 2) a combination of temporal and spatial features employed as model inputs to capture key physical phenomena influencing thermal behavior. Temporal features describe how each node interacts with the moving heat source over time, while spatial features capture the effect of the surrounding geometry on local heat conduction and boundary conditions. The model is validated on 2.5D test geometries as much as ~10 times larger than the training cases with different geometric complexity. On large-scale test blocks, STEAM achieves an R² of 0.97, and 29.61 °C mean absolute error, with a 1237–3648× speedup compared with FEM using four GPUs (i.e., 309–912× per GPU). Importantly, performance improves with increasing model complexity, which is promising for further scaling STEAM to larger and more complex LPBF domains. These results demonstrate that the STEAM framework provides a promising path toward fast and accurate thermal prediction in real-world LPBF parts. • A scalable data-driven surrogate model for part-scale scanwise thermal process simulation of LPBF is presented. • The training set consists of compact yet geometrically diverse blocks that capture representative heat transfer patterns. • Temporal features describe node–heat source interactions, while spatial features reflect the influence of surrounding geometry on local heat conduction and boundary conditions. • The model is evaluated on unseen test blocks that are as much as ~10 times larger than the training blocks. • The model predicts thermal histories with an R² of 0.97 across millions of nodes and thousands of time steps.