A central challenge in machine learning-based thermal modeling for Laser Powder Bed Fusion (LPBF) is achieving reliable generalization across different part geometries and toolpaths. These capabilities are crucial for practical manufacturing, in order to avoid defects caused by poor process planning. Finite element simulations provide relatively accurate results but are too computationally intensive, highlighting the need for fast, generalizable surrogate models. This work presents a U-Net Convolutional Neural Network (CNN) that addresses the generalization challenge by leveraging intelligent feature engineering utilizing signed distance fields to represent geometry, time fields to capture laser scanning dynamics, and the time gradient field to encode heat diffusion patterns. This enables a single trained model to generalize effectively to new part shapes and toolpath orientations. The resulting surrogate model achieves around 1000 times speedup relative to finite element analysis while maintaining high accuracy, enabling real-time thermal prediction for process optimization and intelligent toolpath planning in additive manufacturing.
Demir et al. (Sun,) studied this question.