Scalable and transferable graph neural networks for predicting temperature evolution in laser powder bed fusion

Abstract

Predicting temperature distributions in laser powder bed fusion (L-PBF) processes is essential for mitigating thermal distortions and ensuring the structural integrity of manufactured parts. Traditional finite element analysis (FEA) methods, while accurate, are computationally intensive and struggle to scale to larger domains. To address these limitations, this study proposes novel predictive models based on Graph Neural Networks (GNNs) to simulate thermal dynamics in L-PBF processes. The models leverage high-fidelity FEA data from small-scale domains to generalize effectively to larger domains with minimal retraining. For single-laser setups, the GNN achieves a Mean Absolute Percentage Error (MAPE) of 3.77 %, while significantly reducing computational costs. For instance, a thermomechanical simulation for a 2 mm square domain typically takes about 4 h, whereas the single-laser model predicts thermal distributions almost instantly. When calibrated for larger domains, the models significantly enhance predictive performance, showing notable improvements for square domains of 3 mm and 4 mm. Additionally, the models show a decreasing trend in Root Mean Square Error when tuned to larger domains, suggesting the potential for becoming geometry-agnostic. The interaction of multiple lasers complicates heat transfer, necessitating larger model architectures and advanced feature engineering. Using hyperparameters from Gaussian process-based Bayesian optimization, the best multi-laser surrogate model demonstrates a 46.4 % improvement in MAPE over the baseline model. By providing scalable and efficient predictive tools alongside FEA, this work paves the way for thermal modeling in L-PBF.