A Reduced Gaussian Process Heat Emulator for Laser Powder Bed Fusion

Laser Powder Bed Fusion (LPBF) is a promising additive manufacturing technique used for realizing complex and bespoke designed metal parts. Despite its good performance, its quality assurance is still hampered by the absence of in-process optimization and control. In this sense, real-time thermal analysis can facilitate fault predictions and rectifications. High-fidelity three-dimensional thermal modelling with the Finite Element Method (FEM) is generally time-consuming since the heat transfer equation is nonlinear and high-dimensional. The challenge is thus to compute fast, reliable and accurate thermal predictions that capture the nonlinearity triggered by the phase changes of the part during printing. Gaussian Process (GP) with Isomap dimension reduction is investigated to find and predict the low-dimensional representations of the high-dimensional thermal profiles in FEM without intricate processing. Based on these representations, the high-dimensional predictions are then approximated using localized radial basis functions. To validate the performance of this reduced GP heat emulator, a heat simulation during fabricating an Aluminum object is performed to compare FEM-based temperature calculations against reduced GP emulations. Retaining 0.06% of the original model dimension the execution time per temperature profile is 0.70s on average achieving a 95.07% reduction, while maintaining at least 85% accuracy (with respect to the FEM simulation) for 96.80% of the thermal profile queries and at least \(80\%\) for 89.38% of the thermal history queries. With this encouraging performance, the reduced GP heat emulator can be a step forward in online defect prediction, process optimization and closed-loop control in LPBF.