Rapid layerwise process model calibration through a surrogate model for interpass temperature history in laser powder bed fusion

Layerwise thermal process simulation is widely used for predicting heat buildup and residual stress in parts built by the laser powder bed fusion (LPBF) process. However, calibration of layerwise process simulation model parameters (such as absorptivity and convection coefficients) using in-process thermal measurements typically requires hundreds of process simulations to reach convergence due to the large model parameter space. To significantly reduce the number of process simulations needed for calibration, this work proposes a surrogate model (SGM) to accurately model the relationship between simulated interpass temperature time history and model parameters needing calibration using simple functions. Specifically, the SGM is constructed by fitting multi-variate polynomials to the temperature histories mapped to a low-dimensional model space through the principal component analysis (PCA), as a function of the process model parameters. Since the SGM is 4-5 orders of magnitude faster than finite element method based process simulations, it can be used to calibrate the model parameters in several minutes. The SGM specifically developed for calibrating three process model parameters requires only ten process simulations to fit its coefficients and is tested to be accurate using 18 simulated temperature histories to fit ten model coefficients, yielding a mean absolute percentage error (MAPE) between 0.12 and 2.04% across four test geometries. Applying the fitted surrogate models to calibrate absorptivity, top surface convection coefficient, and part-side surface convection coefficient to match the temperature histories measured by an infrared camera, a MAPE between 2.84 and 3.46% is achieved across the four test geometries using the same set of model parameters. Two unseen geometries are tested with the fully calibrated FEM model with errors under 3.7% MAPE. The proposed SGM approach is demonstrated to be a valuable tool for accurate part-scale thermal prediction by significantly reducing calibration time and computational resources.