BAYESIAN MODEL UPDATING FOR LASER POWDER BED FUSION (LPBF) ADDITIVE MANUFACTURING: UNCERTAINTY QUANTIFICATION OF HEAT SOURCE AND MATERIAL PARAMETERS

Numerical simulation of metal additive manufacturing processes is a key technology for revealing the intrinsic relationship between process parameters, melt pool dynamics, and forming quality, thereby enabling controlled shaping and property preservation during fabrication. However, due to the complexity of physical processes such as heat conduction and the limited availability of experimental data, various sources of uncertainty inevitably exist in model building, casting doubt on the predictive accuracy of model. In order to enhance the predictive accuracy of the heat transfer model and fully quantify the uncertainty of the model parameters, this paper addresses the challenge of high-fidelity simulation modelling for the heat transfer process in laser powder bed fusion (LPBF) process. A Bayesian correction method for heat transfer models is developed, which quantifies the uncertainties in heat source parameters and material parameters based on melt pool size data. Furthermore, in order to solve the problem of high computational cost in the heat transfer model correction process, an adaptive learning Gaussian process regression model training method is developed, which can significantly improve the correction efficiency while maintaining the computational accuracy. Finally, the proposed method is used to obtain the correction results of heat source and material parameters under different experimental observation datasets, and the results demonstrate that the method developed in this paper can effectively update the uncertainty of heat source and material parameters at the same time, and the printing power and scanning velocity exert significant influence on the uncertainty of thermal conduction model parameters. Research has revealed that the deviations between model predictions and experimental values cannot be resolved only through model correction. To address this issue, this paper further incorporates model bias. By introducing a bias function related to printing power and scanning speed, the accuracy of the corrected molten pool depth predictions is significantly enhanced.