Data-augmented modeling in laser powder bed fusion: A Bayesian approach

The laser powder bed fusion (LPBF) technique has become increasingly prominent in metal additive manufacturing. However, tuning parameters for printing defect-free parts requires labor-intensive experimental work and computationally expensive simulations. Moreover, to calibrate LPBF models against the experimental data, typically MCMC methods or similar methods are used which is also time-consuming. These procedures are viable when calibrating LPBF models against data for individual chemistries but are not efficient for alloy design. A rapid method to calibrate LPBF models is needed to design for printable alloys. We address this challenge by integrating a low-fidelity analytical thermal model, a machine learning model, and proxy experimental data to create an accurate and rapidly-trained model that leverages the principles of Bayesian updating. As a case study in ‘printability extrapolation’, a dataset of 195 single-tracks on 16 unique chemistries was used to probe the method’s ability to predict melt-pool dimensions on ‘unseen’ chemistries. As a case study in ‘printability interpolation’ the framework was deployed on two compositions that were studied rigorously in the literature for their printability, namely, the ultra-high strength martensitic steel alloy AF9628 and the nickel super alloy 718. The interpolative/extrapolative abilities of the proposed method were compared to a set of 4 control models under data sparse conditions.