Additive Manufacturing Melt Pool Prediction and Classification via Multifidelity Gaussian Process Surrogates

It is well known that the process parameters chosen in metal additive manufacturing (AM) are directly related to the melt pool dimensions, which can be related to microstructure characteristics, properties, and printability. Thus, the determination of melt pool dimensions and resulting printability, given a set of process parameters, is crucial to understanding the performance of AM parts. Unfortunately, experiments relating process parameters to the melt pool have a high execution and data collection cost. On the other hand, simulations do not suffer this cost penalty but can never be as good as the ground truth experiments. A number of capabilities exist to predict melt pool geometry from process parameters that range from highly accurate codes with long simulation times to less accurate analytical solutions that are nearly instantaneous. This work leverages multifidelity Gaussian process (GP) surrogates to examine how the fusion of information from different fidelities, including experiments, influences the resulting predictive model. Both a multifidelity GP regression and a novel multifidelity GP classification are examined. The multifidelity models are compared to standard GPs and a withheld set of test data. The results show good predictive accuracy for the melt pool width and depth, but suggest multifidelity models may not significantly improve classification of printability.