UC Davis

Laser powder bed fusion (LPBF) additive manufacturing is a mesoscale process that uses small scale processing to create large scale parts. During this process, unconventional microstructures can form, which can have both favorable and adverse effects on part performance. Due to the complexity of the LPBF process and the large number of process parameters, controlling the resulting microstructures on a part specific basis has largely been impractical. While conventional simulation capabilities have shown promise for making predictions about the resulting microstructures, avoiding the time and monetary costs of printing and experimentally characterizing LPBF microstructures, these simulations are too slow for practical process tuning. This research aimed to develop a surrogate model, capable of making predictions about the microstructures produced by LPBF, fast enough to be used for process optimization.To achieve this, a framework for the creation of machine learning based surrogate models capable of making localized predictions about the statistical distribution of grain morphology metrics of interest was developed. This framework was then used to create models that can predict spatially defined grain size distributions for LPBF printed stainless steel, using a handful of thermal characteristics, calculated by a fast thermal model, that are representative of the LPBF process. The model is shown to not only be accurate in process regimes that are well represented in the training data, but also to improve on the accuracy of the training data by removing the stochastic error present in the data. The speed of the model, combined with the improved accuracy, allows for the prediction of the outcome of ensembles several thousand conventional microstructure simulations in a small fraction of the time. To ensure the framework and model are behaving as desired, a variety of methodologies for the analysis of the model were developed. These analyses provide insight into the behavior of the model with respect to the training data, the influence of the thermal characteristics chosen to represent the LPBF process, and the underlying assumptions used to create the model framework. The findings of these analyses support the validity of the framework and the usefulness of surrogate models created using it, showing promise for their future application to LPBF process optimization.