Abstract
This paper presents a machine-learning based predictive modeling approach on how process input parameters affect melt-pool volumes for a Laser Powder Bed Fusion (LPBF) additive manufacturing (AM) process. A physics-informed approach is adopted to define input features for the machine-learning models, and a two-level architecture is defined for the model training and validation. Specifically, a so-called initial (pre-deposition) temperature at the deposition point is identified as one key variable in characterizing thermal history for predicting melt-pool sizes. At the lower-level of the two-level modeling architecture, a hybrid model consisting of an analytical computation and a Gaussian process is developed to predict the pre-deposition temperature using process input parameters. Then at the higher-level of the modeling architecture, eight machine learning algorithms (including machine-learning based regression models and a two-layer neural network) are evaluated in predicting melt-pool volumes using the pre-deposition temperature and process parameters. For this proof-of-concept study, simulation data generated from the Autodesk Netfabb Local Simulation are used for model training and validation. The study shows that the proposed two-level machine learning model achieves high prediction performance and its prediction accuracy improves significantly compared to the one-level machine learning without using pre-deposition temperature as an input feature.
