Machine Learning Framework Development for Rapid Input Space Mapping for Laser Powder Bed Fusion

Author(s):

Alexander Butler and Brandon McWilliams

Abstract:

Process optimization, alloy development, and other research efforts all possess broad input spaces that take considerable effort to characterize and understand. These lengthy and costly efforts hinder progress and knowledge application. ARL researchers developed an ML framework that aims to reduce the characterization and analytical burden of mapping a broad input space. This framework uses a greedy sampling algorithm and Latin hypercube sampling to assist with initial design of experiments and sample selection. The initial dataset is then prepared and used to train various ML models, using a hyperparameter grid search and k-fold cross validation. Using the trained models, process maps and correlation coefficients can be generated to better understand the given input space. The model can be queried to provide a predicted output for a given input combination or, conversely, provide a set of input combinations that would likely result in a given output. This data-driven framework is process and material agnostic, making it ideal for rapid implementation and transfer learning. The development and implementation of this framework is described for the laser powder bed fusion metal additive manufacturing process of aluminum alloy (AlSi10Mg).