Machine learning for knowledge transfer across multiple metals additive manufacturing printers

Abstract

Adopting new metals 3D printers introduces time and cost obstacles to printing parts with the same quality as was attained on existing printers. A large number of trial-and-error experiments or computationally intense simulations for property prediction and process optimization are often required. However, machine learning (ML) promises the ability to accelerate the adoption of new printing technologies. Specifically, it should enable transfer to the new printer prior knowledge built up from existing data, resulting in a data-informed starting point that is closer than would be possible without the statistics-based model. To enable the reuse of previously acquired knowledge, this study proposes a data-mining-assisted ML knowledge transfer framework. Bayesian models are found to more effectively model process-property relations and outperform support vector machine and logistic regression models. This framework is verified through 3 “industry-use” inspired scenarios for laser powder bed fusion of Ti-6Al-4V: 1) adopting a new model of printer from the same manufacturer that uses similar technology, 2) adopting a new model of printer from a different manufacturer that uses similar technology, and 3) adopting a new model of printer from a new manufacturer that uses different technology. The multi-property optimization experiments demonstrate the feasibility of cross-machine knowledge transfer to accelerate the adoption of new metals AM printing technologies.