Using machine learning to aid in the parameter optimisation process for metal-based additive manufacturing

Abstract

Purpose Metal-based additive manufacturing is a relatively new technology used to fabricate metal objects within an entirely digital workflow. However, only a small number of different metals are proven for this process. This is partly due to the need to find a new set of parameters which can be used to successfully build an object for every new alloy investigated. There are dozens of variables which contribute to a successful set of parameters and process parameter optimisation is currently a manual process which relies on human judgement. Design/methodology/approach Here, the authors demonstrate the application of machine learning as an alternative method to determine this set of process parameters, the subject of this test is the processing of pure copper in a laser powder bed fusion printer. Data in the form of optical images were collected over the course of traditional parameter optimisation. These images were segmented and fed into a convolutional autoencoder and then clustered to find the clusters which best represented a high-quality result. The clusters were manually scored according to their quality and the results applied to the original set of parameters. Findings It was found that the machine-learned clustering and subsequent scoring reflected many of the observations which were found in the traditional parameter optimisation process. Originality/value This exercise, as well as demonstrating the effectiveness of the ML approach, indicates an opportunity to fully automate the approach to process optimisation by applying labels to the data, hence, an approach that could also potentially be suited for on-the-fly process optimisation. Graphical abstract

Control of Density and Grain Structure of a Laser Powder Bed-Fused Superelastic Ti-18Zr-14Nb Alloy: Simulation-Driven Process Mapping

“…Finally, the recent development of AI looks to be a promising solution for many modeling issues faced in additive manufacturing. Although the use of this approach is currently limited to in-process tuning of the processing parameters [35], it could eventually be able to solve such multiprinter/multipowder modeling issues. It must also be mentioned that the as-printed Ti-Zr-Nb alloy cannot be used without postprocessing treatment.…”

“…The latter is necessary to allow benefiting from the functional properties Finally, the recent development of AI looks to be a promising solution for many modeling issues faced in additive manufacturing. Although the use of this approach is currently limited to in-process tuning of the processing parameters [35], it could eventually be able to solve such multiprinter/multipowder modeling issues.…”

“…The latter is necessary to allow benefiting from the functional properties of shape memory and superelasticity of these materials. It is known that for conventional mechanically deformed metastable near-beta Ti-based alloys, this thermal treatment involves 30-60 min annealing in the 450-550 • C temperature range, followed by rapid cooling [35]. However, since the thermal histories of deformed alloys and their laser powder bed-fused equivalents differ, the influence of postprocessing thermal treatments on the structure and properties of printed specimens must be studied and optimized accordingly.…”

Machine Learning for Additive Manufacturing

“…A summary of the parameters studied as well as respective quality indicators are shown in Figure 2, with porosity and fatigue life identified as having the largest variety of process parameters linked to them. In addition, porosity is the leading quality indicator in the research, with at least five studies implementing machine learning algorithms to optimise process parameters [28][29][30][31][32]. Liu et al [32] built on previous efforts to develop a "physics-informed" model rather than a conventional "setting" model.…”

“…cator in the research, with at least five studies implementing machine learning algorithms to optimise process parameters [28][29][30][31][32]. Liu et al [32] built on previous efforts to develop a "physics-informed" model rather than a conventional "setting" model.…”