(664g) Active Learning and Symbolic Regression for Interpretable Modeling of Laser Power Bed Fusion (L-PBF) Additive Manufacturing Processes

Laser power bed fusion (L-PBF) is an emerging additive manufacturing (AM) technology in which metal parts are built layer-by-layer through laser sintering of metal powder from computer-aided design models [1], [2]. AM technology has progressed beyond prototyping in the past few decades and is now used to reliably produce parts for various industrial sectors. Additive manufacturing has the benefit over subtractive and formative manufacturing methods in that it can produce parts much more quickly while wasting little material. Recently, pulsed wave (PW) laser emission, as compared to the traditional continuous wave (CW) emission, has been proposed to refine the microstructure and improve the density and porosity of the manufactured parts. At the same time, PW emission brings a much larger parameter space, which can be difficult to traverse without proper techniques [3]. In the L-PBF process, there are generally three melting modes: (1) the conduction mode, where heating is dominated by the conduction of heat through the metal; (2) the keyhole mode, where heating is dominated by radiative heat transfer; and (3) the transition mode, where both phenomena compete for dominance. Previous work has shown that the transition mode is where the highest density is found/manufactured [1], [4]. In this work, we develop a process map for Ti-6V-4Al to identify where the transition mode region exists in the process parameter space. We do so by performing single scan tracks (SSTs) on titanium substrates and measuring the melt pool aspect ratios from cross-sectional imaging.

Working towards this goal, we use active learning (AL) through an iterative design of experiments (DOE) approach to maximize information gain from the process. Through this iterative DOE approach, we hope to gain as much information as possible from the parameter space in relatively few experiments. We combine this experimental case with other simulated cases to investigate whether specific AL strategies can be applied more generally. Among the AL strategies we implement are Gaussian process uncertainty sampling and leave-one-out sensitivity sampling, which we compare with random sampling and distance-based sampling. Both of these AL strategies attempt to sample from regions of the parameter space in which the output is either not well characterized or is highly sensitive to the process parameters [5], [6]. Finally, we use symbolic regression to obtain physically-interpretable models from the acquired data, so that we can gain insight into the physical phenomena occurring in the process. In this work, we show that we can use AL and symbolic regression to derive robust, physically-interpretable models in relatively few experiments. We show that we can extend this approach to other application areas to efficiently discover models that extract physical meaning from the systems of interest.

[1] S. Patel and M. Vlasea, “Melting modes in laser powder bed fusion,” Materialia (Oxf), vol. 9, Mar. 2020, doi: 10.1016/j.mtla.2020.100591.

[2] S. S. Razvi, S. Feng, A. Narayanan, Y.-T. T. Lee, and P. Witherell, “A Review of Machine Learning Applications in Additive Manufacturing,” in International design engineering technical conferences and computers and information in engineering conference, Aug. 2019.

[3] E. Toyserkani, D. Sarker, O. O. Ibhadode, F. Liravi, P. Russo, and K. Taherkhani, “Basics of Metal Additive Manufacturing,” in Metal Additive Manufacturing, Wiley, 2021, pp. 31–87. doi: 10.1002/9781119210801.fmatter.

[4] S. Patel and M. L. Vlasea, “Melting Mode Thresholds in Laser Powder Bed Fusion and their Application Towards Process Parameter Development,” 2019.

[5] T. Lookman, P. V. Balachandran, D. Xue, and R. Yuan, “Active learning in materials science with emphasis on adaptive sampling using uncertainties for targeted design,” npj Computational Materials, vol. 5, no. 1. Nature Publishing Group, Dec. 01, 2019. doi: 10.1038/s41524-019-0153-8.

[6] Q. Zhou et al., “An active learning radial basis function modeling method based on self-organization maps for simulation-based design problems,” Knowl Based Syst, vol. 131, pp. 10–27, Sep. 2017, doi: 10.1016/j.knosys.2017.05.025.