Active learning is a type of machine learning that takes a more active role in the acquisition of new data or the design of new experiments. While the goal of machine learning is generally to find patterns in the data that allow the prediction of selected outcomes, active learning exploits these patterns to indicate which unlabeled samples may contain the most new information (e.g., the highest uncertainty in a classification setting). For a phase diagram problem, this could indicate regions where the phase is uncertain [2]. We can use active learning through an iterative design of experiments (DOE) approach to efficiently build a phase diagram or process model. This can take some human assumptions or even errors out of process mapping and allow for more efficient optimization and discovery [3], [4], [5].
In this work, we look at several machine learning models, including support vector machines, random forest classifiers, and Gaussian process classifiers, for use with active learning algorithms. We investigate in particular the exploration-exploitation trade-off and suggest certain hybrid strategies that optimize this trade-off and provide robust search strategies. We also show which strategies hold up best against deterioration when shifting from sequential sampling to batch sampling. We test these strategies against phase diagrams which are known to have complex boundaries. Finally, we discuss how we have applied these methods through iterative experiments to build a process map for a laser powder bed fusion (L-PBF) additive manufacturing (AM) process.
[1] 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,” Dec. 01, 2019, Nature Publishing Group. doi: 10.1038/s41524-019-0153-8.
[2] K. Terayama et al., “Efficient construction method for phase diagrams using uncertainty sampling,” Phys Rev Mater, vol. 3, no. 3, Mar. 2019, doi: 10.1103/PhysRevMaterials.3.033802.
[3] M. Zhang, A. Parnell, D. Brabazon, and A. Benavoli, “Bayesian Optimisation for Sequential Experimental Design with Applications in Additive Manufacturing,” Jul. 2021, ArXiv. Available: http://arxiv.org/abs/2107.12809
[4] R. Yuan et al., “Accelerated Discovery of Large Electrostrains in BaTiO3-Based Piezoelectrics Using Active Learning,” Advanced Materials, vol. 30, no. 7, Feb. 2018, doi: 10.1002/adma.201702884.
[5] G. Lambard, T. T. Sasaki, K. Sodeyama, T. Ohkubo, and K. Hono, “Optimization of direct extrusion process for Nd-Fe-B magnets using active learning assisted by machine learning and Bayesian optimization,” Scr Mater, vol. 209, Mar. 2022, doi: 10.1016/j.scriptamat.2021.114341.