Machine Learning Augmented Experimental Design in Materials Processing Research

As our understanding of materials processing and its parameters improve, so does our capabilities to tailor specific materials properties. However, the processing spaces themselves are extremely vast, and optimizing it with conventional experimental methods is too laborious and time consuming. To address this, we integrated machine learning, both supervised learning and active learning, into materials processing to accelerate materials discovery and optimization. In this thesis, we researched two manufacturing methods with these new workflows. First, we analyzed laser powder bed fusion (LPBF), a non-equilibrium additive manufacturing process with an ability for printing geometries with free-form features, in producing Bi2Te3-based thermoelectrics. We developed an iterative, augmented learning experimental-ML workflow where we predicted and optimized melt pool geometries and Seebeck coefficient as a function of the LPBF process parameters. With guidance from machine learning, our experimental collaborators successfully identified parameters that produced crack-free and highly dense (>99) samples. Second, a novel pool-based active learning workflow was developed for producing large area and full coverage copper-based HKUST-1 metal-organic framework thin films through a flow coating process called solution shearing. Pool-based active learning workflow enabled careful selection of batches of data points for experimental validation and rational expansion of the training data. In both projects, the machine learning driven workflow contributed to reducing the time in navigating the vast space in materials processing.