Felix Adams1,A. Gilad Kusne2,Ichiro Takeuchi1
University of Maryland1,National Institute of Standards and Technology2
Felix Adams1,A. Gilad Kusne2,Ichiro Takeuchi1
University of Maryland1,National Institute of Standards and Technology2
Autonomous experimentation uses machine learning and automation to intelligently choose the order of and perform experiments on the fly to achieve goals specified by a user (for example, finding synthesis parameters which optimize a materials property). Autonomous experimentation reduces the time and cost of materials research by achieving research goals with fewer experiments than exhaustive searching and traditional prototyping. While autonomous experimentation can incorporate many sources of information, qualitative human judgement remains an important part of decision making in some individual experiment tasks, which are difficult to automate. We present a method for integrating human input into an autonomous experimental campaign which maps structural phase regions in a high-throughput combinatorial library, demonstrated on the Fe-Ga-Pd system. X-ray diffraction patterns from the Fe-Ga-Pd library are used to predict potential phase maps. During the campaign, the user can indicate phase regions and boundaries based on their materials science knowledge (e.g., known specific material phases or the phase rule). This user input is integrated as a probabilistic prior into a Gaussian Process classifier, the posterior of which is used to choose the next experiment. We will describe the overall procedure and present the campaign results.