Accelerating High-Throughput Material Experimentation via Human-AI Collaborations

Experiment-based materials discovery has remained a “hard problem” because of the number and complexity of experiments required to efficiently search high-dimensional parameter spaces, along with the prohibitive time and cost constraints. In the past, physical intuition based on an engineer’s knowledge and experience has been absolutely critical to narrow the search sufficiently to yield viable workflows. But in other fields such as computer programming, development of large language model-based artificial intelligence “assistants” over the past few years have enabled software engineers to efficiently generate well-defined boilerplate code. We argue that there need to be, and can be, similar AI assistants for experimental labs to which scientists can delegate specific aspects of a complex experiment protocol. The goal of this task division, which must be clearly delineated and strictly defined, should enable the scientist or engineer to step back to a higher level and only steer the direction of experiments using their understanding of the topic and goals. The main challenge in the design of such AI-assisted experimental workflow lies in the question of where to draw the boundary so both the scientist’s and AI’s strengths can be exploited and maximized.<br/><br/>In our study, we have designed a Bayesian optimization-based autonomous experimental framework that allows seamless AI-human collaboration for targeted material synthesis and material property optimization. The AI agents incorporates on-the-fly probabilistic analysis of high-throughput characterization data (e.g. X-ray diffraction and spectroscopy), Gaussian processes, and advanced numerically stable acquisition functions for robust experimental decision making. In addition, we designed a family of acquisition function that allow the supervising scientist to tune the workflow dynamically, modifying in real-time the objective of the AI in order to guide the experiment direction. As an example, we demonstrate a seamless flow from exploration of potential metastable phases in complex alloys, to exploitation of an identified phase for specific properties. We show that such workflow can dramatically reduce the number of experiments required to approach the objective, and can be executed at experimental loop times consistent with continuously developing automated high-throughput experimentation.