Enhancing Active Learning Framework for Material Discovery and Optimization Through Incorporation of Physical Insights and Multimodal Data

Recently, artificial intelligence (AI) and active learning (AL) have been adapted by the experimental material science community for on-the-fly decision making and experiment planning, enabling autonomous research. Although researchers have demonstrated the ability of AL to efficiently explore materials spaces and identify compositions with extraordinary properties, existing methods mostly "learn" from the experimental data without significant physical insight. Additionally, inputs to the AI are typically based on single characterization data (unimodal). These constraints limit the physical reasoning ability of current AL agents. The incorporation of specific physical characteristics from the experiments as inputs to the AI models, and including data not directly related to the objective properties, has the potential to significantly improve the overall exploration and exploitation efficacy.<br/><br/>In this work, we demonstrate these improvements through two case studies. In the first, we incorporate concurrent probabilistic labeling of complex XRD patterns generated during laser spike annealing of composition gradients. A new phase labeling algorithm provides quantitative identification of potential multiphase structures with probabilistic assessment that is used by the AL agents to develop full temperature/time/composition phase processing maps. This has enabled us to readily expand the active learning AI from simple phase space exploration tasks to the more complex task of searching for specific targeted phase(s). We demonstrate the utility of this approach in the Bi-Ti-O system where phase maps for multiple structures were autonomously generated during an X-ray synchrotron run, with the AI transitioning from exploration mode to exploitation mode mid-experiment to optimize the anatase structure.<br/><br/>Second, we have incorporated additional "low-cost" measurements (e.g. optical microscopy) as indirect indicators of the objective property into the data analytic workflow, further accelerating the property optimization process. Using these additional data sources, the AI efficiency is shown to be significantly enhanced with minimal experimental overhead. In the first demonstration, this additional data was used to more efficiently utilize limited X-ray synchrotron time to maximize the AI learning. In second work, these data enable optimal use of more complex experimental protocols such as contact-mode electrical measurements. For both examples, we carefully examine the efficacy of the data processing and active learning methods, and demonstrate that the inclusion of these elements indeed improves the overall autonomous workflow.