Gated Active Learning for Scientific Exploration in Autonomous Experiments

Recent advancements in autonomous experimentation and machine learning (ML) have transformed materials research including materials synthesis, characterization, and simulation. However, existing autonomous frameworks are often limited in their ability to accommodate the vast variability of materials experiments. This limitation hampers the efficiency of discovery processes. Therefore, we propose a gated active learning (GAL) approach to enhance traditional Bayesian Optimization-driven experimentation by incorporating dynamic gating mechanisms to streamline exploration and optimize experimental efficiency. In conventional autonomous experiments, ML algorithms typically treat all areas of the parameter space equally when searching for optimal material properties. However, this method can be inefficient, as certain areas of the parameter space may quickly become irrelevant. GAL addresses this issue by introducing a gating mechanism that actively adjusts the information flow within the autonomous experiment loop. GAL allows the primary optimization workflow to prioritize and focus on the most meaningful experimental results, the most promising parameter space, etc. This leads to more efficient optimization and accelerates the discovery of high-performing materials. The versatility of GAL allows for its application across various materials science domains and platforms, as well as integrating theory and human expertise in autonomous experiment loops. We have implemented it in autonomous pulsed laser deposition, automated liquid handler robot, and autonomous scanning probe microscopy for optimizing both materials syntheses and characterizations.

Acknowledgments: This work was performed at the Center for Nanophase Materials Sciences, a US Department of Energy Office of Science User Facility.