AI-Driven Self-driving Laboratories for Hard-to-Automate Experiments: Powders, Mechanochemistry, and Closed-Loop X-Ray Analysis
ORAL · Invited
Abstract
Recent advances in AI agents and autonomous control are opening a new frontier in experimental physics—one in which complex laboratory workflows traditionally dependent on tacit human skills can be executed, optimized, and generalized by machines. In this talk, I will present our group’s development of compact, modular autonomous experimental platforms that integrate (i) robotic powder handling and mechanochemical synthesis, (ii) autonomous X-ray diffraction (XRD) characterization, and (iii) AI-driven data analysis and optimal experimental design. These systems are designed to perform experimental tasks that are easy for humans but intrinsically difficult for robots, such as handling powders, kneading high-viscosity materials, or conducting force-controlled solid-state reactions under ultra-dry, oxygen-free conditions.
Our robotic mechanochemical synthesis platform demonstrates precise force-controlled grinding using a soft-gel-based elastic element, achieving reproducibility that surpasses both manual grinding and ball-milling, and enabling reproducible synthesis of perovskite materials. Autonomous characterization is realized through an integrated robotic XRD systemcapable of sample preparation, measurement, and automated Rietveld refinement via black-box Bayesian optimization, outperforming human experts in both speed and accuracy. Furthermore, we introduce new AI methodologies—such as PhaseDifformer, a diffusion-transformer model for single-observation multiphase XRD decomposition, and Bayesian optimal experimental design and automated stopping criteria for spectroscopy—creating a closed-loop pipeline that tightly couples experiment, analysis, and decision-making.
Together, these platforms illustrate a pathway toward AI agents that autonomously plan, execute, interpret, and adapt physical experiments, enabling closed-loop discovery in materials research. I will discuss the underlying physical challenges, the role of human expertise in training robotic intelligence, and the broader prospects for self-driving laboratories in physics and materials science.
Our robotic mechanochemical synthesis platform demonstrates precise force-controlled grinding using a soft-gel-based elastic element, achieving reproducibility that surpasses both manual grinding and ball-milling, and enabling reproducible synthesis of perovskite materials. Autonomous characterization is realized through an integrated robotic XRD systemcapable of sample preparation, measurement, and automated Rietveld refinement via black-box Bayesian optimization, outperforming human experts in both speed and accuracy. Furthermore, we introduce new AI methodologies—such as PhaseDifformer, a diffusion-transformer model for single-observation multiphase XRD decomposition, and Bayesian optimal experimental design and automated stopping criteria for spectroscopy—creating a closed-loop pipeline that tightly couples experiment, analysis, and decision-making.
Together, these platforms illustrate a pathway toward AI agents that autonomously plan, execute, interpret, and adapt physical experiments, enabling closed-loop discovery in materials research. I will discuss the underlying physical challenges, the role of human expertise in training robotic intelligence, and the broader prospects for self-driving laboratories in physics and materials science.
*This work is partly supported by the JST-Mirai Program (Grant Number JPMJMI19G1), the MEXT Program: Data Creation and Utilization-Type Material Research and Development Project (Digital Transformation Initiative Center for Magnetic Materials) (Grant Number JPMXP1122715503), the MEXT Program: Developing a Research Data Ecosystem for the Promotion of Data-Driven Science, the JSPS Grant-in-Aid for Transformative Research Areas (A) 22H05109 and 23H04483, and the JST Moonshot R&D (Grant Number JPMJMS2236).
–
Presenters
-
Kanta Ono
- The University of Osaka