Autonomous Experimentation

In the lab, much of the work — preparing reagents, synthesizing samples, measuring properties — still relies on manual effort. Autonomous experimentation aims at more than simply automating these tasks.

Its essence is a loop that a machine runs on its own: experiment → measurement → AI analysis → proposal of the next conditions → experiment again — so-called closed-loop optimization. This lets us efficiently gather the large datasets that materials informatics requires, and reach the best conditions in far fewer trials.

As AI and robotics advance, research on automating and autonomizing materials experiments is moving quickly. In 2020, B. Burger and colleagues reported “A mobile robotic chemist” in Nature: over eight days it autonomously carried out 688 experiments, drawing wide attention. A stream of similar reports followed, making this one of the central themes in materials science today.

B. Burger et al., “A mobile robotic chemist,” Nature 583, 237 (2020).

Other efforts include the A-Lab at Lawrence Berkeley National Laboratory, the robot “Maholo” at Japan's AIST, and NIMS-OS, which integrates materials exploration — autonomous experimentation is being developed around the world.

Our group, too, has built an automated solvent-mixing system combining an off-the-shelf robot arm with electronic pipettes, pursuing a dramatic improvement in research throughput.

As one example, in the automated synthesis of ZIF-8 — a metal–organic framework (MOF) — we built a machine-learning model (CatBoost) that predicts particle size from synthesis parameters (reagent concentration, dispensed volume, mixing speed, etc.), and used SHAP analysis to identify the factors governing particle size, demonstrating high-precision synthesis of particles with targeted sizes.