Berkeley Lab's Autonomous A-Lab Synthesizes 36 of 57 Target Inorganic Materials in 17 Days

The A-Lab at Lawrence Berkeley National Laboratory is an autonomous laboratory that combines robotics with ab initio databases, machine-learning data interpretation, synthesis heuristics learned from text-mined literature and active learning to synthesise inorganic materials in powder form, described in a Nature paper published in November 2023. "Accelerating the experimental segment of materials discovery requires not only automation but autonomy - the ability of an experimental agent to interpret data and make decisions based on it," the authors write. Targets were identified from large-scale ab initio phase-stability data from the Materials Project and an analogous database from Google DeepMind, and screened for predicted reactivity with oxygen, carbon dioxide and water because samples are handled in open air. For each target, up to five initial recipes are proposed by a model that assesses similarity to known materials through natural-language processing of a large database of literature syntheses, and a second ML model trained on literature heating data proposes the temperature. If these recipes do not reach more than 50% yield, an active-learning algorithm, ARROWS3, which combines computed reaction energies with observed outcomes, proposes new recipes. The targets were new to the lab, meaning they were not in the training data of the recipe-proposing algorithms. The bespoke robotic platform has a precursor preparation station with a robot arm for powder dispensing and mixing built with Labman Automation, four box furnaces, an in-house sample-handling station and X-ray diffraction. Diffraction patterns are analysed by ML models trained on structures from the Materials Project and the ICSD, and the phases are confirmed by automated Rietveld refinement. Over 17 days of continuous operation the A-Lab obtained 36 of 57 targets, a 63% success rate, spanning a range of oxides and phosphates, although only 30% of the 353 recipes tested produced their targets. Analysis of the 17 unobtained targets identified synthetic and computational failure modes, several of which could be overcome by minor adjustments to the lab's decision-making. An author correction to the paper was published on 19 January 2026.