materials-chemistry/ai produced the result/Nature 2023 · v2
Robotic lab with machine learning proposes and runs 353 inorganic synthesis experiments
Researchers built a laboratory in which machine-learning models drew up recipes for making new solid materials from published literature, robots carried them out, and further models read the X-ray measurements that showed what had formed.
spectrum · one line per step, placed by what the step does · bright lines used AI
An autonomous laboratory for the accelerated synthesis of inorganic materials
Nature, 2023
doi:10.1038/s41586-023-06734-w · record aix-00010 v2 · checked 2026-10-07
- AI was for
- Candidate generation, Property prediction, Classification, Experimental design
- Model family
- Convolutional neural network, Gradient-boosted trees
- Checked by
- Experimental57 tested, 36 worked
- Code
- available
The finding the paper is about came from the AI.
What this research was about
Most of the solid materials in batteries, magnets and catalysts are inorganic powders, made by mixing starting chemicals and baking them until they react. Computer calculations can predict which combinations of elements ought to form a stable compound, and such predictions now run far ahead of what anyone has actually made. Turning a prediction into a powder is the slow part. A chemist must choose which starting chemicals, called precursors, to mix, and at what temperature to heat them. That choice rests on experience and on scattered recipes buried in decades of published papers, and a wrong guess yields a jumble of unwanted products instead of the target.
The researchers set out to close that loop without a person in it. They picked 57 target compounds that calculations suggested should be stable and that did not appear in records of known substances, then let an automated laboratory, the A-Lab, attempt to make them. Over 17 days of continuous running and 353 experiments, it obtained 36 of the 57. Afterwards the authors went back over the X-ray data by hand to check the conclusions the lab had reached on its own.
Where AI came in
Learned models made every decision a chemist would normally make. One model, trained on synthesis procedures text-mined from published papers, ranked known materials by how similar their preparation ought to be to each target and borrowed their precursors; a second filled in any missing ingredients; and a third predicted the heating temperature. Up to five recipes per target were produced this way, standing in for a chemist's judgement about where to start.
Reading the results was also automated. After the robots ground each product and measured its X-ray diffraction pattern, the fingerprint of peaks that reveals which crystals are present, a convolutional neural network, a model that learns to recognise patterns in signals, named the phases and estimated how much of each there was. A second model, trained by trial and error, then handled the detailed pattern-fitting that crystallographers usually do by hand. When a target's yield stayed low, an active-learning algorithm worked out which unhelpful reactions had taken place and proposed fresh precursor sets to avoid them, sending them back to the robots.
Written by AIxSci from the checked record below, to give context for readers outside the field. It is not part of the record.
The work
Technical · from the record
The A-Lab is an autonomous laboratory for solid-state synthesis of inorganic powders, in which machine-learning models proposed synthesis recipes from text-mined literature data, robotics carried them out, and further models interpreted the resulting X-ray diffraction patterns. Over 17 days and 353 experiments, the platform obtained 36 of 57 computationally screened target compounds; 30 of the 36 came from the literature-derived recipes and the active-learning cycle found higher-yield routes for nine targets. The authors manually re-refined the diffraction data afterwards to check the autonomous conclusions, and classified the 17 unobtained targets into slow reaction kinetics, precursor volatility, amorphization and errors in the computed stability.
How AI was used
Targets were drawn from ab initio phase-stability data by a rule-based screening algorithm, after which a learned encoding model ranked literature materials by synthesis similarity to each target and supplied their precursors, with a masked precursor completion model filling in missing elements and an XGBoost regressor predicting a synthesis temperature from precursor properties, target composition and computed pairwise driving forces; one averaged temperature per target was used so recipes could be batched. Robotic stations dosed, mixed, heated and ground the samples and measured X-ray diffraction. A convolutional neural network, trained per target on patterns simulated from ICSD structures and the volume-corrected computed target structure and evaluated with Monte Carlo dropout over an ensemble of 100 networks, identified phases and estimated weight fractions, and an actor/critic agent trained with proximal policy optimization drove automated multiphase Rietveld refinement through GSAS-II. When target yield stayed at or below 50%, the ARROWS active-learning algorithm inferred which pairwise reactions had occurred, recorded them in a growing database, and proposed new precursor sets at stepped temperatures that avoided intermediates leaving only a small computed driving force to the target, feeding those recipes back to the robotic synthesis stage.
The shape of the work
Structural · the record, drawn
no AI
Screen air-stable target materials
Reducing a candidate set by filtering or ranking, in a single pass.
The 57 targets evaluated by the A-Lab were identified from the Materials Project database (version 2022.10.28).where the paper describes this · verbatim
AI
Propose precursor sets and synthesis temperature
Producing candidate objects that did not previously exist. The AI stood in for expert judgement.
up to five initial synthesis recipes are generated by a ML model that has learned to assess target ‘similarity’ through natural-language processingwhere the paper describes this · verbatim
no AI
Execute recipes with robotic dosing, mixing and heating
Physical execution, by hand or by robot.
A robotic arm from the second station loads these crucibles into one of four available box furnaces to be heatedwhere the paper describes this · verbatim
no AI
Grind products and collect XRD patterns
Obtaining raw data, whether by measurement, download or retrieval.
The UR5e arm transfers each flattened sample into the diffractometer for X-ray measurements, which are performed using 8-min scanswhere the paper describes this · verbatim
AI
Identify phases in diffraction patterns
Running a trained model over new data to predict, classify or score. The AI stood in for expert judgement.
we apply XRD-AutoAnalyzer to identify the constituent phases and estimate their weight fractionswhere the paper describes this · verbatim
AI
Confirm phases by automated Rietveld refinement
Testing outputs against ground truth. The AI stood in for expert judgement.
we use an automated approach to multiphase Rietveld refinement. An agent with two deep neural networks (actor/critic) were trained using reinforcement learningwhere the paper describes this · verbatim
AI
Propose improved reaction routes with ARROWS
Iterative search over a space. The AI stood in for exhaustive search. Its result feeds back into an earlier step.
New synthesis experiments are then proposed on the basis of sets of precursors expected to avoid such reactionswhere the paper describes this · verbatim
no AI
Manual refinement of autonomous results
Testing outputs against ground truth.
we manually performed Rietveld refinement on the XRD data that was acquired from the original (autonomous) experimentswhere the paper describes this · verbatim
What the record says
Technical · every part carries its own basis
+ in the paper~ our reading− not reported
How to read the quotations. A quotation shows where the paper describes something. It does not quote every value beside it: one passage locates a part of the work, and values without their own quotation are our reading of that passage.
Learned models proposed every synthesis recipe tested, interpreted the resulting diffraction patterns, and drove the optimisation loop, so the reported synthesis outcomes are products of the AI components
This algorithm relies on a convolutional neural network (CNN) consisting of six convolutional layerswhere the paper describes this · verbatim
the A-Lab synthesized 36 of the 57 target compounds over 17 days of continuous experimentationwhere the paper describes this · verbatim
The algorithm we used for identifying potential synthesis targets is available on GitHubwhere the paper describes this · verbatim
What this paper did not report
Technical · absence is published deliberately
- Trained model weightsWhether the trained model is available is not stated.
- ComputeThe hardware or time used is not stated.
- Version of XRD-AutoAnalyzer convolutional neural network for phase identificationWhich version of the model was used is not stated.
- Version of XGBoost synthesis-temperature regressorWhich version of the model was used is not stated.
- Version of Synthesis-context-based target encoding model (precursor recommendation by similarity)Which version of the model was used is not stated.
- Version of Masked precursor completion modelWhich version of the model was used is not stated.
- Version of Actor/critic Rietveld-refinement agent (proximal policy optimization)Which version of the model was used is not stated.
About this article
Record aix-00010, version 2, checked by a person on 2026-10-07. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY-4.0; quotations are at most 25 words. How we work · Report an error