materials-chemistry/ai produced the result/Nature Communications 2023 · v2

Algorithm picks starting powders and temperatures for making inorganic materials

Researchers built a decision loop, ARROWS, that chooses which starting chemicals and temperatures to try when making a solid material. A neural network read the X-ray patterns after each heating step and told the loop which compounds had formed.

spectrum · one line per step, placed by what the step does · bright lines used AI

Autonomous and dynamic precursor selection for solid-state materials synthesis
Nature Communications, 2023

doi:10.1038/s41467-023-42329-9 · record aix-00026 v2 · checked 2026-10-07

ai-resultrole of AI

AI was for: Classification, Experimental design
Model family: Convolutional neural network
Checked by: Experimental87 tested, 10 worked
Code: not reported

The finding the paper is about came from the AI.

read as

The science is explained before the AI appears. Switch to field specialist to go straight to the method.

Assumes the discipline and goes straight to the method.

The diagram, the record and what the paper did not report are identical in both modes. Only the framing changes — never the evidence.

Introduction by AIxSci · plain language

What this research was about

Making a solid inorganic material is not like mixing a solution. Powders are ground together and heated, and the atoms must rearrange by creeping through solid grains. Along the way, the mixture usually passes through other compounds first, called intermediates. Some of these are stubborn, and once they form, little energy is left to drive the final product into being. So the choice of starting powders, the precursors, and the temperature they are heated at decides whether the intended material appears cleanly or is swamped by unwanted phases. Chemists have long leaned on experience and trial and error here, because the sequence of intermediate steps is hard to predict in advance.

The researchers set out to automate that choice. Their system enumerates every combination of available precursors whose amounts balance against the target compound, ranks those combinations by how much energy is released in forming the target, and proposes experiments across a range of temperatures. After each round it works out which pairs of compounds reacted and what they made, then uses that growing picture to guess which intermediates untested combinations would form, and re-ranks them by how much driving force would remain at the final, target-forming step.

Where AI came in

Two learned pieces sit inside a loop that is otherwise driven by thermodynamics and fixed rules. The first is a published machine-learned formula that estimates how a compound's free energy changes with temperature; combined with calculated formation energies from the Materials Project database and measured gas data from NIST, it supplied the energies used to rank candidate precursor sets. The second is XRD-AutoAnalyzer, a convolutional neural network. X-ray diffraction sends X-rays through a powder and produces a pattern of peaks that acts as a fingerprint of the crystalline compounds inside. Reading those overlapping fingerprints is normally expert work; here the network did it, naming the phases present and allowing their proportions to be estimated.

Those phase assignments were what the decision loop learned from, so the AI stood in for the human eye on each diffraction pattern and kept the cycle running without a chemist in the middle. On a retrospective set of experiments for the superconductor YBa2Cu3O6.5, ARROWS found all 10 precursor sets that made the target with no detectable impurities after 87 queried experiments, where Bayesian optimisation, a genetic algorithm and D-optimal design — three general-purpose search methods the authors ran as comparisons — needed on average 164, 167 and 165.

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

ARROWS ranks stoichiometrically balanced precursor sets for a target inorganic material by their calculated reaction energy, proposes experiments across a range of temperatures, and uses a convolutional neural network (XRD-AutoAnalyzer) to identify the phases present in the resulting diffraction patterns. From the identified intermediates it reconstructs pairwise reactions and re-ranks untested precursor sets by the driving force remaining at the target-forming step. On a dataset of 188 YBa2Cu3O6.5 experiments the algorithm located all 10 precursor sets that formed YBCO without detectable impurities after 87 queried experiments, while D-optimal design required 165 and Bayesian optimization and a genetic algorithm required on average 164 and 167. Campaigns guided by the algorithm produced a sample containing 94% Na2Te3Mo3O16 by weight after an 8 h hold, and a triclinic LiTiOPO4 route giving 54% target yield with no detectable orthorhombic polymorph.

How AI was used

Two learned components sit inside an otherwise thermodynamic decision loop. First, temperature-dependent free energies used to rank candidate precursor sets were approximated with a published machine-learned descriptor, combined with DFT formation energies from the Materials Project and experimental gas-phase data from NIST. Second, the XRD patterns measured after each heating step were analysed with XRD-AutoAnalyzer, a convolutional neural network that maps a pattern onto its constituent phases; for the YBCO campaign the network was trained on ICSD phases within the Y–Ba–Cu–O space, and weight fractions were then estimated from relative peak intensities. The phase assignments were fed into a rule-based pairwise reaction analysis that balances candidate two-phase reactions to explain each newly observed intermediate, builds a database of reactive and inert pairs with onset temperatures, predicts the intermediates expected for untested precursor sets, and re-ranks those sets by the reaction energy remaining at the target-forming step; the next experiment or batch of experiments was selected from that updated ranking. As baselines on the retrospective YBCO design space, the authors also ran Bayesian optimization, a genetic algorithm and D-optimal design over one-hot encodings of the precursors.

The shape of the work

Structural · the record, drawn

AI stepNo AI↤ what the AI stood in for

1Preparation
no AI

Enumerate balanced precursor sets and temperature grid

Cleaning, filtering, normalising or labelling data already obtained.

all unique precursor combinations are enumerated and those that can be stoichiometrically balanced with the target are recorded as possible precursor sets for itwhere the paper describes this · verbatim

in the paper

2Screening
AI

Rank precursor sets by DFT reaction energy

Reducing a candidate set by filtering or ranking, in a single pass.

along with temperature-dependent free energies approximated using the machine-learned descriptor developed by Bartel et alwhere the paper describes this · verbatim

in the paper

3Experiment
no AI

Run solid-state synthesis experiments and measure XRD

Physical execution, by hand or by robot.

These compounds were combined to form 47 different precursor sets, listed in Supplementary Table 1, that were each tested at four synthesis temperatureswhere the paper describes this · verbatim

in the paper

4Inference
AI

Identify phases and weight fractions from XRD patterns

Running a trained model over new data to predict, classify or score. The AI stood in for expert judgement.

We used XRD-AutoAnalyzer to analyze the resulting XRD patterns and identify any crystalline phases present.where the paper describes this · verbatim

in the paper

5Interpretation
no AI

Infer pairwise reactions from temperature-resolved phases

Extracting understanding from model behaviour.

ARROWS then determines which pairwise reactions led to the formation of each observed intermediate phasewhere the paper describes this · verbatim

in the paper

6Optimisation
no AI

Predict intermediates and re-rank by remaining driving force

Iterative search over a space. Its result feeds back into an earlier step.

it leverages this information to predict the intermediates that will form in precursor sets that have not yet been testedwhere the paper describes this · verbatim

in the paper

7Validation
AI

Benchmark against black-box optimisers on the YBCO dataset

Testing outputs against ground truth.

We also applied two active learning algorithms, Bayesian optimization (BO) and a genetic algorithm (GA), to the same taskwhere the paper describes this · verbatim

in the paper

8Validation
no AI

Confirm optimised routes with longer hold or ball milling

Testing outputs against ground truth.

we prepared a new sample containing the same precursors (Na2O, MoO3, and TeO2)where the paper describes this · verbatim

in the paper

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.

~Role of AI

AI produced the resultour reading

The reported result is the performance of the ARROWS decision algorithm, which selected which precursor sets and temperatures were tested; a convolutional neural network supplied the phase assignments the algorithm learned from. ARROWS itself is a thermodynamic, rule-based optimiser rather than a trained model, so this classification rests on the learned components embedded in its loop.

+What the AI was for

Classification Experimental designin the paper

This algorithm relies on a convolutional neural network to map each pattern onto a set of constituent phases.where the paper describes this · verbatim

+Model families

Convolutional neural networkin the paper

+How it was taught

SupervisedActive learningin the paper

+Models named

XRD-AutoAnalyzer · Trained from scratchBartel et al. machine-learned free-energy descriptor · Off the shelfBayesian optimization (BO) baseline · Trained from scratchin the paper

+How results were checked

Experimental87 tested, 10 workedin the paper

ARROWS successfully identified all 10 optimal routes from 87 experimentswhere the paper describes this · verbatim

+Code · weights · data

code not reportedweights not reporteddata availablein the paper

all data reported in this work is made publicly availablewhere the paper describes this · verbatim

+Compute

not reportedin the paper

What this paper did not report

Technical · absence is published deliberately

Reported as not stated — 8 items

CodeWhether the code is available is not stated.
Trained model weightsWhether the trained model is available is not stated.
ComputeThe hardware or time used is not stated.
Version of XRD-AutoAnalyzerWhich version of the model was used is not stated.
Version of Bartel et al. machine-learned free-energy descriptorWhich version of the model was used is not stated.
Version of Bayesian optimization (BO) baselineWhich version of the model was used is not stated.
What step 2 replacedThe paper gives no basis for what the AI stood in for.
What step 7 replacedThe paper gives no basis for what the AI stood in for.

About this article

Record aix-00026, 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