materials-chemistry/ai produced the result/Nature Communications 2023 · v2
Machine learning picked catalyst recipes for a loop of 44 lab cycles
Researchers tested 255 new catalyst compositions for turning carbon dioxide into carbon monoxide. A machine learning model trained on measured activity chose which mixtures to make, and the lab results were fed back to retrain it.
spectrum · one line per step, placed by what the step does · bright lines used AI
Accelerated discovery of multi-elemental reverse water-gas shift catalysts using extrapolative machine learning approach
Nature Communications, 2023
doi:10.1038/s41467-023-41341-3 · record aix-00178 v2 · checked 2026-10-09
- AI was for
- Property prediction, Experimental design
- Model family
- Random forest, Clustering
- Checked by
- Experimental255 tested
- Code
- available
The finding the paper is about came from the AI.
What this research was about
The reverse water-gas shift reaction takes carbon dioxide and hydrogen and turns them into carbon monoxide and water. Carbon monoxide is a useful building block for fuels and chemicals, so the reaction is one route to putting waste carbon dioxide back to work. It needs a catalyst: a solid material that speeds the reaction up without being consumed. Catalysts of this kind are usually made by depositing small amounts of metals onto a support, here titanium dioxide. The difficulty is choice. Add up to five extra elements drawn from dozens of candidates, each at some loading, and the number of possible recipes runs far beyond what any laboratory can synthesise and measure one by one.
The authors set out to search that space of mixtures without testing it exhaustively. They began with a small set of measured activities for supported platinum catalysts, 45 data points, and built a loop: a model proposed promising compositions, the laboratory made and tested them, and the measurements went back into the model. The aim was to find mixtures producing carbon monoxide faster than a catalyst they had reported earlier.
Where AI came in
The model was an Extra-Trees regression, a method that averages many simple decision trees. Rather than being told which elements a catalyst contained, it saw each recipe as a set of eight general elemental properties, such as electronegativity, melting point, density, position in the periodic table and how strongly the oxide binds carbon dioxide, each weighted by how much of that element was present. From the model's prediction and its uncertainty, every recipe on a grid of candidates was scored by expected improvement, a measure that favours mixtures likely to beat the best result so far. Similar candidates were grouped by clustering into a shortlist of 100; the researchers then chose by hand which to make.
This scoring stood in for exhaustively trying compositions, and the clustering step stood in for sifting the ranked list by hand. The loop ran 44 times, with the model retrained on the growing dataset each cycle. The composition reported as best contained niobium, an element not present in the starting data. Afterwards the same model was examined with feature-importance scores and SHAP analysis, a technique that apportions a prediction among the inputs, to see which elemental properties its predictions leaned on.
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 authors ran a closed loop in which an Extra-Trees regression model, trained on catalyst activity data and fed only composition-scaled elemental descriptors, scored a grid of multi-elemental Pt/TiO2 catalyst compositions by expected improvement, and the shortlisted compositions were synthesised and tested for the reverse water-gas shift reaction. Starting from 45 data points and running 44 cycles, 300 catalysts were tested in total, including 255 newly predicted ones, and more than 100 showed higher activity than the previously reported Pt(3)/Mo(10)/TiO2 catalyst. The composition reported as optimal was Pt(3)/Rb(1)-Ba(1)-Mo(0.6)-Nb(0.2)/TiO2, which contains niobium, an element absent from the initial dataset. Cross-validated prediction accuracy reached R = 0.81 after 44 cycles, and SHAP and feature-importance analyses were used to relate predicted activity to descriptors such as periodic group, electronegativity and density.
How AI was used
Each catalyst was encoded not by its elemental identities but as a sum-pooled vector of eight composition-scaled elemental descriptors (electronegativity, melting point, enthalpy of fusion, density, periodic group, oxide band gap, oxide oxidation number and CO2 adsorption energy); a composition-only representation and a combined representation were also prepared. An Extra-Trees regression model implemented in scikit-learn was fitted to the current activity dataset, with hyperparameters chosen by grid search under 5-fold cross-validation and accuracy estimated by nested cross-validation using 100 repeated leave-20%-out external splits. The fitted model's predicted mean and standard deviation were used to compute an expected-improvement score for every point in a manually specified composition grid over five additive elements drawn from 50 candidate elements; candidates were grouped by clustering, typically with K = 100, into a top-ranking list from which compositions were picked by hand for diversity. Selected compositions were synthesised by sequential impregnation and tested in a fixed-bed reactor, and the measured CO formation rates were appended to the dataset to retrain the model for the next of 44 cycles. The same model was then analysed with feature-importance scores and SHAP summary and waterfall plots over the final 300-point dataset.
The shape of the work
Structural · the record, drawn
no AI
Assemble initial catalyst activity dataset
Obtaining raw data, whether by measurement, download or retrieval.
The initial dataset consisting of 45 data points was constructed using the catalysts reported in our previous experimental studywhere the paper describes this · verbatim
no AI
Encode catalysts as elemental-descriptor vectors
Encoding data into features, descriptors, embeddings or graphs.
each catalyst is represented as the sum of vectors of each elemental descriptor scaled by its compositional fractionwhere the paper describes this · verbatim
AI
Fit Extra-Trees surrogate on current dataset
Fitting model parameters, including fine-tuning an existing model.
We then trained the explorative ML model based on Extra-Trees regression (ETR) with the initial dataset (45 data points)where the paper describes this · verbatim
AI
Score composition grid by expected improvement
Running a trained model over new data to predict, classify or score. The AI stood in for exhaustive search.
calculated the expected improvement (EI) for all the test points in the catalyst composition gridwhere the paper describes this · verbatim
AI
Cluster candidates and shortlist for synthesis
Reducing a candidate set by filtering or ranking, in a single pass. The AI stood in for manual curation.
Clustering was typically performed to group very similar candidates into K clusters.where the paper describes this · verbatim
no AI
Synthesise and test catalysts, update dataset
Physical execution, by hand or by robot. Its result feeds back into an earlier step.
synthesized the catalysts using the sequential impregnation method, performed the RWGS reaction, and updated the dataset to close the loopwhere the paper describes this · verbatim
AI
Attribute predictions to elemental descriptors
Extracting understanding from model behaviour. The AI stood in for expert judgement.
Feature-importance score and SHapley Additive exPlanations (SHAP) analyses were used to understand the importance of the descriptors for ML predictionwhere 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.
All newly tested catalyst compositions, including the reported optimal composition, came from the model's expected-improvement rankings within a closed experimental loop; the paper's central claim is the discovery itself.
We used ETR as an ML model.where the paper describes this · verbatim
Through experimental testing of 255 ML-predicted new catalysts corresponding to 44 cycles of the closed loop discovery systemwhere the paper describes this · verbatim
All experimental data used for machine learning are available in Excel format on the URL and can be freely usedwhere 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.
- Version of Extra-Trees regression (ETR) surrogate, explorative elemental-property representationWhich version of the model was used is not stated.
- Version of Extra-Trees regression (ETR) surrogate, exploitative composition-plus-property representationWhich version of the model was used is not stated.
- Version of Extra-Trees regression (ETR) surrogate, naive composition-only representation (comparison only)Which version of the model was used is not stated.
- Version of Clustering of candidate compositions (K = 100)Which version of the model was used is not stated.
- What step 3 replacedThe paper gives no basis for what the AI stood in for.
About this article
Record aix-00178, version 2, checked by a person on 2026-10-09. The record describes the paper; it does not assess whether the paper's findings are right. The paper is published under CC-BY; quotations are at most 25 words. How we work · Report an error