1. The short version / before you read / where this sits in the series

• In catalyst search, "adsorption energy" is the yardstick: how strongly a molecule sticks to a surface lowers activity if it is too strong or too weak (a volcano-shaped relation). Calculation and AI estimate this strength in bulk and narrow the candidates Sourced [Ref. 1, 6].
• There are examples found by calculation and confirmed by experiment: hydrogen evolution (BiPt), CO₂ to methanol (Ni–Ga) and CO₂ to ethylene (Cu–Al). On the other hand, for CO₂ reduction the computed features could hardly explain the experiments, even with the same approach Sourced [Ref. 3, 4, 5, 7].
• An error in calculation acts on rate exponentially: a barrier off by 0.1 eV makes the rate estimate off by about 48 times near room temperature. The baseline error of Open Catalyst in 2020 (around 0.5 eV) is too large to use for predicting rates as it stands Our calculation / Sourced.

2. What "searching for catalysts by calculation" means — adsorption energy and the volcano plot

(1) First, an everyday analogy

Think of what happens on a catalyst surface as "receiving a ball and passing it to the next person". If your grip is too weak, you cannot catch the ball at all (the reactant does not stay on the surface and the reaction does not start). If your grip is too strong, you cannot let go (the product stays on the surface and the next reactant cannot land). The right strength gives the fastest cycling. This is the Sabatier principle. Limit of the analogy: on a real surface, several steps proceed at once and the state of the surface changes during the reaction. It is not as simple as passing a ball.

(2) Turning the "right strength" into a number

The number that expresses "how readily something sticks" to a surface is the adsorption energy (the more negative, the stronger the bond). If you plot the adsorption energy of a key intermediate on the horizontal axis and activity on the vertical axis, a mountain shape appears. Because the shape resembles a volcano, it is called a volcano plot. From electronic-structure calculations, Nørskov et al. explained the trend in how fast oxygen reduction proceeds on various metals in terms of the adsorption energies of oxygen and hydroxyl Sourced [Ref. 6]. In other words, the adsorption energy plays the role of a "business card" (a descriptor) that expresses a catalyst as a number.

We check how a volcano plot arises with a small model we built (Fig. 3).

View original SVG, values and sources

Original SVG reference: Concept diagram (vector drawing). The flow of searching for catalysts with MI. The four steps on the left are prediction by calculation and AI; the two on the right are confirmation by experiment. The division is this article's own; the order and number of steps differ by method.

• Step that activates the reactant: the stronger the adsorption, the lower the barrier (barrier = 0.9 + 0.5×E, where E is the adsorption energy in eV)
• Step in which the product leaves: the stronger the adsorption, the higher the barrier (barrier = 0.2 + max(0, −E))
• We assume the slower step sets the overall rate (500 K). Activity is higher the lower the barrier

In this model, activity is highest at E ≈ −0.47 eV, where the two barriers are equal (Our calculation). On the left (too strong), activity changes by about 10 orders of magnitude per 1 eV of E; on the right (too weak), by about 5, so the slopes of the two sides differ. All numbers are assumptions for illustration, not values for a real reaction.

(3) Another guide: the d-band model

For transition-metal surfaces, a widely used idea (the d-band model) is that the position of the energy distribution of the metal's d electrons (the d band) is a guide to adsorption strength. The 1995 paper by Hammer and Nørskov, titled "Why gold is the noblest of all the metals", is known as one of the origins of this idea [Ref. 21] (we checked it only at the level of its title and bibliographic data). A fuller explanation is left to our explainer on the fundamentals of density functional theory.

3. Examples found by calculation and confirmed by experiment

Attempts to search for catalysts by calculation began before AI became common. Here we set out three examples in which calculation narrowed the candidates and experiment confirmed them, based on the papers' abstracts (Fig. 4).

• Hydrogen evolution (Greeley et al., 2006): they evaluated the activity and stability of more than 700 binary surface alloys by DFT and predicted that BiPt would have an activity equal to or better than Pt, the standard catalyst for hydrogen evolution. They synthesized and tested BiPt and report better hydrogen-evolution performance than pure Pt Sourced [Ref. 3].
• CO₂ to methanol (Studt et al., 2014): through descriptor-based analysis and calculation they identified Ni–Ga intermetallic compounds as stable candidates, and synthesized and tested several catalysts. They report that Ni₅Ga₃ was particularly active and selective, with methanol-synthesis activity equal to or better than the conventional Cu/ZnO/Al₂O₃ catalyst and considerably less CO by-product, at ambient pressure Sourced [Ref. 4].
• CO₂ to ethylene (Zhong et al., 2020): using DFT with active learning (a way of choosing candidates efficiently with few calculations) they found Cu–Al alloys. They report a Faradaic efficiency (the share of the electricity that went to the target product) above 80% (about 66% for pure Cu) at a current density of 400 mA/cm², and, by in situ X-ray absorption, a Cu coordination environment that favors C–C bond formation Sourced [Ref. 5].

4. Large datasets — the Open Catalyst Project and others

(1) What is public

The Open Catalyst Project is a joint effort of FAIR (Fundamental AI Research) at Meta and the Department of Chemical Engineering at Carnegie Mellon University. Its official site states the aim of using AI to model and discover catalysts for renewable-energy storage Sourced [Ref. 12]. The public resources are summarized in the table (checked 22 October 2026).

Sources: the abstracts and text of each paper, the official site, and the README of fairchem (checked 22 October 2026). Sourced.

(2) Licenses of models and data

The README of fairchem (Meta FAIR's collection of code) states that the code is under the MIT license and that model and weight licenses vary by application area. To use the UMA models, you must apply for access to the model on Hugging Face and log in Sourced [Ref. 13]. We could not confirm the terms of use of each dataset (whether commercial use is allowed and so on). Check the official statements before using them in your own work.

5. Reading accuracy figures — how an energy error acts on rate

(1) The 2020 baseline figures

The OC20 paper tried three graph neural networks (CGCNN, SchNet, DimeNet++) as baselines. On the task of predicting the relaxed energy from the initial structure (IS2RE), the results when trained on all the training data were as follows Sourced [Ref. 1].

• Mean absolute error (MAE) of energy: 0.50 to 0.92 eV (depending on the model and on the test split)
• The "answer with the median" baseline: 1.68 to 1.89 eV
• Fraction within 0.02 eV error (EwT): about 2 to 7%
• The test has four splits (same range as training / unseen adsorbate / unseen catalyst composition / both unseen). The more unfamiliar the split, the larger the error tends to be

For the S2EF metric, the paper says the results are "still far from practically useful" [Ref. 1]. From the fact that no upper limit on model size was seen in most tasks, it also points to the possibility of improvement with larger models [Ref. 1]. What we stress here is that these are baselines from 2020. Since then, data (OC22, OC25) and models (UMA and others) have been released, and present accuracy may be much better. But we did not check the latest figures. Check the date and conditions on the official leaderboard.

(2) Our calculation: by what factor is the rate off for a 0.1 eV error?

We calculate how serious the size of an energy error is, from the rate equation. The rate of a reaction is roughly proportional to exp(−Ea / kT) for a barrier (activation energy) Ea (k: the Boltzmann constant, T: temperature). If the barrier differs by ΔE, the rate differs by a factor of exp(ΔE / kT). We look at this numerically (Fig. 6). Our calculation

• 300 K (near room temperature): 0.05 eV → 6.9 times, 0.1 eV → 48 times, 0.2 eV → about 2,300 times, 0.5 eV → about 250 million times
• 600 K (an example of gas-phase thermal catalysis): 0.05 eV → 2.6 times, 0.1 eV → 6.9 times, 0.2 eV → 48 times, 0.5 eV → about 16,000 times

So a model with an error of around 0.5 eV cannot be used for a quantitative prediction such as "this catalyst is so many times faster than that one". What it can be used for is a screen that "ranks candidates roughly and reduces the number sent to experiment". However, the error of one method is often largely shared among the catalysts being compared (a systematic shift), and a ranking can be more accurate than an absolute value (our commentary; how far this holds has to be checked for each task).

(3) How the "structure" is chosen is also a source of error

Adsorption energy changes with where, and in what orientation, the molecule sits on the surface (the adsorption configuration). Finding the lowest-energy configuration is itself laborious work. AdsorbML is a method that searches for low-energy configurations with machine-learning potentials; in a balanced setting it found the lowest-energy configuration 87.36% of the time and made the calculation about 2,000 times faster. For this evaluation they also built the Open Catalyst Dense dataset, with about 1,000 surfaces and about 100,000 configurations Sourced [Ref. 8]. Put the other way, in the remaining 13% or so, the lowest configuration is missed (our calculation: 100% − 87.36%). Such a miss becomes an error that overestimates the energy (our commentary).

Kolluru et al. also state that machine-learning potentials for catalysts, when focused on specific chemistries and compositions, do well at interpolating between data but find it hard to generalize across chemical space, and position OC20 as an attempt to change that Sourced [Ref. 9].

6. Between calculation and experiment — what OCx24 showed

(1) What they did

To close the gap between the predictions of AI-based catalyst calculation and experimental results, Abed et al. built an experimental dataset (OCx24) (a 2024 preprint) Sourced [Ref. 7].

• Experiment: with two synthesis methods (chemical reduction and spark ablation) they synthesized 572 samples of 13 elements, and confirmed composition, purity and structure by XRF and XRD. In 441 gas diffusion electrodes (including replicates) they evaluated CO₂ reduction (CO2RR) and hydrogen evolution (HER) at up to 300 mA/cm². There were 230 independent preparations, which became 179 targets when similar compositions were grouped.
• Calculation: for about 19,400 materials (those in Materials Project, OQMD and Alexandria that could be thermodynamically stable under reaction conditions), they computed adsorption energies for six adsorbates (OH, CO, CHO, C, COCOH, H), combining AI and DFT. It took 685 million structural relaxations and about 20 million DFT single points. The paper calls it the largest computational catalyst screening.

(Our calculation: 685 million / about 20 million ≈ 34. That is the sense of scale: about 34 AI relaxations per DFT single point.)

(2) Result: hydrogen evolution worked; CO₂ reduction did not

• Hydrogen evolution (HER): a linear model with only the adsorption energies of hydrogen (H) and hydroxyl (OH) as features could explain the experimental voltage (R² = 0.59 under LOCO, which removes a composition entirely). The experimental scatter (297 replicates) is σ = 0.043 V. When this model was used to estimate the activity of about 19,400 materials, Pt appeared near the top of the volcano, even though Pt (including alloys) was not in the training data Sourced. The paper also lists 436 materials without Pd, Pt, Rh, Ir, Ru, Au, Os and Ag (282 of them containing Se) as candidates for low-cost HER catalysts. This is a prediction, not a result confirmed by experiment [Ref. 7].
• CO₂ reduction (CO2RR): production rates of each product were predicted with random forests. Under the evaluation where the same composition is on the training side (LOO), R² was 0.46 for H₂, 0.44 for CO and 0.43 for liquid products. But under the evaluation that removes the composition entirely (LOCO), it was 0.22 for H₂, 0.18 for liquid products, 0.00 for CO and 0.01 for ethylene, hardly explained at all. The paper concludes that the energies of isolated adsorbates plus elemental features are not enough to predict the complexity of this chemistry Sourced [Ref. 7].

From the trend that correlation improves as data grow, the paper expects that increasing the dataset to the 10,000 to 100,000 scale would allow much more predictive models [Ref. 7] (Not yet confirmed; this is the authors' expectation).

7. Learning from the literature — the bias in catalyst data

There is also learning from experimental data reported in the past rather than from calculation. This has also been attempted in catalysis. Suzuki et al., using literature data for the oxidative coupling of methane (OCM), the water-gas shift (WGS) and CO oxidation, showed that feeding in elemental features instead of catalyst compositions directly improves prediction accuracy; gradient boosting (XGBoost) was best, and for OCM they proposed the top 20 candidate catalysts Sourced [Ref. 16]. The same paper says that literature data come with obstacles: little compositional overlap, bias from previously published reports and low sample counts for many elements [Ref. 16].

For OCM there are also a report that combined high-throughput experiments with catalyst informatics [Ref. 17] and one that determined reaction conditions by machine learning [Ref. 18] (we checked them at the level of titles). A review that broadly organizes the application of machine learning to catalysis is by Toyao et al. [Ref. 19] (checked at the title level).

8. Running experiments automatically

Another way is to run the experiment itself automatically and let AI choose the next condition. Burger et al. used a mobile robot to search photocatalyst formulations for producing hydrogen from water. The robot operated autonomously for eight days, performing 688 experiments in a ten-variable experimental space, and with a batched Bayesian search it found photocatalyst mixtures six times more active than the initial formulations Sourced [Ref. 20]. (This is a search for photocatalyst mixtures, a different object from the search for metal catalysts.) The general explanation of robots and Bayesian optimization is in our explainers on autonomous laboratories and on Bayesian optimization and active learning. OCx24 also ran a high-throughput flow from synthesis, through characterization by XRD and XRF, to electrolysis tests [Ref. 7].

9. Telling prediction from experiment — the ladder of confirmation

We line up the reports so far by "how far they were confirmed". For catalysts, it helps to divide the stages into five rungs (this article's own arrangement).

1. Calculated prediction: prediction of adsorption energy, volcano plot or stability by calculation or AI
2. Synthesized, structure and composition confirmed: actually make it and confirm by XRD, XRF, in situ X-ray absorption and so on
3. Activity and selectivity at standard conditions: measure activity normalized by area or mass, and product fractions
4. Near-real conditions and stability: high current density, pressure, impurities, long tests
5. Independent check, scale-up: replication by other groups, reproduction in real devices or mass production

As Fig. 9 shows, OC20 is a dataset, a tool at rung 1 (prediction). The three examples from Greeley et al. to Zhong et al. went through rungs 1 to 3, and OCx24 covered rungs 1 to 3 and part of rung 4 (up to industrially meaningful current densities). But no report among those covered here could be confirmed through rung 5 (independent verification, scale-up). Long-term stability could not be confirmed within the abstracts.

10. A materials engineer's view

What can be left to AI, and where expert judgment is needed

The table is this article's own arrangement (commentary).

11. Limits, what could not be confirmed, and summary

(1) Latest accuracy

The Open Catalyst accuracy figures are baselines from the 2020 paper. We did not check the accuracy of later models (UMA and others) or the latest values on the official leaderboard.

(2) Papers cited from the abstract or title only

Greeley et al., Studt et al., Zhong et al., Burger et al. and Nørskov et al. are cited from their abstracts; Hammer and Nørskov (1995), Toyao et al., Nguyen et al. and Ohyama et al. from their titles and bibliographic data. OCx24 was checked in the text of the preprint (arXiv); we did not check whether a peer-reviewed published version exists.

(3) Licenses

We confirmed that the fairchem code is under the MIT license, but we did not check the terms of use of each dataset (whether commercial use is allowed and so on) or the licenses of each model's weights.

(4) What we do not cover

We do not cover microkinetic models of reaction rates, models of catalyst degradation and poisoning, homogeneous and enzyme catalysts, photocatalysis (except the one example of Burger et al.), or company case studies. For the supply and recycling of platinum-group metals, see our explainer on platinum-group metals and catalysts.

(5) Developments in Japan

We did not survey data infrastructure for catalyst research in Japan or the efforts of institutions and companies there.

12. How to read the yardsticks and standards

How good MI for catalysts is depends on what is taken as the correct answer (a DFT value or an experiment), how the evaluation data are split (whether compositions similar to training are included or removed: LOO or LOCO), what it is compared against (the median, a baseline catalyst) and how far it was confirmed (Section 9). We could not find an international standard specific to MI for catalysts. Here are the community benchmarks and shared procedures for experimental evaluation that we could confirm (checked 22 October 2026).

The arrangement of the table is Our calculation (arrangement). Each figure was checked in the paper's abstract or text, or on the official site (22 October 2026).

Cautions when reading (for materials engineers)

• A benchmark ranking is not the ability on your catalyst: a benchmark like OC20 evaluates the prediction of adsorption energies on calculated data. The activity, selectivity and stability of a real catalyst must be confirmed separately.
• Tell prediction from experimental confirmation: always check which rung of the ladder in Section 9 a result is at. "Found by AI" is often a rung 1 story.
• How the data are split: figures change a lot between LOO and LOCO. If you are searching for new compositions, look at the LOCO figure.
• How strong a principle or standard is: the FAIR principles are recommendations. The protocol of McCrory et al. is a proposal for a shared procedure, not a standard with legal force. The authors' expectation (that growing the data to 10,000 to 100,000 would improve prediction) is treated as Not yet confirmed.

13. Glossary

14. References

1. Chanussot L. et al. "Open Catalyst 2020 (OC20) Dataset and Community Challenges", ACS Catal. 11, 6059–6072 (2021), DOI 10.1021/acscatal.0c04525 (the preprint arXiv:2010.09990 was used to check the text) — arxiv.org/abs/2010.09990
2. Tran R. et al. "The Open Catalyst 2022 (OC22) Dataset and Challenges for Oxide Electrocatalysts", ACS Catal. 13, 3066–3084 (2023), DOI 10.1021/acscatal.2c05426 (abstract checked at arXiv:2206.08917) — arxiv.org/abs/2206.08917
3. Greeley J. et al. "Computational high-throughput screening of electrocatalytic materials for hydrogen evolution", Nat. Mater. 5, 909–913 (2006), DOI 10.1038/nmat1752 — doi.org/10.1038/nmat1752
4. Studt F. et al. "Discovery of a Ni-Ga catalyst for carbon dioxide reduction to methanol", Nat. Chem. 6, 320–324 (2014), DOI 10.1038/nchem.1873 — doi.org/10.1038/nchem.1873
5. Zhong M. et al. "Accelerated discovery of CO₂ electrocatalysts using active machine learning", Nature 581, 178–183 (2020), DOI 10.1038/s41586-020-2242-8 — doi.org/10.1038/s41586-020-2242-8
6. Nørskov J. K. et al. "Origin of the Overpotential for Oxygen Reduction at a Fuel-Cell Cathode", J. Phys. Chem. B 108, 17886–17892 (2004), DOI 10.1021/jp047349j — doi.org/10.1021/jp047349j
7. Abed J. et al. "Open Catalyst Experiments 2024 (OCx24): Bridging Experiments and Computational Models", arXiv:2411.11783 (November 2024, preprint) — arxiv.org/abs/2411.11783
8. Lan J. et al. "AdsorbML: a leap in efficiency for adsorption energy calculations using generalizable machine learning potentials", npj Comput. Mater. 9, 172 (2023), DOI 10.1038/s41524-023-01121-5 (open access) — nature.com/articles/s41524-023-01121-5
9. Kolluru A. et al. "Open Challenges in Developing Generalizable Large-Scale Machine-Learning Models for Catalyst Discovery", ACS Catal. 12, 8572–8581 (2022), DOI 10.1021/acscatal.2c02291 (abstract checked at arXiv:2206.02005) — arxiv.org/abs/2206.02005
10. Sahoo S. J. et al. "Insights into CO dimerization at electrified Cu interfaces from large-scale machine learning simulations" (introducing Open Catalyst 2025 [OC25]), arXiv:2509.17862 — arxiv.org/abs/2509.17862
11. Wood B. M. et al. "UMA: A Family of Universal Models for Atoms", arXiv:2506.23971 — arxiv.org/abs/2506.23971
12. Open Catalyst Project "Open Catalyst Project" (official site) — opencatalystproject.org
13. Meta FAIR Chemistry "fairchem" (README on GitHub; used to check the code license and the UMA task names; checked 22 October 2026) — github.com/facebookresearch/fairchem
14. Winther K. T. et al. "Catalysis-Hub.org, an open electronic structure database for surface reactions", Sci. Data 6, 75 (2019), DOI 10.1038/s41597-019-0081-y (open access) — doi.org/10.1038/s41597-019-0081-y
15. McCrory C. C. L. et al. "Benchmarking Hydrogen Evolving Reaction and Oxygen Evolving Reaction Electrocatalysts for Solar Water Splitting Devices", J. Am. Chem. Soc. 137, 4347–4357 (2015), DOI 10.1021/ja510442p — doi.org/10.1021/ja510442p
16. Suzuki K. et al. "Statistical Analysis and Discovery of Heterogeneous Catalysts Based on Machine Learning from Diverse Published Data", ChemCatChem 11, 4537–4547 (2019), DOI 10.1002/cctc.201900971 — doi.org/10.1002/cctc.201900971
17. Nguyen T. N. et al. "High-Throughput Experimentation and Catalyst Informatics for Oxidative Coupling of Methane", ACS Catal. 10, 921–932 (2020), DOI 10.1021/acscatal.9b04293 (checked at the title level) — doi.org/10.1021/acscatal.9b04293
18. Ohyama J. et al. "Data Driven Determination of Reaction Conditions in Oxidative Coupling of Methane via Machine Learning", ChemCatChem 11, 4307–4313 (2019), DOI 10.1002/cctc.201900843 (checked at the title level) — doi.org/10.1002/cctc.201900843
19. Toyao T. et al. "Machine Learning for Catalysis Informatics: Recent Applications and Prospects", ACS Catal. 10, 2260–2297 (2020), DOI 10.1021/acscatal.9b04186 (checked at the title level) — doi.org/10.1021/acscatal.9b04186
20. Burger B. et al. "A mobile robotic chemist", Nature 583, 237–241 (2020), DOI 10.1038/s41586-020-2442-2 — doi.org/10.1038/s41586-020-2442-2
21. Hammer B., Nørskov J. K. "Why gold is the noblest of all the metals", Nature 376, 238–240 (1995), DOI 10.1038/376238a0 (checked at the level of title and bibliographic data) — doi.org/10.1038/376238a0
22. Wilkinson M. D. et al. "The FAIR Guiding Principles for scientific data management and stewardship", Sci. Data 3, 160018 (2016), DOI 10.1038/sdata.2016.18 — doi.org/10.1038/sdata.2016.18

15. Claim-to-source audit

Last updated 22 October 2026. Sources are limited to primary material (peer-reviewed papers, arXiv papers, and the official pages of the Open Catalyst Project and fairchem). The Open Catalyst accuracy figures are baselines from 2020. The calculation examples are from our toy model and formulas and do not show the performance of any real catalyst. The latest accuracy, the terms of use of each dataset, an international standard specific to MI for catalysts, reports confirmed through independent verification or scale-up, and experimental confirmation of long-term stability could not be confirmed and are not stated here. All figures are explanatory concept graphics or drawings that include our calculations. Figs. 1 and 8 are ImageGen-generated conceptual illustrations, and none of them shows a real instrument or measurement result.