COMPANY PROFILE / MATERIALS INFORMATICS
Google DeepMind
The research group that predicted 2.2 million "new" crystal structures with AI, working on the computing side of materials discovery
Google DeepMind is the AI research arm of Alphabet, Google's parent company. In materials science, it announced in 2023 that its deep-learning model GNoME had predicted 2.2 million new crystal structures to be stable on paper, which drew wide attention. That figure, however, is a prediction made inside a computer. It is not a count of materials that have been made in a laboratory. The announcement has since been followed by criticism from other researchers and, in January 2026, by a correction to the paper from its partner A-Lab. This profile separates the company's claims, the computed predictions, the experimental confirmations and the criticism, and uses only what primary sources let us confirm.
- Google DeepMind in 30 seconds, and where it stands in materials informatics
- The logo and how the name is written
- Where the company sits in the materials informatics field
- Scale (Alphabet as a whole, in brief)
- Resources committed to materials informatics
- What it has done so far
- What comes next
- What the company is aiming for in materials informatics
- Map of materials informatics partnerships
- Technology, products and services (how to read the GNoME numbers, and the criticism)
- The risks this company carries
- Glossary, references and claim-to-source audit
1. Google DeepMind in 30 seconds, and where it stands in materials informatics
Materials informatics (MI) means using data, computation and AI to find new materials more efficiently. Its workflow has three broad steps: "data, then prediction and design, then experimental confirmation". Google DeepMind mainly works on the prediction and design step.
California, U.S.Mountain View. Listed on Nasdaq (GOOGL, GOOG)
(primary type)An automated laboratory is planned as a secondary type (not yet confirmed)
380,000 of them the most stableJudged stable by DFT calculation. Not an experimental confirmation
Data: CC BY-NC 4.0The data is for non-commercial use (as stated on GitHub)
In the headline "AI discovers 2.2 million new materials", the word "discovers" refers to the number of crystal structures that were predicted to be stable, inside a computer Sourced. Stability was judged with density functional theory (DFT), a physics-based calculation. It is not a count of samples that were made and measured. The company itself says in the paper that synthesizability, meaning whether a material can actually be made, remains an open problem. Separately, 736 of the structures have been reported to match structures that other researchers obtained in laboratories. Throughout this profile we keep computed predictions and experimental confirmation apart.
2. The logo and how the name is written
A note on naming: in this article, "Google DeepMind" means the AI research organisation within Alphabet. We keep it separate from Alphabet Inc. (the parent company and the filer of the SEC documents), Google LLC (the operating company behind Google's products) and Google Research (a different research organisation; one co-author of the GNoME paper belongs to it). The 10-K reports figures for Alphabet as a whole, not for Google DeepMind alone.
Official website: https://deepmind.google/3. Where the company sits in the materials informatics field
As Fig. 1 shows, Google DeepMind belongs to the "foundation models and compute" type. Of the MI workflow (data, then prediction and design, then experiment), its centre of gravity is building AI from existing public data and large-scale computing, and producing predictions. GNoME's training data came from the Materials Project, a public materials database Sourced. The experimental check has been done by others: A-Lab at Lawrence Berkeley National Laboratory (Berkeley Lab) in the United States, and outside research groups.
A foundation model is an AI model trained on a very large amount of data that can then be reused for many purposes. In materials science, a model that answers "given how the atoms are arranged, how stable is this crystal?" is one example. Compute means the large amount of computing power needed to build such a model. The GNoME paper says it relied on hundreds of millions of first-principles calculations, of the kind explained below as DFT. For the technical detail, see our explainers on graph neural networks, crystal structure prediction and stability, and high-throughput computation.
4. Scale (Alphabet as a whole, in brief)
Google DeepMind is part of Alphabet. The documents filed with the SEC (the U.S. Securities and Exchange Commission) cover Alphabet as a whole, and revenue, headcount and budget for Google DeepMind alone are not disclosed Not yet confirmed. The figures below show only the size of the group Sourced.
| Metric (Alphabet group, US$ million) | 2023 | 2024 | 2025 |
|---|---|---|---|
| Total revenues | 307,394 | 350,018 | 402,836 |
| Research and development | 45,427 | 49,326 | 61,087 |
| Purchases of property and equipment | 32,251 | 52,535 | 91,447 |
| Item | Detail |
|---|---|
| Employees (31 December 2025, Alphabet group) | 190,820 |
| Headquarters | Mountain View, California, U.S. |
| Listing | Nasdaq Global Select Market (Class A: GOOGL, Class C: GOOG) |
| Mentions of Google DeepMind | In the 10-K and in the most recent 10-Q (quarter to June 2026), we could not find any mention of Google DeepMind, GNoME or materials science |
The figures above describe Alphabet as a whole. How much of the roughly US$61.1 billion of 2025 R&D went to Google DeepMind or to materials research is not disclosed. The text of the 10-K discusses investment in AI and the Other Bets (Waymo, and Isomorphic Labs, an AI drug-discovery company, among others), but it does not mention materials AI. The size of the company and the size of its commitment to materials AI are two different things.
5. Resources committed to materials informatics
Amounts and headcounts are not disclosed, but we can set out what official sources do and do not show, under four headings.
| Heading | What we could confirm | Type |
|---|---|---|
| People | Five of the six authors of the GNoME paper are affiliated with Google DeepMind (Mountain View) and one with Google Research. The A-Lab paper lists two co-authors from Google DeepMind (London). The total number of people working on materials AI is not disclosed | Sourced / not disclosed |
| Money | R&D spending, capital expenditure and computing investment for materials AI are not disclosed. For the UK automated laboratory, official sources give no amount and no headcount | Could not confirm |
| Investments, acquisitions, joint ventures, divestments | We could not confirm any investment, acquisition, joint venture or divestment concerning materials AI. Alphabet's 10-K lists Isomorphic Labs among the Other Bets, but that is a drug-discovery company, not a materials one | Could not confirm |
| Universities, research institutes and public funding | (1) Berkeley Lab (A-Lab): its paper appeared in Nature on the same day, 29 November 2023. (2) Materials Project: provider of the GNoME-derived data. (3) UK government (DSIT): a December 2025 memorandum covering support for launching an automated laboratory. (4) U.S. Department of Energy (DOE): in backing the Genesis Mission, the company says it will explore research collaboration that includes new materials discovery (no specific agreement was found). We could not confirm receipt of public research funding | Sourced / could not confirm |
Status of the main models, datasets and facilities
| Name | Kind | Main specifications and conditions | Release and licence | Stage and date |
|---|---|---|---|---|
| GNoME (model) | Graph neural network (predicts crystal energy and stability) | The model on GitHub was trained on the 2018 Materials Project snapshot, with an error of 21 meV/atom (README). The final model in the paper reaches 11 meV/atom and a stable-prediction hit rate above 80% (paper) | Model definitions and Colab code released (Apache-2.0). From the README we could not confirm release of trained weights | Released (November 2023) |
| GNoME (dataset) | Database of crystal structures judged stable by calculation | About 381,000 entries (most stable side). In August 2024 it was expanded to more than 520,000 entries by adding materials within 1 meV/atom of the convex hull (GitHub). DFT done with VASP, mainly the PBE functional | CC BY-NC 4.0 (non-commercial). A disclaimer says the data is for theoretical modelling only and should be used with caution | Released (November 2023 to August 2024) |
| NequIP (JAX implementation) | Architecture for machine-learning interatomic potentials | Used for the interatomic potentials trained on the GNoME data (paper). The original architecture is by Batzner et al. (2022) | JAX implementation and basic settings released (Apache-2.0) | Released (November 2023) |
| DM21 | Neural-network density functional | Claimed state-of-the-art accuracy on large parts of chemistry at the time (Science, 2021). Code released | The official blog says the code was open-sourced. We did not check the licence for this article | Released (December 2021) |
| Automated laboratory (UK) | Gemini-integrated robotic synthesis and characterisation facility | Described as able to synthesise and characterise hundreds of materials per day. We could not confirm the site, start date or equipment | Not applicable | Planned (announced for establishment in 2026) |
GNoME's training and evaluation both rely on DFT results as the yardstick Sourced. The stability DFT gives is a calculated value at absolute zero (0 K), with one particular functional. It does not include real temperature, atmosphere, impurities, defects or atoms swapping places (disorder). A practical way for a materials engineer to read a candidate is to check (1) its distance from the convex hull, meaning how far on the stable side it lies, (2) whether its composition is robust against disorder, and (3) whether a separate prediction exists for the function you are after, such as ionic conduction or superconductivity. The data is licensed CC BY-NC 4.0 (non-commercial), so any company should check the terms before using it in its own research.
6. What it has done so far
| When | What happened | What it means |
|---|---|---|
| December 2021 | Published DM21, a neural-network density functional, in Science (official blog) | Work to improve DFT accuracy, which underpins calculations in materials and chemistry |
| 29 November 2023 | Published GNoME in Nature. 2.2 million crystal structures predicted stable, 381,000 of them on the updated convex hull. Announced on the official blog | The company's claim. Mainly a computed prediction |
| 29 November 2023 | Berkeley Lab's A-Lab paper appeared in Nature the same day. It chose synthesis targets using calculations that included GNoME's and synthesised them automatically | The experimental side of the chain (a paper by an outside research group) |
| March 2024 | Leeman et al. published a critical paper on the A-Lab work in PRX Energy | Questioned the analysis of the experiments and the novelty claim |
| April 2024 | Cheetham and Seshadri published a paper examining the GNoME work in Chemistry of Materials | Questioned novelty, credibility and utility |
| August 2024 | Expanded the dataset on GitHub (more than 520,000 entries, including those within 1 meV/atom of the hull) | Gave the convex hull more context |
| December 2024 to June 2026 | GNoME-derived materials added to the Materials Project in stages, with more accurate r2SCAN calculations. A June 2026 update brought the total to 117,000 | Integration into a public database is under way (Materials Project change log) |
| 10 December 2025 | Announced a partnership with the UK government. Plan to establish its first automated laboratory in the UK in 2026 | A move from prediction alone to making and measuring in-house (a plan) |
| 11 December 2025 | Memorandum of understanding with the UK government (DSIT) published. Term of 36 months | Includes support for launching the automated laboratory |
| 18 December 2025 | Announced support for the U.S. DOE's Genesis Mission, including exploring collaboration on new materials discovery | We could not confirm a specific agreement |
| 19 January 2026 | Author correction to the A-Lab paper (Nature). Success count updated to "36 of 57 targets"; novelty wording corrected | Explained in detail in section 10 |
7. What comes next
All of these are company plans and targets, not accomplished facts Not yet confirmed. The December 2025 announcement says the UK automated laboratory will be established in 2026. On the date of this check (22 October 2026), we could not confirm from official sources that it is operating or has produced any first results.
None of the following appeared in official sources Not yet confirmed.
The laboratory's location, its start date, its equipment (types and numbers of instruments), investment and headcount, and specific materials it will target first. We also found nothing in official sources about a successor to GNoME or about new materials-specific releases.
8. What the company is aiming for in materials informatics
This section includes points the company has not stated directly. We give them as inferences drawn from primary sources, together with the evidence.
| What we read from the sources | Primary evidence | Type |
|---|---|---|
| The aim is to move from "prediction only" to "prediction plus its own experiments" | In December 2025 it announced a Gemini-integrated automated laboratory in the UK. GNoME was computation only, and the experiments were done by others (A-Lab and others) | Inference |
| Materials discovery is treated as one more application area for general-purpose AI such as Gemini | The laboratory is described as fully integrated with Gemini. The DOE blog says AlphaEvolve could be transformative in areas such as materials science | Inference |
| It releases data for the research community to use, but restricts commercial use | It supplied GNoME data to the Materials Project and published it on GitHub, under CC BY-NC 4.0 | Sourced (the release policy) / Inference (the intent behind the commercial restriction) |
| Materials AI looks like a relatively small effort within Alphabet as a whole | The 10-K does not mention materials, while its AI coverage is extensive | Inference |
The conclusion of the GNoME paper says that synthesizability (whether a material can actually be made) is an unresolved problem Sourced. The December 2025 announcement, for its part, says the automated laboratory will synthesise and characterise hundreds of materials per day and thereby "significantly shorten the timeline for identifying transformative new materials". From this we read that the company recognises that the next bottleneck after prediction, "making and measuring", is something it needs to own Not yet confirmed.
9. Map of materials informatics partnerships
10. Technology, products and services
What Google DeepMind has released for materials is model designs, datasets and Colab code. We could not find any paid service or product for materials. This section explains how GNoME works, the conditions behind its numbers, and the criticism and re-evaluation.
How GNoME works (illustration)
According to the paper, GNoME has two ways to generate candidates. (1) The structural route modifies known crystals by swapping elements (with adjusted substitution probabilities and a new method called symmetry-aware partial substitution). (2) The compositional route starts from chemical formulas without fixing a structure. In both, the AI narrows the field, DFT checks the survivors, and the results go back into training. This loop, called active learning, was repeated six times Sourced.
How to read the numbers: what was measured, and under what conditions
| Figure (claimed by the company and paper) | Conditions and definition | Nature of the figure |
|---|---|---|
| 2.2 million crystal structures | Structures that DFT judged stable relative to snapshots of the Materials Project, OQMD and similar databases taken in March 2021. The paper's wording is "stable with respect to previous work" | Computed prediction |
| 381,000 on the updated convex hull; 421,000 in total | The number left on the most stable side, because competing structures push the convex hull down. Previously known stable crystals numbered about 48,000 (external data as of July 2023) | Computed prediction |
| Energy prediction error of 11 meV/atom | The split is done by hashing the composition string, 85% for training and 15% for testing, so no composition appears on both sides. The initial model had 21 meV/atom | Model accuracy (calculation against calculation) |
| Stable-prediction hit rate above 80% (structural) and 33% per 100 trials (composition only) | The paper gives 1% for earlier work. The blog also describes the hit rate on Matbench Discovery as rising from about 50% to 80% | Model efficiency (calculation against calculation) |
| 736 structures independently realised in experiments | Structure matches against the ICSD (a database of experimentally determined structures, queried in January 2023). Of 4,235 composition matches, 4,180 could be parsed for structure and compared. The paper calls this likely a lower bound. 184 are new discoveries made after the project began | Experimental confirmation (but by outside studies) |
| 84% remain stable under a higher-accuracy calculation | Binary and ternary crystals recomputed with the more accurate r2SCAN functional | Computed prediction (checked with a different functional) |
The "2.2 million new crystals" include many that are the same structure type as a known one with a different composition. The paper itself counts about 45,500 novel structure types (prototypes) Sourced. In materials science, some people call a material "new" when it has a new composition, and others only when it has a new structure type. Unless you know which meaning is being used, the numbers cannot be compared.
Criticism and re-evaluation (reported neutrally)
| Source | Subject | Main points (as each source states) |
|---|---|---|
| Cheetham and Seshadri (Chem. Mater., April 2024) | The GNoME paper | Found scant evidence for compounds that combine novelty, credibility and utility. While saying the methods look promising, they argue that domain expertise in synthesis and crystallography needs to be built in |
| Leeman et al. (PRX Energy, March 2024) | The A-Lab paper (automated synthesis of targets chosen from GNoME and other predictions) | Examined the 43 reported products and concluded that no new materials had been discovered. Argued that automated Rietveld analysis is not yet reliable and that disorder (atoms swapping sites) is neglected. Claimed that two thirds of the successful materials are likely known, compositionally disordered versions |
| Author correction to the A-Lab paper (Nature, 19 January 2026) | The A-Lab paper | The authors explain that the novelty claim meant "new to the prediction platform", not necessarily new to science. Their re-analysis of the diffraction patterns found the platform's conclusion correct in 36 of 40 reported successes, with 4 inconclusive. The success count was updated to "36 of 57 targets", and one compound was removed from the discussion |
| The text of the GNoME paper | The GNoME paper | Lists synthesizability, competing polymorphs and dynamic stability as open problems. Also notes that structures could be pushed off the hull by future discoveries (it says GNoME itself displaced at least 5,000 previously "stable" materials) |
These are matters on which opinions differ, and this article does not decide who is right. What can be confirmed is the following. (1) For the A-Lab paper, the authors themselves corrected both the numbers and the wording. (2) On the check date, the official blog still contained the phrase "41 new materials". That differs from the paper's own figure (36 of 57 targets), so the blog figure should be read as belonging to an older version. (3) For the GNoME paper itself, we found no correction by its authors within the scope of this check.
Three points matter directly to a materials engineer Sourced.
(1) Use it as a candidate list. The data is a list of structures that look stable on paper, not a list of things you can make. To make one, you still have to work out precursors, firing temperature, atmosphere and how uniformly to mix, none of which the calculation includes. The A-Lab paper sorted the 17 of 57 targets it could not make into slow reaction kinetics (11 targets), precursor volatility, amorphisation and computational inaccuracy. A "stable" prediction is no guarantee that you can make the material.
(2) Watch out for disorder. As the critical papers point out, the structures an AI proposes are ordered ones in which every atom sits at a fixed site. In a real material, atoms may swap places and form a solid solution, which turns the same composition into a different substance.
11. The risks this company carries
| Risk | Detail | Type |
|---|---|---|
| Doubt over the "new materials" claim | Two outside peer-reviewed papers questioned novelty and the analysis of experiments, and the A-Lab paper was corrected by its authors. We found no correction of the GNoME paper itself, but how "novelty" is defined is a point of dispute | Sourced |
| Experimental confirmation depends on others | GNoME's validation rests on comparison with existing databases and on outside experiments. The company's own laboratory is still at the planning stage | Sourced |
| Limits on commercial use of the data | The GNoME data is CC BY-NC 4.0 (non-commercial). A company would have to check separately before using it in product development | Sourced |
| The automated laboratory has not been confirmed to be running | It is planned for 2026, but we could not confirm operation from official sources. Whether it proceeds as planned is unknown | Not yet confirmed |
| Continued investment in materials AI is invisible from outside | Alphabet's disclosures do not mention materials AI, and spending and headcount are not disclosed. Its standing as a business cannot be checked externally | Inference |
| Limits built into the calculations | The yardstick is DFT on ordered structures at 0 K. Temperature, disorder, defects and interfaces are not part of the evaluation | Sourced |
12. Glossary
- Materials informatics (MI)
- Using data, computation and AI to find new materials efficiently.
- GNoME
- Graph Networks for Materials Exploration. Google DeepMind's model for predicting crystal stability, and the database of its results.
- Graph neural network (GNN)
- An AI that takes a "graph" as input, with atoms as points and bonds as lines. Well suited to predicting crystals.
- DFT (density functional theory)
- A physics-based method that calculates the energy of an arrangement of atoms from the behaviour of its electrons. Accurate, but slow.
- Convex hull
- The line joining the most stable combinations for a given set of elements. A substance above the line would be more stable if it split into others.
- Active learning
- A loop in which an AI makes predictions, calculations check them, and the results go back into training, raising accuracy with fewer calculations.
- A-Lab
- A laboratory at Lawrence Berkeley National Laboratory in the U.S. that synthesises powder materials automatically with robots. It is not a Google DeepMind facility.
- ICSD
- A database of inorganic crystal structures that have been confirmed by experiment.
- Materials Project
- A large U.S. public database of calculated materials data.
- Machine-learning interatomic potential (MLIP)
- A way to calculate the forces between atoms quickly with AI, used in molecular dynamics simulation.
- CC BY-NC 4.0
- A licence that allows free use with attribution but not commercial use.
- 10-K
- The annual report that U.S.-listed companies file with the SEC.
13. References
- Alphabet Inc. Form 10-K for the fiscal year ended December 31, 2025 (filed with the U.S. SEC on 5 February 2026). Group-wide revenue, R&D, capital expenditure and headcount, business description and the Other Bets passage. Also used to confirm that Google DeepMind, GNoME and materials science are not mentioned. https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm
- Google DeepMind Millions of new materials discovered with deep learning (official blog, 29 November 2023). Overview of GNoME, the 2.2 million and 380,000 figures, 736 independent experimental realisations, and the A-Lab passage. https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/
- Nature Merchant, A. et al. “Scaling deep learning for materials discovery”, Nature 624, 80–85 (29 November 2023, DOI 10.1038/s41586-023-06735-9, CC BY 4.0). Methods, evaluation conditions and validation details. https://www.nature.com/articles/s41586-023-06735-9
- Google DeepMind (GitHub) README and DATASET.md of google-deepmind/materials_discovery. Contents of the release, licences (code Apache-2.0, data CC BY-NC 4.0) and the August 2024 data expansion. https://github.com/google-deepmind/materials_discovery
- Nature Szymanski, N. J. et al. “An autonomous laboratory for the accelerated synthesis of inorganic materials”, Nature 624, 86–91 (29 November 2023, DOI 10.1038/s41586-023-06734-w; the version reflecting the January 2026 correction was checked). https://www.nature.com/articles/s41586-023-06734-w
- Nature Author Correction: An autonomous laboratory for the accelerated synthesis of inorganic materials, Nature 650, E1 (19 January 2026). Wording of the novelty claim, re-analysis of the results and the updated success count. https://www.nature.com/articles/s41586-025-09992-y
- Chemistry of Materials (ACS) Cheetham, A. K.; Seshadri, R. “Artificial Intelligence Driving Materials Discovery? Perspective on the Article: Scaling Deep Learning for Materials Discovery” (8 April 2024, DOI 10.1021/acs.chemmater.4c00643). Bibliographic details checked through the Crossref API. https://doi.org/10.1021/acs.chemmater.4c00643
- PRX Energy (APS) Leeman, J. et al. “Challenges in High-Throughput Inorganic Materials Prediction and Autonomous Synthesis”, PRX Energy 3, 011002 (7 March 2024, DOI 10.1103/PRXEnergy.3.011002). https://doi.org/10.1103/PRXEnergy.3.011002
- Google DeepMind Our partnership with the UK government (official blog, 10 December 2025). Plan for an automated materials laboratory in the UK. https://deepmind.google/blog/strengthening-our-partnership-with-the-uk-government-to-support-prosperity-and-security-in-the-ai-era/
- GOV.UK Memorandum of Understanding between the UK and Google DeepMind on AI opportunities and security (published 11 December 2025). Support for launching automated R&D and laboratory facilities; term of 36 months. https://www.gov.uk/government/publications/memorandum-of-understanding-between-the-uk-and-google-deepmind-on-ai-opportunities-and-security/memorandum-of-understanding-between-the-uk-and-google-deepmind-on-ai-opportunities-and-security
- Google DeepMind Google DeepMind supports the U.S. Department of Energy on Genesis (official blog, 18 December 2025). Access programme for all 17 national laboratories and possible research collaboration including new materials discovery. https://deepmind.google/blog/google-deepmind-supports-us-department-of-energy-on-genesis/
- Materials Project Database Versions (change log). Additions of GNoME-derived materials (December 2024 to June 2026) and the BY-NC licence note. https://docs.materialsproject.org/changes/database-versions
- Google DeepMind Simulating matter on the quantum scale with AI (official blog, December 2021). Overview of the DM21 density functional (Science, DOI 10.1126/science.abj6511). https://deepmind.google/blog/simulating-matter-on-the-quantum-scale-with-ai/
- Google DeepMind Official website (source of the logo: the logo image used in the site navigation). https://deepmind.google/
14. Claim-to-source audit
| Claim in the text | Type | Source |
|---|---|---|
| Google DeepMind is part of Alphabet, and neither the 10-K nor the latest 10-Q mentions it by name, GNoME or materials science | Sourced | Reference 1 https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm |
| Alphabet's revenue, R&D and capital expenditure (2023 to 2025), 190,820 employees, headquarters and listing | Sourced | Reference 1 https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm |
| Revenue, headcount and budget for Google DeepMind alone are not disclosed | Not yet confirmed | Not in the 10-K. Reference 1 https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm |
| GNoME overview: 2.2 million crystals, 380,000 of them most stable, 52,000 layered compounds, 528 candidate lithium-ion conductors | Sourced | Reference 2 https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/ |
| 736 structures independently realised in experiments by outside teams | Sourced | References 2, 3 https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/ |
| Energy error of 11 meV/atom, hit rate above 80% (structural) and 33% (composition only), 1% in earlier work | Sourced | Reference 3 https://www.nature.com/articles/s41586-023-06735-9 |
| Composition-hash split of 85% training and 15% test, with no composition on both sides | Sourced | Reference 3 https://www.nature.com/articles/s41586-023-06735-9 |
| Six rounds of active learning giving 381,000 stable crystals; about 45,500 novel prototypes | Sourced | Reference 3 https://www.nature.com/articles/s41586-023-06735-9 |
| ICSD matching: 4,235 composition matches, 4,180 parsed, 736 structure matches, 184 of them new since the project began | Sourced | Reference 3 https://www.nature.com/articles/s41586-023-06735-9 |
| 84% of binary and ternary crystals remain stable under r2SCAN recalculation | Sourced | Reference 3 https://www.nature.com/articles/s41586-023-06735-9 |
| Synthesizability, polymorphs and dynamic stability as open problems; at least 5,000 existing stable materials displaced | Sourced | Reference 3 https://www.nature.com/articles/s41586-023-06735-9 |
| Author affiliations of the GNoME paper (five Google DeepMind, one Google Research) | Sourced | Reference 3 https://www.nature.com/articles/s41586-023-06735-9 |
| The GitHub model was trained on the 2018 Materials Project snapshot; model definitions and Colabs released; release of weights not confirmable from the README | Sourced | Reference 4 https://github.com/google-deepmind/materials_discovery |
| Code under Apache-2.0, data under CC BY-NC 4.0, theoretical-use disclaimer, August 2024 expansion to more than 520,000 entries | Sourced | Reference 4 https://github.com/google-deepmind/materials_discovery |
| GNoME-derived materials added to the Materials Project (December 2024 to June 2026, 117,000 in total, BY-NC) | Sourced | Reference 12 https://docs.materialsproject.org/changes/database-versions |
| A-Lab paper: 36 of 57 targets synthesised in 17 days of continuous operation (corrected version) | Sourced | Reference 5 https://www.nature.com/articles/s41586-023-06734-w |
| Author correction to the A-Lab paper (19 January 2026): novelty wording, four compounds excluded, count updated to 36 of 57 | Sourced | Reference 6 https://www.nature.com/articles/s41586-025-09992-y |
| Two Google DeepMind (London) authors on the A-Lab paper | Sourced | Reference 6 https://www.nature.com/articles/s41586-025-09992-y |
| The official blog still says “41 new materials” (as of the check date) | Sourced | Reference 2 https://deepmind.google/blog/millions-of-new-materials-discovered-with-deep-learning/ |
| Cheetham and Seshadri's doubts about novelty, credibility and utility | Sourced | Reference 7 https://doi.org/10.1021/acs.chemmater.4c00643 |
| Leeman et al.'s critique of the A-Lab paper (no new materials, neglected disorder) | Sourced | Reference 8 https://doi.org/10.1103/PRXEnergy.3.011002 |
| DM21: a neural-network density functional (December 2021) | Sourced | Reference 13 https://deepmind.google/blog/simulating-matter-on-the-quantum-scale-with-ai/ |
| Plan for a UK automated laboratory (2026, hundreds of materials per day, example target materials) | Not yet confirmed | Reference 9 https://deepmind.google/blog/strengthening-our-partnership-with-the-uk-government-to-support-prosperity-and-security-in-the-ai-era/ |
| UK government (DSIT) memorandum: support for automated R&D and laboratory facilities, 36 months | Sourced | Reference 10 https://www.gov.uk/government/publications/memorandum-of-understanding-between-the-uk-and-google-deepmind-on-ai-opportunities-and-security/memorandum-of-understanding-between-the-uk-and-google-deepmind-on-ai-opportunities-and-security |
| DOE Genesis Mission support: 17 national laboratories, collaboration on new materials discovery to be explored, AlphaEvolve's possible value | Sourced | Reference 11 https://deepmind.google/blog/google-deepmind-supports-us-department-of-energy-on-genesis/ |
| Laboratory location, start date, equipment, investment, headcount, and any successor to GNoME | Not yet confirmed | Not in official sources. Not estimated in this article. Reference 9 https://deepmind.google/blog/strengthening-our-partnership-with-the-uk-government-to-support-prosperity-and-security-in-the-ai-era/ |
| Investments, acquisitions, joint ventures or divestments in materials AI, and receipt of public research funding | Not yet confirmed | Could not confirm. We do not claim they do not exist either. Reference 1 https://www.sec.gov/Archives/edgar/data/1652044/000165204426000018/goog-20251231.htm |
| The logo image (navigation logo on the official website) | Sourced | Reference 14 https://deepmind.google/ |
| The four-type grouping of companies and the arrangement of Fig. 1 and Fig. 2 | Our calculation | Our own arrangement. Company names are examples from the companies covered in this series |
| The reading that the company wants to own the next bottleneck, "making and measuring" | Inference | Our own interpretation of the laboratory announcement and the conclusion of the GNoME paper |
| The reading that materials AI is a relatively small effort within Alphabet | Inference | Our own interpretation of the absence of materials in the 10-K |
| The reading of the intent behind the CC BY-NC restriction | Inference | Our own interpretation of the licence type |
Last updated 22 October 2026 / Troy Technical
The figures in this article come from Alphabet's 10-K filed with the U.S. Securities and Exchange Commission, Google DeepMind's official materials, peer-reviewed papers such as those in Nature, and publications by public bodies. We do not use market estimates from research firms or figures from news reports. Passages marked "Inference" are our own readings of primary sources and are not statements by the company. Figures such as "so many new materials" are presented as claims by the company or research group, and we keep computed predictions apart from experimental confirmation. We say nothing about share prices, valuation or investment decisions.