Skip to content
ContentLora

    Tip: press / anywhere to search.

    Analysis

    Has AI discovered new materials? The debate over GNoME and A-Lab

    Google DeepMind's GNoME predicted 2.2 million crystal structures and the autonomous A-Lab reported making dozens of targets, but chemists dispute how many are genuinely new or useful.[1][2][3] As of late 2025, reporting found no convincing big win for AI materials discovery, even as start-ups raised hundreds of millions to pursue it.[4][5]

    Editor reviewedUpdated AI for scienceMaterials scienceArtificial intelligenceScience
    Show:

    In materials science, AI discovery claims have drawn detailed pushback from chemists.[3][6] The dispute is less about whether the models work than about what counts as a “discovery”.

    The claims

    Google DeepMind‘s GNoME paper, published in Nature in November 2023, reported 2.2 million crystal structures below the known stability hull, including 381,000 newly discovered stable materials.[1] It described this as an order-of-magnitude improvement in discovery efficiency, building on 48,000 previously known stable crystals.[7] At publication, 736 of its stable structures had been made independently.[8]

    The A-Lab, an autonomous synthesis lab described in Nature the same day, used robots guided by computation, literature data and machine learning.[9] Its targets were picked using large-scale stability data from the Materials Project and Google DeepMind.[10] The current version of the paper reports 36 compounds realised from 57 targets in 17 days.[2]

    Nature’s one-line summary of the article still describes 41 novel compounds from 58 targets, the figures in the original version.[11][12] A January 2026 author correction explains the change. It says the novelty claims were meant to indicate materials new to the prediction platform, not necessarily new to science. A manual re-analysis of the diffraction data confirmed 36 of the 40 reported successes, with four inconclusive, and one compound was dropped because it had been in the training data.[12] Nature says that re-analysis was peer-reviewed after publication.[13]

    The pushback

    In March 2024 a PRX Energy perspective concluded that no new materials had been discovered in the A-Lab work, and that about two-thirds of claimed successes were likely known, compositionally disordered versions of the predicted compounds.[3] The authors also argued that automated Rietveld analysis of powder X-ray diffraction data is not yet reliable.[14]

    MIT Technology Review reported in December 2025 that researchers found scant evidence of GNoME compounds combining novelty, credibility and utility, describing some as merely trivial variations of known ones. It also reported that AI materials discovery had yet to produce a convincing big win.[6][4]

    Two things can be true at once. GNoME’s stability calculations may be accurate and useful as a map, while few individual predictions turn out to be materials anyone needs. “Stable at zero kelvin in a simulation” is a weaker claim than “new, made, characterised and better at something”. The A-Lab dispute shows the experimental side has its own failure modes: automated analysis can mislabel known alloys as new compounds. The 2026 correction partly concedes the critics’ point on wording, since “new to the platform” is a much weaker claim than “new to science”, while defending most of the syntheses themselves.

    The next wave

    Generative models now target properties directly: MatterGen, published in January 2025, produced structures more than twice as likely to be new and stable as earlier generators, and one synthesised example landed within 20% of its design target.[15][16] Investors are funding autonomous labs. Lila Sciences launched with $200 million in committed seed capital in March 2025, then closed a $350 million Series A in October 2025 with NVIDIA‘s venture arm among its backers.[17][18] Periodic Labs raised a $300 million seed in September 2025, aiming first at better superconductors.[5] In September 2026 Lila reported that its AI-directed lab had proposed, made and screened 2,942 green-hydrogen catalysts in three months; that result is company-reported and not yet peer-reviewed.[19] The US Genesis Mission asks DOE to build an AI platform that automates experiment design.[20]

    The money is betting that the bottleneck is experimental throughput and data quality, not model architecture. That is a reasonable reading of the A-Lab dispute, but it raises the bar: a self-driving lab is only as good as its characterisation, and its claims need the same independent checks as any human lab’s.

    What would count as proof

    The clearest evidence would be a material that was first proposed by an AI model, independently synthesised and characterised by another group, and shown to beat existing materials on a property that matters, such as a battery electrolyte or superconductor. We think such a result is plausible within two to three years given current funding, but expect early claims to be contested again. Watch for peer-reviewed results from Periodic Labs, Lila Sciences and Genesis Mission projects, and for independent replications. The superconductor angle is covered in the room-temperature superconductivity debate.

    Competing views

    A genuine step change

    Graph networks found 2.2 million structures below the known stability hull, an order-of-magnitude expansion, and 736 had already been made independently. Generative models such as MatterGen now hit target properties. Scale itself is the breakthrough.[1][8][16]

    Predictions are not discoveries

    Critics found that some predicted compounds are trivial variants of known ones, and that the A-Lab's claimed new materials were likely known disordered alloys. The A-Lab authors' 2026 correction narrowed 'novel' to 'new to the prediction platform'. Without careful experimental validation, numbers overstate progress.[6][3][14][12]

    The lab is the bottleneck, so build better labs

    The fix is to close the loop with autonomous synthesis and characterisation at scale. Start-ups and the US Genesis Mission are funding exactly that.[5][18][19][20]

    Questions readers ask

    How many new materials did GNoME find?

    It counted 381,000 newly discovered stable materials among 2.2 million predicted structures; 736 had been made independently at publication.[1][8]

    What is the criticism of the A-Lab?

    A 2024 PRX Energy perspective concluded no new materials were discovered and that two-thirds of claimed successes were likely known disordered compounds.[3]

    Did the A-Lab authors respond?

    Yes. A January 2026 correction said "novel" meant new to the prediction platform, not necessarily new to science, and a peer-reviewed re-analysis confirmed 36 of 40 reported successes, with four inconclusive.[12][13]

    Has AI produced a commercially important material?

    MIT Technology Review reported in December 2025 that there had been no convincing big win yet.[4]

    Sources

    Each numbered claim is a statement we checked against the sources listed with it. Status shows how well established it is.

    1. [1]

      Google DeepMind's GNoME work, published in Nature in November 2023, used graph networks to identify 2.2 million crystal structures, of which it counted 381,000 as newly discovered stable materials. confirmedas of 2023-11-29

    2. [2]

      The A-Lab, an autonomous lab for solid-state synthesis described in Nature in November 2023, realized 36 compounds from 57 targets over 17 days of continuous operation, according to the current version of the paper. confirmedas of 2026-10-10

    3. [3]

      A March 2024 PRX Energy perspective concluded that no new materials had been discovered in the A-Lab work and that about two-thirds of claimed successes were likely known disordered versions of predicted compounds. disputedas of 2024-03-07

    4. [4]

      MIT Technology Review reported in December 2025 that AI materials discovery had not yet produced a convincing big win or a new miracle material. reportedas of 2025-12-15

    5. [5]

      Periodic Labs emerged from stealth in September 2025 with a $300 million seed round, with a first goal of inventing better superconductors. confirmedas of 2025-09-30

    6. [6]

      MIT Technology Review reported that researchers who examined GNoME's predicted compounds found scant evidence of novelty, credibility and utility, describing some as merely trivial variations of known compounds. reportedas of 2025-12-15

    7. [7]

      GNoME built on 48,000 stable crystals identified in earlier studies and claimed an order-of-magnitude improvement in materials-discovery efficiency. confirmedas of 2023-11-29

    8. [8]

      The GNoME paper reported that 736 of its predicted stable structures had already been independently realized experimentally. confirmedas of 2023-11-29

    9. [9]

      The A-Lab combines computations, literature data, machine learning and active learning to plan and interpret experiments carried out by robots. confirmedas of 2023-11-29

    10. [10]

      The A-Lab's synthesis targets were identified using large-scale ab initio phase-stability data from the Materials Project and Google DeepMind. confirmedas of 2026-01-19

    11. [11]

      The A-Lab article's one-line summary on Nature's site describes 41 novel compounds discovered and synthesised from 58 targets in 17 days, more than its current abstract reports. confirmedas of 2026-10-10

    12. [12]

      A January 2026 author correction to the A-Lab paper said its novelty claims meant materials new to the prediction platform, not necessarily new to science; a manual re-analysis confirmed 36 of the 40 reported successes, with four inconclusive, and one compound was removed because it had been in the training data. confirmedas of 2026-01-19

    13. [13]

      Nature says the A-Lab diffraction re-analysis behind the 2026 correction was peer-reviewed after publication. confirmedas of 2026-01-19

    14. [14]

      The PRX Energy critics argued that automated Rietveld analysis of powder X-ray diffraction data is not yet reliable. confirmedas of 2024-03-07

    15. [15]

      MatterGen, a generative model for inorganic materials published in Nature in January 2025, produced structures more than twice as likely to be new and stable as earlier generative models. confirmedas of 2025-01-16

    16. [16]

      The MatterGen team synthesized one generated structure and measured its target property within 20% of the design value. confirmedas of 2025-01-16

    17. [17]

      Flagship Pioneering unveiled Lila Sciences in March 2025 with $200 million in committed seed capital to build a scientific superintelligence platform and fully autonomous labs for life, chemical and materials sciences. confirmedas of 2025-03-10

    18. [18]

      In October 2025 Lila Sciences closed a $350 million Series A, bringing its total funding to $550 million, with NVIDIA's venture arm NVentures among the investors. confirmedas of 2025-10-10

    19. [19]

      Lila Sciences reported in September 2026 that its AI-directed lab proposed, synthesised and screened 2,942 catalysts for green hydrogen in three months and identified six high-performing material families; the result is company-reported. reportedas of 2026-09-25

    20. [20]

      The Genesis Mission tasks the Department of Energy with integrating its supercomputers and datasets into a closed-loop AI platform intended to automate experiment design, accelerate simulations and generate predictive models. confirmedas of 2025-11-24

    Revision history (3)
    1. Page created.
    2. Added the January 2026 A-Lab author correction, which resolves the 36-versus-41 discrepancy; added Lila Sciences' Series A and its company-reported catalyst screen.
    3. Linked the Google DeepMind entity page.

    Created Oct 10, 2026. Last reviewed by an editor on Oct 10, 2026. Next scheduled review: Jan 10, 2027.

    Cite this page

    "Has AI discovered new materials? The debate over GNoME and A-Lab." ContentLora, updated Oct 10, 2026. https://contentlora.com/analysis/ai-materials-discovery-debate

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.

    Keep exploring

    ☰ Contents