GNoME
Last reviewed
Jun 3, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,592 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,592 words
Add missing citations, update stale details, or suggest a clearer explanation.
GNoME (Graph Networks for Materials Exploration) is a deep-learning system from Google DeepMind that predicts the thermodynamic stability of inorganic crystals and uses those predictions to search for new materials. It was introduced in a paper titled "Scaling deep learning for materials discovery," published in Nature on 29 November 2023.[1] The work reported the prediction of roughly 2.2 million stable crystal structures, of which about 381,000 were classed as genuinely new and on or near the energy convex hull, a result the authors framed as an order-of-magnitude increase in the number of stable inorganic materials known to science.[1][2] The release was widely covered as a landmark for AI in the physical sciences, and it also drew pointed criticism from materials chemists who questioned how many of the predicted compounds are truly novel, credible, and useful.[3][4]
Most of the materials that power batteries, solar cells, superconductors, and microelectronics are inorganic crystals: ordered arrangements of atoms in a repeating lattice. Finding a new one that is both useful and physically realizable is hard. A candidate composition can be arranged in many ways, and only some arrangements are thermodynamically stable enough to form and persist. Historically, chemists narrowed the search using intuition, known structural templates, and slow trial-and-error synthesis, supplemented from the 2000s onward by quantum-mechanical screening.
A central tool in that computational effort is the Materials Project, an open database founded in 2011 at Lawrence Berkeley National Laboratory that catalogs the computed properties of known and predicted materials.[2] Before GNoME, such databases held on the order of 48,000 computationally stable inorganic compounds.[1] The bottleneck was cost: the standard way to judge stability is density functional theory (DFT), an accurate but expensive quantum simulation, which makes brute-force exploration of the vast space of possible compositions impractical. GNoME was designed to use machine learning to triage that space, reserving DFT for the most promising candidates. The approach is conceptually adjacent to DeepMind's earlier success with AlphaFold in protein structure, in that a learned model is used to predict a physical property far faster than first-principles calculation.
GNoME is built on graph neural networks (GNNs), a class of model well suited to crystals because atoms and the bonds between them map naturally onto the nodes and edges of a graph.[2] The networks use a message-passing formulation in which information propagates between neighboring atoms, allowing the model to estimate a structure's formation energy and therefore its stability.[1]
The defining feature of the project is an active-learning loop. The model proposes candidate structures and predicts which are likely to be stable; the most promising are then verified with DFT calculations (run with the VASP software), and the verified results are folded back into the training set for the next round.[1][2] Over six iterative rounds, this cycle drove the prediction error down to about 11 meV per atom and raised the discovery hit rate substantially.[1]
GNoME generates candidates through two complementary pipelines:
| Pipeline | Method | Reported hit rate |
|---|---|---|
| Structural | Modifies known crystal templates via ionic substitution, including symmetry-aware partial substitutions (SAPS) | greater than 80% |
| Compositional | Starts from chemical formulas with randomized structures, less reliant on existing templates | about 33% |
The structural pipeline is more reliable because it leans on the geometry of materials already known to exist, while the compositional pipeline ranges more freely and so is riskier but more exploratory. For comparison, the authors noted that an earlier composition-only approach had a hit rate near 1%.[1]
DeepMind reported that GNoME identified about 2.2 million structures that are stable with respect to previous work. Filtering for those that are genuinely new yielded 381,000 fresh entries, bringing the total catalog of stable inorganic crystals to roughly 421,000.[1][2] The paper characterized this as "an order-of-magnitude expansion in stable materials known to humanity."[1]
| Quantity | Figure |
|---|---|
| Structures predicted stable vs. prior work | about 2.2 million |
| New stable entries (on or near the convex hull) | about 381,000 |
| Total stable inorganic crystals after GNoME | about 421,000 |
| Approximate prior baseline | about 48,000 |
| Predictions already realized experimentally (concurrent work) | 736 |
The authors also pointed to external corroboration: at the time of publication, 736 of the predicted structures had already been independently created in laboratories around the world through separate, concurrent research.[1][2] DeepMind further highlighted candidate materials of practical interest, including hundreds of thousands of potential layered compounds and a set of candidate lithium-ion conductors relevant to battery research.[2]
Rather than keep the predictions internal, DeepMind released them to the research community and contributed the 380,000 most stable candidates to the Materials Project, where they joined a decade of openly available computational data.[2] Kristin Persson, the founder and director of the Materials Project, presented the contribution as a way to accelerate work on energy and climate challenges by giving experimentalists a much larger pool of vetted candidates to draw from.[2]
The contribution coincided with a separate but related demonstration. An autonomous laboratory at Berkeley known as the A-Lab, led by Gerbrand Ceder and reported in its own Nature paper, attempted to synthesize a set of target compounds using recipes proposed by computation. Over 17 days it produced 41 compounds out of 58 attempts, a 71% success rate, and was offered as evidence that AI-proposed materials could be made in practice.[2] DeepMind also published the GNoME code and dataset on GitHub. The dataset was later expanded; as of August 2024 it included more than 520,000 materials within 1 meV per atom of the convex hull, distributed under a Creative Commons Attribution-NonCommercial license.[5]
GNoME's claims attracted scrutiny on two related fronts.
The first concerns the GNoME predictions themselves. In a perspective published in Chemistry of Materials in April 2024, Anthony Cheetham and Ram Seshadri of the University of California, Santa Barbara argued that the predictions, while numerous, showed "scant evidence for compounds that fulfill the trifecta of novelty, credibility, and utility."[3][4] They stressed that the 2.2 million entries are crystalline inorganic compounds rather than functional materials in a working sense, since none had a demonstrated application, and that the survey excludes large classes of real-world materials such as polymers, glasses, metal-organic frameworks, heterostructures, and composites.[4] In their framing, a predicted stable compound is a starting point, not a finished material.
The second front concerns the A-Lab synthesis results, which are distinct from GNoME but were presented alongside it. In an analysis posted to the ChemRxiv preprint server on 31 January 2024, Robert Palgrave of University College London and Leslie Schoop of Princeton University, with collaborators, contested the claim that the A-Lab had produced large numbers of genuinely new materials.[6][7] They argued that many of the reported compounds were ordered versions of materials already known in disordered form, that the analysis overlooked compositional disorder and atomic site mixing, and that the automated interpretation of X-ray diffraction data fell short of competent human refinement.[6] Palgrave went so far as to say the main discovery claim was wrong. Ceder defended the work in a December 2023 response, supplying additional data while acknowledging that a human could perform higher-quality structural refinement.[7] The dispute was long-running: in January 2026 Nature issued a correction to the A-Lab paper, softening the language so that the synthesized materials are no longer described as necessarily new to science.[8]
A separate, practical limitation runs through both lines of criticism: thermodynamic stability is not the same as synthesizability. A compound can sit on the convex hull yet have no known route to being made, and turning the predicted catalog into laboratory materials remains an open problem that later work has tried to address with dedicated synthesizability models.
DeepMind has stood by the original paper. A spokesperson noted that hundreds of the predicted materials had already been independently synthesized by other scientists and said the company stands by all the claims in its Nature publication.[3]
Even allowing for the disputes, GNoME marked a shift in scale for computational materials discovery. It demonstrated that a graph network coupled to an active-learning loop and DFT verification could enlarge a curated database of stable inorganic crystals by roughly an order of magnitude, and that releasing those predictions openly could seed downstream experimental and modeling work. The criticism, in turn, sharpened a distinction that matters for the whole field: between a structure that is predicted to be stable, a compound that is genuinely new, and a material that can be made and put to use. GNoME advanced the first of these dramatically, and made the gap to the others a central question for AI-driven materials science.