GNoME (DeepMind)

This article is about the Google DeepMind materials-discovery system. For the free software desktop environment, see GNOME. For the mythological creature, see Gnome.

GNoME

Developer(s)	Google DeepMind
Initial release	November 29, 2023 (2023-11-29)
Repository	github.com/google-deepmind/materials_discovery
Engine
Type	Machine learning system for materials science
License	Apache License 2.0 (code); CC BY-NC 4.0 (dataset)
Website	deepmind.google/discover/blog/millions-of-new-materials-discovered-with-deep-learning/

Search GNoME (DeepMind) on Amazon.

GNoME (Graph Networks for Materials Exploration) is an artificial intelligence system developed by Google DeepMind for materials discovery. It uses graph neural networks and active learning to predict stable inorganic crystal structures, with candidates evaluated using density functional theory calculations.

The system was announced in November 2023, when Google DeepMind reported that GNoME had identified approximately 2.2 million candidate crystal structures, including about 380,000 predicted stable materials. The results were published in Nature and the predicted stable materials were made available through the Berkeley Lab Materials Project, an open-access database for computed materials properties.

Background

The discovery of stable inorganic crystals is a major problem in materials science, solid-state chemistry, and computational materials science. New materials are important for development of semiconductors, solar cells, catalysis, thermoelectric materials, , nuclear fuels, superconductivity, lithium-ion batteries, and other battery technologies. Traditional discovery methods often rely on experimental trial and error, while computational approaches use physics-based calculations to estimate whether a proposed material is likely to be stable.

Before GNoME, databases such as the Materials Project, the Inorganic Crystal Structure Database, and other computational materials databases were used to identify and screen candidate structures. GNoME was developed to expand this search by using machine learning models trained on crystal-structure and stability data.

Methodology

GNoME treats candidate crystal structures as graphs. In this representation, atoms are modeled as nodes and relationships between neighboring atoms are modeled as edges. A simplified graph representation can be written as:

$G=(V,E)$

where $V$ is the set of atomic sites and $E$ is the set of edges representing neighboring atomic interactions. This makes graph neural networks suitable for modeling crystals, because a crystal structure can be described by both the identities of the atoms and their geometric arrangement.

The system used two main candidate-generation pipelines. A structural pipeline generated new candidates from known crystal structures, while a compositional pipeline searched over chemical formulas without requiring a known structure as input. Candidate materials were filtered by the machine-learning model and then evaluated with density functional theory. The results of those calculations were added back into the training data, creating an iterative active-learning loop.

Stability notation

A central objective of GNoME is to predict whether a candidate material is thermodynamically stable. For a compound with total energy $E_{\text{total}}$, its formation energy can be written in simplified form as:

$E_f=E_{\text{total}}-\sum _i{n}_i{\mu }_i$

where $E_f$ is the formation energy, $n_i$ is the number of atoms of element $i$, and ${\mu }_i$ is the chemical potential of that element. In computational materials screening, candidate compounds are often compared against other possible phases. A material is considered stable if it lies on the convex hull of lowest-energy competing phases.

The energy above the convex hull can be expressed as:

$E_{\text{hull}}=E_f-E_{\text{hull}}^{\min }$

where $E_{\text{hull}}^{\min }$ is the minimum formation energy available from competing stable phases at the same overall composition. Materials with $E_{\text{hull}}=0$ are on the convex hull and are predicted to be stable, while materials above the hull may be metastable or unstable.

Applications

The predicted materials from GNoME were described as candidates for further screening in areas such as electronics, energy storage, photovoltaics, and solid-state ionics. The Nature paper highlighted layered materials and lithium-ion conductors as examples of material families that could be explored using the larger catalogue of predicted stable crystals.

Some GNoME-derived data was also used in connection with A-Lab, an autonomous materials-synthesis laboratory at Lawrence Berkeley National Laboratory. A related Nature paper reported that A-Lab used data from the Materials Project and Google DeepMind in an automated workflow for synthesizing inorganic materials.

Reception and limitations

GNoME was covered as part of a wider trend toward AI-assisted scientific discovery and automated materials research. The project was also compared to other large-scale uses of machine learning in science, such as AlphaFold in protein-structure prediction.

Some researchers later raised concerns about duplication and novelty in AI-generated crystallographic datasets. In 2025, Chemical & Engineering News reported that researchers had identified possible exact and near-duplicate crystal structures in the GNoME database and argued that such issues should be labeled or corrected in crystallographic databases. The related A-Lab study was also corrected by Nature in 2026 after criticism of its original claims about newly synthesized materials.

See also

• Artificial intelligence in chemistry
• Cheminformatics
• Computational chemistry
• Crystal structure prediction
• Density functional theory
• Drug discovery
• Generative artificial intelligence
• History of materials science
• Machine learning in physics
• Materials Project
• Protein design
• Timeline of materials technology

External links

• Official website
• materials discovery on GitHub
• GNoME Explorer at the Materials Project

This article "GNoME (DeepMind)" is from Wikipedia. The list of its authors can be seen in its historical and/or the page Edithistory:GNoME (DeepMind). Articles copied from Draft Namespace on Wikipedia could be seen on the Draft Namespace of Wikipedia and not main one.

---