GNoME: Google DeepMind's 380,000 New Materials, and How to Pull Them Free

GNoME is a Google DeepMind deep learning tool that predicted 2.2 million new crystal structures, of which 381,000 are newly stable. About 380,000 of the most stable were added to the Materials Project, a free public database you can query today.

You pull them two ways. Query the Materials Project with a free API key and the Python client mp-api, or download DeepMind's raw files from a public Google Cloud bucket, then filter by property.

What is GNoME and what did it actually find?

GNoME stands for Graph Networks for Materials Exploration. It is a deep learning model that predicts whether a new material will be stable, meaning it will hold its shape instead of falling apart.

DeepMind trained it, then used it to scan far more candidate crystals than people could test by hand. The result, published in Nature on November 29, 2023, was 2.2 million structures below the known convex hull. Of those, 381,000 were newly discovered stable materials.

The paper calls this an order-of-magnitude expansion in the stable materials known to humanity. Earlier work had identified about 48,000 stable crystals. With GNoME, the count of stable materials known now sits at roughly 421,000.

Sources: DeepMind | Nature

Where can I get the GNoME data for free?

There are two doors, and both are open to the public.

• The Materials Project. This is a free, open database run by Berkeley Lab. DeepMind contributed roughly 380,000 of the most stable GNoME materials to it. You browse it in the Materials Explorer app or query it with code.
• DeepMind's raw release. The full dataset lives in a public Google Cloud Storage bucket at gs://gdm_materials_discovery. You download it with the Google Cloud CLI command: gcloud storage cp --recursive gs://gdm_materials_discovery/ data/. Inside you get stable_materials_summary.csv plus CIF structure files grouped by ID, formula, and composition.

One rule to note. The code is Apache 2.0, but the data is Creative Commons Attribution NonCommercial 4.0.

That last part matters if you plan to build a product on top of it, so read the license before you ship. In August 2024, DeepMind changed the stability cutoff to 1 meV per atom, which significantly increased the number of materials in the release.

Sources: GNoME dataset (GitHub) | GNoME README (license) | Materials Project

How do I query the Materials Project for real leads?

You need a free key and the Python client. Here is the short version.

1. Log in at materialsproject.org, open your dashboard, and copy your free API key.
2. Install the client with: pip install mp-api.
3. In Python, import MPRester from mp_api.client and open it with your key.
4. Call mpr.materials.summary.search() and filter. For example, ask for materials that contain silicon and oxygen with a band gap between 0.5 and 1.0 eV.

Here is what one run looks like. You pass the search a pair of elements and a band gap window, and it hands back a short list of material IDs that fit. You then feed each ID into get_structure_by_material_id to see the atoms, or get_bandstructure_by_material_id to pull the deeper electronic data on a single hit.

That is the same discover-then-verify loop showing up across AI for science, from crystals to biology.

Sources: Materials Project docs

Which columns matter for battery, solar, and electronics leads?

The GNoME summary CSV carries the fields you need to screen without opening every file. Match the column to the job.

• Solar and electronics. Sort by Bandgap. Solar cells and semiconductors live in specific band gap ranges, so this one number filters most of the noise fast.
• Batteries. The Nature work also produced learned interatomic potentials that give zero-shot predictions of ionic conductivity, which is the property that moves lithium and other ions through a battery. Start from lithium-containing candidates, then test conductivity.
• Everything. Use Decomposition Energy as your reality filter. The lower it goes, the more likely the material actually holds together. I call this the decomposition-energy screen, and it is the fastest way to cut 380,000 rows down to a short list worth a real look.

Other columns in the file include Formation Energy, Density, Crystal System, and space group. Sort by Density when weight matters, since density is mass per volume. Filter by Crystal System or space group when a project needs a specific atomic arrangement rather than any stable match. For a lightweight solar cell, stack a band gap filter on top of a low Density cutoff to screen for the right optics and less mass at once.

Sources: GNoME dataset (GitHub)

How do I know a candidate is real and not just a prediction?

You do not, until someone makes it. That is the honest part. GNoME predicts stability from physics-based math, not from a finished lab sample.

But the predictions have started to hold up. DeepMind reported that 736 of the new materials had already been independently created in labs by other teams around the same time. Berkeley Lab's autonomous facility, called A-Lab, made 41 new materials on its own using GNoME stability data plus the Materials Project.

So treat a hit as a strong lead, then confirm it. A low decomposition energy plus an existing lab report is the closest thing to a green light.

Sources: DeepMind

What is the catch before I bet a project on this?

Three things keep this grounded. First, a prediction is not a product. Making and testing a material still costs time and money.

Second, the noncommercial license on the data limits what you can sell directly on top of it. Third, these are computed results, so a real synthesis can still fail.

The upside is that you are searching a map that did not exist before 2023, and most people have not looked at it yet.

Sources: Nature | GNoME dataset (GitHub)

If you want help wiring a GNoME screen into an actual pipeline, that is the kind of thing we build over at nomadtechnologist.com.

---

Not legal, financial, or tax advice.