GNoME (Graph Networks for Materials Exploration)

What GNoME is

GNoME, short for Graph Networks for Materials Exploration, is a deep-learning tool for discovering inorganic crystals. Google DeepMind announced it on 29 November 2023 together with a paper in Nature.^[1] The problem it addresses is one of scale: the number of ways to arrange the elements of the periodic table into a crystal is enormous, and making and testing each candidate in a laboratory is slow and expensive. GNoME does the first pass on a computer, proposing candidate structures and estimating whether they would hold together.

How it works

GNoME is a graph neural network.^[1] A crystal can be described as a graph: atoms are the nodes, and the connections or close contacts between them are the edges. Trained on known materials, such a network learns relationships between structure and properties, and can then make predictions for combinations it has never seen.

Those predictions were checked with Density Functional Theory (DFT) during active learning.^[1] DFT is a standard computational method in chemistry and condensed-matter physics: it approximates the behaviour of electrons in a material and from that estimates energies and other properties. It is far slower than a neural network but gives more reliable numbers, so the two are used together. In active learning the model proposes candidates in bulk, the slower method evaluates part of them, and the results are fed back into training.

On MatBench Discovery, a benchmark for materials-stability prediction, GNoME raised the discovery rate from around 50% to 80%.^[1]

Results

Google DeepMind said the tool led to the discovery of 2.2 million new crystals, equivalent to nearly 800 years' worth of knowledge.^[1] Of these, 380,000 were predicted to be the most stable, making them candidates for experimental synthesis.^[1] With GNoME, the number of known stable materials rises to 421,000.^[1]

The predictions were also tested outside the company. External researchers had independently created 736 of the new materials in the lab.^[1] In a second Nature paper, a Berkeley Lab robotic laboratory working with Google DeepMind synthesised more than 41 new materials.^[1] Google DeepMind named six of the independently created materials: K2BiCl5, Li4MgGe2S7, Mo5GeB2, KV3Se3, Rb2HfSi3O9 and Tm5Pd9P7.^[1]

Data and where it sits

Google DeepMind said it would contribute the 380,000 predicted-stable materials to the Materials Project database.^[1] The dataset is available on GitHub under google-deepmind/materials_discovery.^[1]

New inorganic materials matter because a great deal of technology depends on finding a compound with the right combination of properties — a battery electrode that holds charge, a catalyst that speeds a reaction, a material that conducts or absorbs light in a particular way. The traditional route is guided by chemical intuition and trial and error, one laboratory experiment at a time. GNoME belongs to a different approach: use machine learning to search candidate structures at scale, rank them by predicted stability, and hand the shortlist to experimentalists. The pairing with a robotic laboratory points at one way this could work in practice, with a model proposing candidates and automated equipment testing them. How many of the predicted materials will turn out to be useful, and in which applications, is a question that only laboratory work can answer.