Scaling deep learning for materials discovery

The problem

A new material is only useful if it holds together. For a candidate crystal structure, the first test is usually thermodynamic stability: whether the structure has a lower formation energy than any combination of competing phases at the same composition. Researchers summarise this with the convex hull, the lower envelope of the formation energies of all known phases; a structure that sits below the hull is predicted to be favoured to form rather than to decompose into something else. Searching the space of possible compounds for such structures is expensive, because each candidate normally has to be checked with its own quantum-mechanical calculation.

The approach

The paper is by Amil Merchant, Simon Batzner, Samuel S Schoenholz, Muratahan Aykol, Gowoon Cheon and Ekin Dogus Cubuk ^[1]. It appeared in Nature, volume 624, issue 7990, pages 80–85, on 29 November 2023, with DOI 10.1038/s41586-023-06735-9 ^[1].

The title names the central move: scaling deep learning for materials discovery. The work uses a deep-learning model to predict the stability of candidate structures, so that a very large pool of compositions can be screened cheaply before anything is computed in detail or made in a laboratory. The authors state the payoff in terms of rate: the abstract says the approach improves "the efficiency of materials discovery by an order of magnitude" ^[1].

Results

The headline is the size of the output. The abstract reports that the work "enables the discovery of 2.2 million structures below the current convex hull" ^[1]. That number is a set of predictions rather than synthesised compounds, but it is not purely hypothetical: of the stable structures, 736 have already been independently realised in experiments ^[1][2]. Those 736 serve as a check on the method, because they were made by other groups rather than by the model.

The Google DeepMind blog post published the same day supplies the accuracy behind the list: the work raised the success rate of materials stability prediction "from around 50%, to 80%" ^[2]. The post also sorts the output: of the 2.2 million predictions, 380,000 are the most stable, "making them promising candidates for experimental synthesis" ^[2].

Why it matters

Materials discovery has long been limited by the cost of evaluation, and both headline numbers attack that limit. An order-of-magnitude improvement in efficiency ^[1] means the same effort can cover far more candidates, and a jump from roughly half to 80% correct ^[2] means fewer of those candidates are wasted. Together they change screening from a step that has to be rationed into one that can be run at scale.

The second consequence is that the result is a list someone can act on. Google DeepMind said it would contribute the 380,000 materials it predicts to be stable to the Materials Project ^[2], a shared database of materials information. That turns a computational result into a starting point for experimental groups, who can pick candidates instead of searching blind.

What it does not settle

A structure below the convex hull is stable in the thermodynamic sense, but that does not guarantee it can be made: whether a crystal actually forms depends on kinetics, temperature, pressure and available precursors, none of which the hull captures. Of millions of predictions, 736 are confirmed by experiment ^[1][2], a small fraction. The 80% figure is also a rate over a test setting, not a promise about any single compound. What the work changes is where the bottleneck sits — from guessing which structures to try to testing the ones already on the list.