Discovered Materials opens benchmark: AI can find chip materials but can rarely turn them into synthesis recipes

Discovered Materials has opened an AI benchmark: of 526 computationally identified chip materials, only one recipe was rated “worth trying”

On August 12, Discovered Materials opened Material Discovery Bench, a benchmark for AI models that search for materials used in multilayer chips. Seven models proposed 526 candidates, but only one recipe was rated “worth trying.” The team is now attempting to synthesize that material.

In these chips, memory is placed above the logic circuits to shorten the distance that signals must travel. The layers require a dielectric material that electrically insulates the circuits while conducting heat away from them. Otherwise, heat builds up inside the assembly and limits the number of layers that can be stacked.

In the benchmark, an AI model searches the scientific literature, uses materials science software, and proposes a crystal structure, which specifies how atoms are arranged within a material. Computational models then select candidates predicted to conduct heat well enough, electrically insulate the circuits, withstand mechanical stress, and retain their structure. The result is a computational candidate for an interlayer material.

The model must then write a recipe for producing a thin film, which is a layer of material deposited onto a wafer. The recipe must specify the starting materials, the deposition method, and the conditions needed to obtain the required atomic arrangement. Thin-film deposition specialists evaluated sample recipes and developed a set of rules. During the benchmark, a separate language model applies those rules.

“A candidate does not count if its recipe is judged not worth attempting.”

Of the 526 published computational candidates, only one was rated “worth trying.” The team is attempting to synthesize that material. The useful result in the table is a recipe that gives an experimentalist a practical starting point.

This check is necessary because an agent can earn points by repeating an old result. In one early run, Claude Fable submitted the same material 58 times, increasing the size of its repeated crystal unit cell with each submission. A unit cell is the smallest repeating fragment that forms a crystal. The earlier novelty check treated these enlarged versions as different materials. The benchmark now requires both predicted properties and a recipe suitable for a laboratory attempt.

In July, CuspAI launched a network of more than 45 partners for materials discovery. A new formula must still be developed into a laboratory sample and then adapted to a manufacturer’s equipment. Material Discovery Bench evaluates the part of that process in which the model must give the laboratory a practical starting point: a synthesis recipe.