A successful AI cameo in the lab

Scientists have identified a new compound with potential uses in photonic devices and biologically inspired computers. Or, to be more precise, they developed the AI algorithm that found it, and stands ready to do so again.

In the latest chapter in the unfolding and occasionally slightly creepy story of what artificial intelligence can do better than us, a multi-institutional team has unveiled the Closed-Loop Autonomous System for Materials Exploration and Optimisation, or CAMEO.

Like all such systems, it’s designed to save time and improve accuracy (in this case in the lab), but unlike others it needs no training or supervision, according to Aaron Gilad Kusne from the US National Institute of Standards and Technology, co-author of the team’s paper in Nature Communications.

“Many types of AI need to be trained or supervised,” he says. “Instead of asking it to learn physical laws, we encode them into the AI. You don’t need a human to train the AI.”

In materials science, researchers often look for new materials that can be used in very specific applications, which usually involves a lot of time-consuming experiments and theoretical searches.

CAMEO uses prediction and uncertainty to determine which experiment to try next, skipping the bits that would give redundant information.

It is designed to contain knowledge of key principles, past simulations and lab experiments, how the equipment works, and physical concepts. For example, the researchers armed it with the knowledge of phase mapping, which describes how the arrangement of atoms in a material changes with chemical composition and temperature.

“The key to our experiment was that we were able to unleash CAMEO on a combinatorial library where we had made a large array of materials with all different compositions,” says Ichiro Takeuchi, from the University of Maryland, US.

In that experiment, CAMEO was given 177 potential materials to investigate, covering a large range of compositional recipes. It performed 19 experimental cycles, which took 10 hours, compared with an estimated 90 hours for a scientist with the full set of materials.

The result is the discovery of the material GST467 – officially ?Ge?_4 ?Sb?_6 ?Te?_(7,) – which, the researchers suggest, is optimal for phase-change applications.

Unlike similar machine-learning approaches, they say, CAMEO discovered it by focusing on the composition-structure-property relationship of crystalline materials. In this way, it navigated the course of discovery by tracking the structural origins of a material’s functions.

“CAMEO has the intelligence of a robot scientist, and it’s built to design, run and learn from experiments in a very efficient way,” Kusne says.

The code for CAMEO is open source and the authors say it will be freely available for use by scientists and researchers.

Please login to favourite this article.