
AI finds better perovskite materials for solar cells

Researchers in Germany have used AI to find new perovskite materials for more efficient solar cells.
The team at KIT’s Institute of Nanotechnology and the Helmholtz Institute Erlangen-Nürnberg (HI ERN) found a material with a near-record conversion efficiency of 26.2%.
“With only 150 targeted experiments, we were able to achieve a breakthrough that would otherwise have required hundreds of thousands of tests. The workflow we have developed will open up new ways to quickly and economically discover high-performance materials for a wide range of applications,” said Professor Christoph Brabec at HI ERN.
- imec shows long term stability of outdoor perovskite solar modules
- Startup improves perovskite solar cells with aerosol process
The starting point at HI ERN was a database with structural formulae for approximately one million virtual molecules that could be synthesized from commercially available substances. From these virtual molecules, 13,000 were selected at random. The KIT researchers used established quantum mechanical methods to determine their energy levels, polarity, geometry and other properties.
From the 13,000 molecules, the scientists chose 101 with the greatest differences in their properties, synthesized them with robotic systems at HI ERN, used them to produce otherwise identical solar cells, and then measured the efficiency of the solar cells. “Being able to produce truly comparable samples thanks to our highly automated synthesis platform, and thus being able to determine reliable efficiency values, was crucial to our strategy’s success,” said Brabec, who headed the work at HI ERN.
With one of the discovered materials, they increased the efficiency of a reference solar cell by approximately two percentage points to 26.2 percent. “Our success shows that enormous amounts of time and resources can be saved by applying skillful strategies for the discovery of new energy materials,” said Pascal Friederich at KIT.
He used the achieved efficiencies and the properties of the associated molecules to train an AI model, which suggested 48 other molecules to synthesize. Its suggestions were based on two criteria: high expected efficiency and unforeseeable properties.
- Dual AI boost for solid state battery materials
- $100m for AI-enabled sustainable semiconductor materials
“When the machine learning model is uncertain about the predicted efficiency, it’s worthwhile to synthesize the molecule and take a closer look at it,” Friederich said, explaining the second criterion. “It might surprise us with a high efficiency level.”
Using the molecules suggested by the AI, it was possible to build perovskite solar cells with above-average efficiency, including some exceeding the capabilities of the most advanced materials currently used. “We can’t be sure we’ve really found the best one of a million molecules, but we’re certainly close to the optimum,” said Friederich.
Since the researchers used an AI that indicates which of the virtual molecules’ properties its suggestions were based on, they were able to gain some insight into the molecules it suggested. For example, they determined that the AI-suggestions are based in part on the presence of certain chemical groups, such as amines, that chemists had previously neglected.
