KIT deploying AI to develop better photovoltaic materials faster

19 Dec 2024

Machine learning speeds search for new semiconductors for perovskite solar cells.

Perovskite solar cells are considered a flexible and sustainable alternative to conventional silicon-based solar cells. Researchers at the Karlsruhe Institute of Technology, Germany, (KIT) are part of an international team that has found new organic molecules within a few weeks that can be used to increase the efficiency of perovskite solar cells.

The team combined the use of AI with fully automated high-throughput synthesis. The strategy developed can be transferred to other areas of materials research, such as the search for new battery materials. Their achievement is described in Science.

If you want to find out from 1,000,000 molecules which make perovskite solar cells particularly efficient as conductors of positive charge, you have to produce and test these million molecules – or proceed as researchers led by Tenure Track Professor Pascal Friederich from the KIT Institute of Nanotechnology and Professor Christoph Brabec from HI ERN have done.

“With just 150 targeted experiments, a breakthrough was achieved that would otherwise have required hundreds of thousands of tests. The workflow developed opens up new possibilities for the rapid and cost-efficient discovery of high-performance materials in a wide range of application areas,” said Brabec, who led the work at HI ERN.

Using one of the materials discovered in this way, they increased the efficiency of a reference solar cell by around two percent to 26.2 percent. “This success shows that with a clever strategy, enormous time and resources can be saved when developing new energy materials,” said Friederich.

The starting point at HI ERN was a database with the structural formulas of around one million virtual molecules that could be produced from commercially available substances. The researchers at KIT used established quantum mechanical methods to calculate energy levels, polarity, geometry, and other characteristics of 13,000 of these virtual molecules, which were randomly selected.

AI training with data from 101 molecules

From these 13,000 molecules, the researchers then selected 101 molecules that differed as much as possible in their characteristics. These were produced automatically at HI ERN using a robot system, thereby producing otherwise identical solar cells. They then measured their efficiency. “The key to the success of our strategy was that, thanks to our highly automated synthesis platform, we were able to produce truly comparable samples and thus determine reliable values for the efficiency,” said Brabec.

The KIT researchers trained an AI model using the achieved efficiencies and the characteristics of the associated molecules. The model then suggested another 48 molecules for synthesis based on two criteria: an expected high efficiency and unpredictable properties. “If the machine learning model is unsure about predicting the efficiency, it is worth producing the molecule to examine it more closely,” said Pascal Friederich, tenure-track professor for artificial intelligence in materials science, explaining the second criterion. “It could surprise with a high efficiency.”

In fact, the molecules proposed by the AI could be used to build above-average efficient solar cells, including ones that outperform other state-of-the-art materials. "We cannot be sure that we have really found the best among a million molecules, but we are certainly close to the optimum," said Friederich.

AI versus chemical intuition

The researchers can understand the AI’s molecule suggestions to a certain extent because the AI used indicates which features of the virtual molecules were decisive for its suggestions. It turned out that the AI’s suggestions were partly based on features, such as the presence of certain chemical groups such as amines, to which chemists had previously paid less attention.

Brabec and Friederich are convinced that their strategy is promising for materials research in other application areas or can be extended to the optimization of entire components. The research results, which were developed in collaboration with researchers from the University of Erlangen-Nuremberg, the Ulsan National Institute of Science in South Korea, the Xiamen University in China and the University of Electronic Science and Technology in Chengdu, China, were recently published in the Science paper.