With the support of AI methods, researchers want to improve the manufacturing processes for highly efficient perovskite solar cells (Photo: Amadeus Bramsiepe, KIT)
This article was first published onwww.kit.edu
Tandem solar cells based on perovskite semiconductors can convert sunlight into electricity more efficiently than conventional silicon solar cells. In order to bring this technology to market, stability and production processes must be further improved. Researchers from the Karlsruhe Institute of Technology (KIT) as well as the Helmholtz Imaging platforms at the German Cancer Research Center (DKFZ) and Helmholtz AI have now found a way to predict the quality of the perovskite layers and thus the solar cells: using machine learning and new artificial methods Intelligence (AI) allows this to be recognized during production from variations in light emission. The team reports on the results from which improved production processes can be derived in Advanced Materials (DOI: 10.1002/adma.202307160).
Perovskite tandem solar cells combine a perovskite with a conventional solar cell, for example based on silicon. They are considered next-generation technology because, with an efficiency of currently more than 33 percent, they are much more efficient than conventional silicon solar cells - with inexpensive starting materials and simple manufacturing methods. The prerequisite for achieving this level of efficiency is a very high-quality and extremely thin perovskite layer, which is only a fraction of the thickness of a human hair. “One of the biggest challenges is to produce these high-quality so-called multicrystalline thin films using cost-effective and scalable processes without defects and holes,” explains tenure-track professor Ulrich W. Paetzold from the Institute of Microstructure Technology and the Light Engineering Institute of KIT. Even under apparently perfect conditions in the laboratory, unknown influences lead to fluctuations in the quality of the semiconductor layers: “This ultimately prevents the rapid start of industrial production of these highly efficient solar cells, which we so urgently need for the energy transition,” says Paetzold.
In order to find out which factors influence the coating, an interdisciplinary team of perovskite solar cell experts from KIT teamed up with machine learning and explainable artificial intelligence (XAI) specialists from Helmholtz Imaging (at the DKFZ in Heidelberg) and Helmholtz AI ( at KIT). The researchers have developed AI methods that train and analyze so-called neural networks using a large data set. The data set includes video recordings of the photoluminescence of the perovskite thin films during the manufacturing process. Photoluminescence refers to the radiant emission of the semiconductor layers after excitation by an external light source. “Since even experts couldn’t see anything remarkable on the thin films, the idea arose to train an AI for machine learning (deep learning) to find hidden indicators of a good or bad coating in the millions of data from the videos,” explain Lukas Klein and Sebastian Ziegler from Helmholtz Imaging at the DKFZ.
In order to filter and analyze the very broad information provided by deep learning AI, the researchers subsequently used methods of explainable artificial intelligence.
The result: In the experiment, the researchers were able to see that the photoluminescence varies during production and this influences the coating quality. “What was crucial in our work was that we specifically used XAI methods to see which factors would have to change for a high-quality solar cell,” said Klein and Ziegler. That is usually not the case. Most of the time, XAI is only used as a kind of guardrail to avoid errors when building AI models: “This is a paradigm shift, and the fact that we can systematically gain highly relevant findings in materials science in this way is new.” Because the answer is based on the variation of the Photoluminescence enabled the researchers to go further. After appropriate training of the neural networks, the AI was able to predict whether the solar cell would achieve low or high efficiency, depending on when and what variation in light emission occurred during production. “These are extremely exciting results,” says Ulrich W. Paetzold. “Thanks to the combined use of AI, we have an idea of which adjustments we need to make first in order to improve production. We can carry out our experiments in a more targeted manner and no longer have to search for a needle in a haystack in the dark. This is a blueprint for follow-up research, including for many other aspects in energy research and materials science.”
Lukas Klein, Sebastian Ziegler, Felix Laufer, Charlotte Debus, Markus Götz, Klaus Maier-Hein, Ulrich W. Paetzold, Fabian Isensee, Paul F. Jäger: Discovering Process Dynamics for Scalable Perovskite Solar Cell Manufacturing with Explainable AI. Advanced Materials, 2023. DOI: 10.1002/adma.202307160