Perovskite solar cells offer a flexible and eco-friendly option compared to traditional silicon solar cells. Researchers at the Karlsruhe Institute of Technology (KIT) are part of an international coalition that successfully identified new organic molecules within just a few weeks, which enhance the efficiency of perovskite solar cells. The team employed a smart blend of artificial intelligence (AI) and automated high-throughput synthesis. This innovative approach could also be utilized in different areas of materials science, like the development of new battery technologies.
To determine which out of a million potential molecules could effectively conduct positive charges and optimize the efficiency of perovskite solar cells, one might typically need to synthesize and evaluate all of them. However, the researchers, under the guidance of Tenure-track Professor Pascal Friederich from KIT’s Institute of Nanotechnology, and Professor Christoph Brabec from the Helmholtz Institute Erlangen-Nürnberg (HI ERN), took a more efficient route. “With just 150 targeted experiments, we achieved a breakthrough that could have otherwise demanded hundreds of thousands of trials. Our developed workflow will pave the way for rapid and cost-effective discovery of high-performance materials across various fields,” Brabec stated. One of the materials they identified improved the efficiency of a reference solar cell by roughly two percentage points, reaching 26.2 percent. “Our success indicates significant time and resource savings can be realized through strategic approaches in discovering new energy materials,” Friedrich commented.
The endeavor at HI ERN commenced with a database containing structural formulas for about one million virtual molecules that could be derived from commercially available chemicals. From this pool, 13,000 molecules were randomly selected. Researchers at KIT utilized established quantum mechanical methods to analyze their energy levels, polarity, geometry, and more.
Training AI with Data from Just 101 Molecules
Out of the 13,000 molecules, the scientists picked 101 that exhibited the greatest variations in their characteristics, synthesized them using robotic systems at HI ERN, then created otherwise identical solar cells to evaluate their efficiency. “Our success hinged on our ability to produce genuinely comparable samples via our highly automated synthesis platform, enabling reliable efficiency assessment,” Brabec explained, who oversaw the work at HI ERN.
The KIT researchers leveraged the efficiencies achieved and the associated molecular properties to train an AI model, which proposed 48 additional molecules for synthesis. The recommendations were founded on two main aspects: anticipated high efficiency and unpredictable characteristics. “If our machine learning model expresses uncertainty about the efficiency prediction, synthesizing that molecule for closer inspection is advantageous,” Friedrich clarified regarding the second criterion. “It might yield unexpectedly high efficiency.”
Indeed, utilizing the AI-suggested molecules allowed for the creation of solar cells with superior efficiency, with some surpassing the capabilities of the cutting-edge materials currently in use. “While we cannot definitively claim to have identified the absolute best of one million molecules, we are likely very close to the ideal,” Friedrich remarked.
AI Versus Chemical Intuition
The researchers benefited from the AI’s transparency regarding which specific properties of the virtual molecules informed its suggestions, allowing them to glean insights into the suggested molecules. Notably, they discovered that the AI recommendations were partly influenced by the presence of specific chemical groups, like amines, which had often been overlooked by chemists.
Brabec and Friedrich are optimistic that their approach could be applicable in other materials science contexts or could even extend to the optimization of entire components.
The results of this research, which involved collaboration with scientists from FAU Erlangen-Nürnberg, South Korea’s Ulsan National Institute of Science, and China’s Xiamen University and University of Electronic Science and Technology, were recently published in the journal Science.
