Engineers use artificial intelligence to expand production scale of advanced solar cells
June 06 13:46:07, 2025
A group of researchers from MIT and Stanford University has developed a novel machine learning-based system that promises to revolutionize the production of perovskite solar cells. Perovskite, a material with great potential to replace traditional silicon-based solar cells, is complex to manufacture due to its sensitivity to numerous variables. This new system combines historical experimental data and personal insights from experienced workers, improving accuracy and leading to perovskite solar cells with an impressive energy conversion efficiency of 18.5%, which is a leading figure in the current market.
This research, published in the journal Joule, was led by Tonio Buonassisi, a professor of mechanical engineering at MIT, along with Reinhold Dauskardt from Stanford University and others. Perovskite is a versatile crystalline compound, with thousands of potential variations and multiple manufacturing methods. Although spin coating is commonly used in labs, it is impractical for large-scale manufacturing. Hence, industries worldwide are exploring ways to adapt lab-developed materials for industrial production.
The team focused on a promising method known as Rapid Spray Plasma Treatment (RSPP), which could potentially be scaled up for mass production. This process involves spraying or inkjet printing the perovskite precursor solution onto moving sheets or roll-to-roll surfaces, followed by a curing stage to ensure fast and continuous output. According to Rolston, now an assistant professor at Arizona State University, this method offers higher throughput compared to other photovoltaic technologies.
However, optimizing this process involves managing at least a dozen variables, some of which are harder to control than others. Factors such as the composition of the starting material, temperature, humidity, nozzle distance, and curing methods all play a role, and their interactions can be intricate. Since testing all possible combinations experimentally is nearly impossible, the team incorporated human experience and external data into the machine learning models using Bayesian optimization techniques.
This integration allowed the researchers to identify trends that were previously unseen. For instance, they initially struggled with uncontrolled humidity but later realized that adjusting the temperature and other parameters could counteract its effects. The system now enables faster experimental guidance, allowing researchers to optimize for specific conditions or outcomes. Currently, the team is focusing on power output optimization, but the system can also be adapted for cost and durability considerations.
The U.S. Department of Energy, which supported the research, encourages commercialization of the technology. Buonassisi mentioned that they are in discussions with companies and making their code freely available on GitHub. Liu, based in Xi'an at Northwestern Polytechnical University, noted that several companies are preparing to produce perovskite-based solar panels, though they are still refining their production techniques. These companies are beginning with high-value applications like building-integrated solar tiles, where appearance is crucial. Some are planning to produce 1-meter by 2-meter rectangular modules within two years.
Despite the progress, there is no consensus on the best manufacturing technology, with RSPP still holding strong potential. Ted Sargent, a professor at the University of Toronto, praised the work, highlighting its significance for other industries like LED and graphene production.
In addition to Buonassisi, Dauskardt, and Rolston, the team included Austin Flick, Thomas Colburn from Stanford, and Zekun Ren from SMART. Funding came from the MIT Energy Program, the National Science Foundation Graduate Research Fellowship Program, and the SMART Program, in addition to the Department of Energy.