In PV cells, perovskite materials would be better than silicon, but scaling up production is a significant challenge. Machine learning can be of assistance.
Perovskites are a group of materials that are vying to replace silicon-based solar photovoltaics, which are now in widespread use. They offer panels that are significantly lighter and thinner, that can be manufactured in vast quantities with ultra-high throughput at room temperature rather than hundreds of degrees, and that are easier and less expensive to transport and install. However, it has taken a long time to turn these materials from minor laboratory studies into a product that can be mass-produced cheaply.
Even within one manufacturing process among numerous options, producing perovskite-based solar cells requires optimizing at least a dozen or so factors at simultaneously. A new system based on a revolutionary approach to machine learning, on the other hand, might hasten the development of optimum manufacturing procedures and aid in the realization of the next generation of solar power.
The technique, which has been created over the last few years by academics at MIT and Stanford University, allows machine learning to include data from previous trials as well as information based on personal observations by experienced personnel. This improves the accuracy of the results and has already resulted in the production of perovskite cells with an energy conversion efficiency of 18.5%, which is competitive in today's market.
Tonio Buonassisi, MIT professor of mechanical engineering, Stanford professor of materials science and engineering Reinhold Dauskardt, recent MIT research assistant Zhe Liu, Stanford doctorate graduate Nicholas Rolston, and three other researchers recently published a study in the journal Joule.
Perovskites are a class of layered crystalline compounds determined by the atoms' crystal lattice structure. There are hundreds of such compounds that may be made in a variety of methods. While spin-coating is used in most lab-scale perovskite material research, it is not viable for larger-scale production, thus corporations and labs throughout the world have been looking for ways to translate these lab materials into a practical, manufacturable product.
“There’s always a big challenge when you’re trying to take a lab-scale process and then transfer it to something like a startup or a manufacturing line,” says Rolston, who is now an assistant professor at Arizona State University. The team looked at a procedure called rapid spray plasma processing, or RSPP, that they thought had the most potential.
The precursor solutions for the perovskite compound would be sprayed or ink-jetted on a moving roll-to-roll surface, or sequence of sheets, as the sheet rolled past. The material would then go through a curing cycle, generating a constant and quick output “with throughputs that are higher than for any other photovoltaic technology,” according to Rolston.
“The real breakthrough with this platform is that it would allow us to scale in a way that no other material has allowed us to do,” he continues. “Even materials like silicon require a much longer timeframe because of the processing that’s done. Whereas you can think of [this approach as more] like spray painting.”
At least a dozen variables may influence the result of the procedure, some of which are more controlled than others. These factors include the starting ingredients' composition, temperature, humidity, processing route speed, nozzle distance utilized to spray the material onto a substrate, and curing techniques. Many of these variables might interact, and if the operation is conducted outside, humidity, for example, may be uncontrollable. Because it is difficult to evaluate all conceivable combinations of these factors through testing, machine learning was used to help guide the experimental procedure.
However, while most machine-learning systems use raw data such as electrical and other properties of test samples, they rarely include human experience such as qualitative observations of the visual and other properties of the test samples made by the experimenters, or information from other experiments reported by other researchers. So, using a probability factor based on a mathematical approach called Bayesian Optimization, the researchers devised a way to include such outside information into the machine learning model.
“Having a model that comes from experimental data, we can find out trends that we weren’t able to see before,” he adds. For example, they initially struggled to compensate for uncontrollable humidity fluctuations in their environment. But the model showed them “that we could overcome our humidity challenges by changing the temperature, for instance, and by changing some of the other knobs.”
The method now allows experimenters to steer their process much more quickly in order to optimize it for a specific set of variables or outcomes. The team concentrated on improving power output in their trials, but the system could also be used to integrate other factors, such as cost and durability, which members of the team are now working on, according to Buonassisi.
The Department of Energy, which funded the research, pushed the scientists to commercialize the technology, and they're now focused on tech transfer to existing perovskite producers. “We are reaching out to companies now,”Buonassisi adds, adding that the code they created has been published open-source and is available for free. “It’s now on GitHub, anyone can download it, anyone can run it,” he explains. “We’re happy to help companies get started in using our code.”
According to Liu, who is currently at the Northwestern Polytechnical University in Xi'an, China, numerous businesses are already planning to manufacture perovskite-based solar panels, even if they are still working out the intricacies of how to do so. Companies there, he adds, aren't currently conducting large-scale production, preferring to focus on smaller, high-value applications like building-integrated solar tiles, where aesthetics are crucial. Three of these startups “are on track or are being pushed by investors to manufacture 1 meter by 2-meter rectangular modules, within two years,”according to him.
"The problem is, they don’t have a consensus on what manufacturing technology to use,” Liu explains.According to him, the Stanford-developed RSPP approach “still has a good chance” of becoming competitive. And the team's machine learning method might be useful in directing the optimization of whichever procedure is eventually used.
“The primary goal was to accelerate the process, so it required less time, less experiments, and less human hours to develop something that is usable right away, for free, for industry,” he explains.
“Existing work on machine-learning-driven perovskite PV fabrication largely focuses on spin-coating, a lab-scale technique,” says Ted Sargent, University Professor at the University of Toronto, who was not involved with this work but says it demonstrates “a workflow that is readily adapted to the deposition techniques that dominate the thin-film industry. Only a handful of groups have the simultaneous expertise in engineering and computation to drive such advances.”
According to Sargent, this method “could be an exciting advance for the manufacture of a broader family of materials” including LEDs, other PV technologies, and graphene, “in short, any industry that uses some form of vapor or vacuum deposition.”