Perovskite is a group of materials that are currently the main contender to replace silicon-based photovoltaic cells. They hold the promise of thinner, lighter panels, can be manufactured to very high productivity at room temperature instead of hundreds of degrees, and are cheaper and easier to transport and install. But bringing these materials from controlled laboratory experiments into a product that can be manufactured competitively has been a long struggle.
The fabrication of perovskite-based solar cells involves optimizing at least a dozen or so variables simultaneously, even within a given fabrication approach among many possibilities. But a new system based on a novel approach to machine learning could speed the development of improved production methods and help make the next generation of solar energy a reality.
The system, developed by researchers at MIT and Stanford University over the past few years, makes it possible to integrate data from past experiences, and information based on personal observations by experienced workers, into the machine learning process. This makes the results more accurate and has already led to the fabrication of perovskite cells with an energy conversion efficiency of 18.5 percent, a level competitive for today’s market.
The research is published today in the journal joulesIn a paper written by MIT mechanical engineering professor Tonio Bonassisi, Stanford professor of materials science and engineering Reinhold Dawskardt, recent MIT research assistant Zhi Liu, Stanford doctoral alumnus Nicholas Rollston, and three others.
Perovskites are a group of layered crystalline compounds defined by the formation of atoms in their crystal lattice. There are thousands of these possible compounds and many different ways to make them. While most lab-scale development of perovskite materials use spin-coating technology, this is not feasible for large-scale manufacturing, so companies and laboratories around the world are looking for ways to translate these lab materials into a workable, manufacturable product.
“There is always a big challenge when you try to take a process on a lab scale and then turn it into something like a startup or a manufacturing line,” says Rollston, who is now an assistant professor at Arizona State University. The team looked at the process they felt had the greatest potential, a method called Rapid Spray Plasma Therapy, or RSPP.
The manufacturing process will involve a moving surface from roll to roll, or a series of plates, on which initial solutions of the perovskite compound are sprayed or sprayed with ink as the plate is rolled. The material will then proceed to the curing phase, providing rapid and continuous output “with higher throughput than any other PV technology,” says Rollston.
“The real breakthrough with this platform is that it will allow us to scale in a way that no other material has allowed us to do,” he adds. “Even materials like silicone require a longer time frame due to the processing that is being done. You can think of [this approach as more] Like spray paint.”
Within this process, at least a dozen variables may influence the outcome, and some are more controllable than others. This includes the composition of the starting material, temperature, humidity, speed of the curing path, distance of the nozzle used to spray the material onto the substrate, and methods for curing the material. Many of these factors can interact with each other, and if the process is outdoors, humidity, for example, may be out of control. It is impossible to evaluate all possible combinations of these variables through experiment, so machine learning was necessary to help guide the experimental process.
But while most machine learning systems use raw data such as measurements of electrical and other properties of test samples, they typically do not incorporate human experience such as qualitative observations made by experimenters of visual and other properties of test samples, or information from other experiments reported by other researchers. Therefore, the team found a way to incorporate such external information into a machine learning model, using a probability factor based on a mathematical technique called Bayesian Optimization.
Using the system, he says, “having a model that comes from empirical data, we can detect trends we haven’t been able to see before.” For example, they initially had trouble adapting to the uncontrolled variations in humidity in their surroundings. But the model showed them “that we can overcome our humidity challenges by changing the temperature, for example, and changing some of the other knobs.”
The system now allows experimenters to direct their process more quickly in order to optimize it for a specific set of desired conditions or outcomes. The team focused their experiments on improving energy production, but the system could also be used to simultaneously incorporate other criteria, such as cost and durability — something team members continue to work on, Bonassisi says.
The researchers were encouraged by the Department of Energy, which sponsored the work, to commercialize the technology, and they are currently focused on transferring the technology to the existing perovskite manufacturers. “We are reaching out to companies now,” Bonassisi says, and the code they developed is freely available through an open source server. “It’s now on GitHub, anyone can download it, and anyone can play it,” he says. “We are excited to help companies get started with our code.”
Already, several companies are preparing to produce perovskite-based solar panels, although they are still working out the details of how to produce them, says Liu, who is currently at Northwestern Polytechnic University in Xi’an, China. He says companies there have yet to manufacture on a large scale, but instead are starting with smaller, higher-value applications such as building integrated solar tiles where appearance matters. Three of these companies are “on track or being pushed by investors to manufacture 1m by 2m rectangular units [comparable to today’s most common solar panels]Within two years.”
“The problem is that they don’t have a consensus on which manufacturing technology to use,” says Liu. He says the RSPP method, which was developed at Stanford University, “still has a good chance” of being competitive. The machine-learning system the team developed could prove important in guiding the optimization of any process that is ultimately used.
“The primary goal was to speed up the process, so it took less time, fewer experiments and fewer human hours to develop something that could be used immediately, for free, for industry,” he says.
Ted Sargent, an undergraduate professor at the University of Toronto, who was not associated with this work, who says it demonstrates a “workflow that readily adapts to the deposition techniques that dominate the thin-film industry. Only a few groups have the concurrent expertise in engineering and computation to drive such developments”. Sargent adds that this approach “could be an exciting advance for fabricating a broader range of materials” including LEDs, other photovoltaic technologies, and graphene, “in short, any industry that uses some form of vapor or vacuum deposition.”
The team also included Austin Flick and Thomas Colborne at Stanford and Zikon Rehn in the Singapore-MIT Alliance for Science and Technology (SMART). In addition to the Department of Energy, work was supported by a fellowship from the MIT Energy Initiative, the National Science Foundation Graduate Research Fellowship Program, and the SMART Program.