Taking the Guess out of Guesswork: Using Machine Learning to Optimize Solar Cells

Material scientist Julia Hsu was listening to a lecture when it hit her. The methodology being described by the presenter was exactly what she needed in her lab. It could help her greatly accelerate advances in her research program, for example improving the efficiency of solar cells. There was only one problem. The method used machine learning — which Hsu didn’t know the first thing about.

“It took me a while to sort through the different types of machine learning to even figure out what might work for my project,” says Hsu, a professor of materials science and engineering in the Erik Jonsson School of Engineering and Computer Science at the University of Texas at Dallas (UTD).

Despite her inexperience with the topic, Hsu wasn’t deterred. She narrowed in on an approach and was ready to tackle the immense job of learning an entirely new discipline, even though she knew it would be a long and challenging process.

But then she learned about the Simons Foundation’s new Pivot Fellowship. The fellowship, launched in 2022, was developed with curious scientists in mind — ones who wanted to learn a new field that was separate from their usual line of study. Open to researchers in natural sciences, mathematics, engineering, data science and computer science, the fellowship gives grantees salary support as well as research, travel and professional development funding during the fellowship year as well as a mentor who is committed to guiding them through a transition to the new field.

“We wanted to create this fellowship to break down some of the barriers that keep researchers from being siloed in one discipline for their whole career,” says Alyssa Picchini Schaffer, vice president and senior scientist for the Simons Foundation’s neuroscience collaborations. Picchini Schaffer spearheaded the development and launch of the Pivot Fellowship program along with Simons Foundation president David Spergel. “So many researchers are highly curious and have a lot to offer a new field, such as different ways of thinking and problem solving as well as unique skillsets.”

By providing each Pivot Fellow a mentor, the program facilitates a fast dive deep into the fellow’s newly chosen subject. The Pivot Fellowship additionally helps foster community, building a support system that connects the fellows with one another during the intense year-long program.

Hsu applied and became one of seven researchers in the 2023 inaugural cohort of Pivot Fellows. In January 2023 she began he work under the mentorship of Tonio Buonassisi, a professor of mechanical engineering at the Massachusetts Institute of Technology. Buonassisi has years of expertise in scientific machine learning, including spearheading the use of artificial intelligence to develop new materials such as solar photovoltaics.

Hsu, who has tackled many problems in material sciences over her career, has recently been focusing on improving solar cells — particularly a type called perovskite solar cells. These cells, unlike the more common silicon-based variety, are made of metal halide perovskites, which can be made at low cost. However, new and efficient methods need to be developed to produce perovskite cells at scale for commercial use.

Solar cells work by absorbing sunlight, which excites electrons from the semiconductor material to produce electricity. By layering the semiconductor with other materials, the electron can be captured, forming an electrical current — or electricity. Perovskite solar cells are made of many layers between a bottom electrode and a transparent conductive electrode on the top that sandwich a perovskite slab and other layers that capture and transport electrons. However, to activate the perovskite and transporting layers, the materials need to be cured to create the right crystalline structure for the job. Typically, this curing requires baking the material with immense heat — some above 250 Celsius — for ten minutes or more, which inflates production costs and raises the product’s carbon footprint. And ultimately to do this commercially would require improbably large ovens.

“Using heat is slow and inefficient, and moreover it is carbon-intensive,” says Hsu. “In 2020, industrial heat accounts for one-third of U.S.’s energy use and 9 percent of carbon emissions. If we’re trying to create a renewable energy source, it doesn’t make sense for it to require a high-carbon manufacturing process.”

To streamline production of these solar cells, Hsu is looking into new ways to make them. Recently, Hsu and her lab at UTD have focused on a method called photonic curing. Instead of sending a cell into an oven to bake, like a pizza, photonic curing uses short but intense blasts of light that impart enough radiant energy to cure the material without heating the entire sample and environment to high temperatures. Exposing the material to a light pulse for just milliseconds in this method can achieve the same crystallization of the material as traditional heat treatments but wastes much less energy. It has the added benefit of only affecting the targeted layer, which ensures no other layers are altered in the process. This opens up the use of new materials that could be more efficient than traditional materials but would melt or burn in the conventional heating process.

In 2020, Hsu and her lab published several papers demonstrating the use of photonic curing to make perovskite solar cells that can be viable for the industrial-scale manufacturing of perovskite solar cells as well as flexible electronics. Since then, she’s switched focus to optimizing photonic curing for flexible transparent conducting electrode, which has applications beyond solar cells in LEDs, touch screens and displays. However, work has proven to be very difficult.

The concept of photonic curing is straightforward — you shine a light on the material to be altered. But it comes with a whole host of variables that can be adjusted, like the light intensity, how long the light pulse should be, and number of light pulses, to name a few. To make matters more complicated, the variables can depend on one another, which makes finding the perfect recipe incredibly difficult.

Imagine you wanted to perfect a recipe for chocolate chip cookies. You might first make some batches with different levels of sugar, before settling on the right amount. Next, you move on to testing for the right amount of butter. But then you realize the amount of butter in a perfect cookie also depends on the number of eggs and the temperature of the oven. All of a sudden, you need to make dozens of batches of cookies to narrow down the perfect combination. This is the issue Hsu faces with her solar cell fabrication using photonic curing, but in her case, making a batch of solar cells is much more difficult, time-consuming and cost-intensive than whipping up a dozen cookies.

After testing a round or two, Hsu typically has an idea of what might be a good combination to try next. This is because she has intuition built upon years of experience. That’s not always the case though. Hsu’s lab is filled with graduate students and postdoctoral researchers working on projects that can vary widely from one year to the next.

“I had one graduate student working on the photonic curing project and he became really good at picking which combination of parameters to try next,” Hsu says. “But when new students come in, it’s really hard to transfer that knowledge.”

An important way in which Hsu has hoped to streamline this process is with machine learning. If she can create a methodology that can pick the best option, she won’t have to trial-and-error her way to perfection. Many machine learning programs — including ones like ChatGPT and DALL-E 2 — are trained by feeding the code tens of thousands of images or examples. By making connections between the samples it’s given, a machine learning code eventually learns what a puppy looks like or how to write a joke. But Hsu doesn’t have the time or money to create tens of thousands of solar cells, so this method doesn’t work.

Instead, Hsu began studying a machine learning strategy called Bayesian optimization. Bayesian optimization, which started to take form in the 1970s, is a method to simultaneously tune multiple parameters to give the best result. It’s widely used in problems that would otherwise have high computational costs and has been applied in experimental physics, chemistry, drug development, robotics and more.

On that fateful day in 2021, as Hsu listened to a talk at a conference, it was Bayesian optimization that she heard about. Nick Rolston, then a postdoc at Stanford, had presented on findings using Bayesian optimization to improve the manufacturing of perovskite solar cells using an open-air processing technique. The research showed a marked increase in production efficiency much faster than would have been possible with traditional experiments. If machine learning could be used for that processing method, Hsu thought, why not use it for photonic curing?

Hsu started her learning journey by first watching YouTube videos, and then playing with other people’s codes. Now, under Buonassisi’s guidance as a Pivot Fellow, Hsu and her lab have begun writing up a Bayesian optimization machine learning code to identify the best settings for photonic curing. Along the way, they’ve been learning machine learning lingo and potential pitfalls.

“Often it seems like machine learning is described as a panacea that can solve all your problems,” Hsu says. “But now that I’m actually doing it, I’m finding all of the intricacies, the subtleties and the difficulties you have to overcome to do it right.”

The mentorship aspect, Hsu says, is key to keeping her on-track. With Buonassisi’s input, Hsu is able to keep clear of strategies and tricks that don’t work or are outdated, which saves valuable time and effort.

“I really like having a mentor,” Hsu says, “It would be much harder to ‘pivot’ without one.”

“We benefit a lot by collaborating with Julia. She and her team bring unique domain expertise in materials processing, which has already benefitted our other projects in the lab,” Buonassisi says. “As somebody new to machine learning, she asks really insightful questions that force us to think deeply, and sometimes, shift our assumptions. For example, I was not aware how much some commercial machine learning platforms have advanced relative to custom-built codes. We might be changing our own platform choice for new learners accordingly.”

The idea of mentorship was part of the fellowship from the start, says Picchini Schaffer. She wanted the researchers to be able to go deeper than just looking at surface level questions.

“The important part was not to only enable researchers to switch fields but provide them the mentoring and training to really dig deep and learn what’s needed to truly make a meaningful contribution to the field,” Picchini Schaffer says.

Hsu’s goal is to have a working code implemented by November and test it on experimental data before her fellowship draws to a close in December. It’s not a lot of time, but enough, she believes, even given she’s continuing her full-time duties as a professor. At the end of the year, Hsu and the other Pivot Fellows will be invited to continue their work with a $1.5 million, 3-year research award.

“So far it’s been fun but very hard,” Hsu says. “I’m very appreciative of having this opportunity because now that I’ve gotten into it, I’ve realized that if I did not have this fellowship, there’s no way I could tackle this project and master machine learning on my own.”