A new method of identifying the most valuable formulations may save this promising new lightweight photovoltaics.
Perovskites, also known as perovskites, are widely promoted as a possible replacement for silicon in solar cells. However, their biggest problem is the tendency to degrade rapidly. Although the lifetime of perovskite-based solar cells has improved from minutes to months to decades, it still needs to catch up to the many decades expected from silicon.
An international interdisciplinary team from MIT has developed a new method to narrow the search for the most promising candidates for long-lasting perovskite formulations. Their system already found one composition that has outperformed all other versions more than ten times in the lab. This perovskite performs three times better in real-world conditions than in different formulations.
These findings are published in the journal matter in a paper written by Shijing Sun (MIT postdoc), John Fisher (MIT professor), Tonio Buonassisi (MIT principal investigator at the Singapore MIT Alliance for Research and Technology), and 16 other researchers from MIT, New York, Germany, Singapore, and Colorado.
Perovskites are a broad category of materials distinguished by the atoms’ arrangement in their crystal lattice. Each layer, commonly called A, B, and X by convention can contain various particles or compounds. It isn’t easy to search through all the possible combinations to find suitable candidates for your goals, such as efficiency, longevity, or availability of raw materials. There needs to be a map to help you.
Buonassisi states that even if you only consider three elements, the most prevalent ones in perovskites people sub in or out are at the A site. Buonassisi’s 1-percent increments can easily modify each piece in its relative composition. It becomes absurd to count the steps. It becomes enormous, and it is impossible to search through them all. Each step involves the complicated synthesis of a new material and testing its degradation. This can be time-consuming, even under accelerated aging conditions.
Their data fusion approach is the key to their success. The iterative process uses an automated system for the production and testing of a variety of formulations. Next, machine learning is used to analyze the results and combine them with the first principle, physical modeling, to guide the next round. The system continues to refine the results by repeating this process.
Buonassisi likes comparing the vast array of possible compositions to an ocean. He says that most researchers have stuck to known formulas with high efficiency. For example, he has achieved high efficiencies by tweaking only slightly with those atomic configurations. Sometimes, however, a mistake is made, or a genius stroke occurs, and someone finds a better composition. It’s a happy accident, and everyone moves over there.” It’s only sometimes a structured thought process.
He says this new approach allows far-off areas to be explored to find better properties. It is also more efficient and systematic. The researchers have identified the most durable perovskite-based solar cell material by combining and testing less than 2% of the possible combinations.
Sun, the coordinator of the international team responsible for the work, said that the story is about the “fusion of all the different tools” used to create the new formula. Sun also developed a high-throughput automated degrading test system that monitors the material’s degradation through changes in its color as it darkens. The team went beyond creating a small chip in the laboratory and integrated the material into a functioning solar cell to confirm its results.
She adds, “Another benefit of this work is that it shows, all the way through the chemical selection up to the actual making of a solar cell in its end.” “It tells us that the machine-learning-suggested chemical is not only stable in its freestanding form. They can also be converted into the real-life solar cell, leading to increased reliability.” Their lab-scale demonstrations reached longevity up to 17 times higher than the baseline formula, but she said even the full-cell rally outlasted the existing materials by over three times.
Buonassisi believes that the same method developed by the team could be used in other areas of materials research with similar composition ranges. It opens up a new mode of study that allows for short loops of innovation at subcomponent and material levels. Once you have found the suitable composition, you can move it into a more extended circle, which involves device fabrication, and then you test it.” at the next level.
He says, “It’s one the biggest promises of the field that you can do this kind of work.” It was one of those moments that will be cherished forever. I can still recall exactly where I was when Shijing called me about the results. It’s when you begin to see the ideas come to life. It was truly unique.”
“What’s particularly interesting about this advance is that the authors use Physics to guide the intuition of [optimization] rather than restricting the search space by hard constraints,” said University Professor Edward Sargent of the University of Toronto. He is a specialist in nanotechnology and was not involved with the research. This approach will be widely used as machine learning advances toward real problems in materials science.
