The quest for advanced solar technology has received a significant boost as researchers leverage artificial intelligence for solar material discovery. Among the most promising candidates for next-generation photovoltaic technology are perovskites, a versatile class of materials that could dramatically increase solar panel efficiency. However, the immense number of potential elemental combinations has traditionally made the identification of optimal formulations a painstakingly slow process.
In a groundbreaking development, scientists from MIT and collaborating institutions have implemented machine learning accelerating perovskite research, achieving an approximately tenfold increase in the speed of synthesizing and analyzing novel compounds. This AI-driven renewable energy materials development approach has already yielded two sets of promising perovskite-inspired materials that demonstrate exceptional potential for further investigation.
The research team's findings, detailed this week in the journal Joule, represent a significant leap forward in materials science. Led by MIT research scientist Shijing Sun and professor of mechanical engineering Tonio Buonassisi, the collaboration includes experts from MIT, Singapore, and the National Institute of Standards and Technology in Maryland.
Surprisingly, while partial automation played a role, the most substantial improvements in processing speed emerged from workflow optimizations rather than technological innovations alone. By meticulously analyzing and refining each step—from compound synthesis to substrate deposition and crystal analysis—the team identified numerous opportunities for efficiency gains that dramatically accelerated their research timeline.
"The imperative for rapid development of new energy materials has never been greater," explains Buonassisi. As global adoption of solar energy accelerates, particularly in space-constrained regions, the traditional two-decade timeline for developing new energy-conversion materials—with its substantial upfront capital investment—has become untenable. "Our objective is to compress this development cycle to under two years," he states.
The researchers' innovation centers on a parallel processing system that enables simultaneous creation and testing of diverse materials. "This platform grants us access to an expansive range of compositions using a unified synthesis approach," Buonassisi notes. "It dramatically expands our ability to explore the vast parameter space of potential materials."
Perovskite compounds comprise three constituent elements, conventionally designated as A, B, and X site ions. This structural flexibility creates an extensive family of materials with varied physical properties. While lead has traditionally dominated the B-site in high-performance perovskite photovoltaics, a major research focus involves developing lead-free alternatives that can match or exceed the efficiency of their lead-containing counterparts.
Despite theoretical predictions identifying over a thousand potentially useful perovskite formulations among millions of possible combinations, experimental verification has lagged significantly behind. This bottleneck underscores the critical need for the accelerated discovery process that neural networks for solar cell optimization can provide.
For their experimental validation, the team selected 75 diverse compositions, processing each in solution before depositing them on substrates for crystallization into thin films—the optimal form for solar cell applications. The resulting structures underwent analysis via X-ray diffraction to reveal atomic arrangements. By implementing a convolutional neural network system, the researchers reduced classification time from 3-5 hours to just 5.5 minutes while maintaining 90% accuracy.
This automated material discovery for photovoltaics approach enabled the team to explore 75 formulations in approximately one-tenth the time previously required. Among these, they identified two novel lead-free perovskite systems exhibiting promising characteristics for high-efficiency solar energy conversion. Additionally, they successfully produced four compounds in thin-film form for the first time and discovered unexpected "nonlinear bandgap tunability" in certain materials—a property that opens new pathways for solar cell development.
The researchers anticipate that further automation could increase processing speeds by an additional factor of 10 to 100 times. "Our ultimate goal is driving solar energy costs as low as possible," Buonassisi emphasizes. With solar technology already experiencing remarkable price reductions, the team aims to achieve economically sustainable prices below 2 cents per kilowatt-hour—a breakthrough potentially achievable through the discovery of a single material with the ideal combination of manufacturability, low cost, and high efficiency.
"We're establishing all the experimental frameworks necessary to explore materials faster than ever before," Buonassisi concludes.
The research received support from Total SA through the MIT Energy Initiative, the National Science Foundation, and Singapore's National Research Foundation via the Singapore-MIT Alliance for Research and Technology.
