Perovskite solar cells have gained huge momentum in the search for cheaper and more efficient solar energy. However, the materials degrade over time, limiting their widespread commercial use.
To overcome one of the biggest barriers to commercializing perovskite solar cells, Marina Leite , a professor of materials science and engineering at the University of California, Davis, and an interdisciplinary team of researchers are harnessing AI.
In a paper published in Advanced Materials , the researchers demonstrated how AI can dramatically accelerate research advancements. Instead of relying solely on trial-and-error experimentation, the team used AI to learn from thousands of automated experiments and accurately predict how new material compositions will respond to heat, a significant environmental stressor, to identify the most promising materials more quickly.
The search for stable perovskites
Compared with conventional silicon solar cells, perovskites are lighter, more flexible, less expensive to manufacture and highly efficient. However, they are unstable in response to environmental stressors such as heat, moisture and light, which limits their scalability.
"Our scientific community is very interested in understanding the chemical and physical processes that drive the stability, or lack thereof, within these materials," Leite said.
Testing every possible perovskite material composition under every single environmental condition would be nearly impossible. Instead, Leite's team asked whether AI could be taught to recognize patterns and predict behavior from a carefully selected subset of experiments.
The team tested 10 perovskite compositions by exposing them to repeated temperature cycles, resulting in 137,000 unique measurements. The measurements were used to train machine learning models that could accurately predict how previously untested material compositions would respond to repeated heating.
From these predictions, the researchers identified which compositions were more thermally stable. Compositions with lower levels of cesium, an extremely reactive alkali metal, generally recovered after repeated heating, while compositions with higher cesium content were more likely to degrade permanently.
AI: A research partner
The findings give researchers a roadmap for developing more durable perovskite solar cells. Instead of experimentally testing thousands of possible recipes, they can now focus their efforts on the candidates most likely to withstand real-world operating conditions.
"AI does not replace experiments in our research," Leite said. "Instead, it can be used to increase the efficiency of scientific discovery."
Leite and her collaborators used AI to learn relationships from a limited set of high-throughput experiments. Using the appropriate algorithms, AI could accurately forecast the stress-response behavior of halide perovskites under previously unmeasured conditions.
Eventually, Leite hopes the same approach can be used to predict how perovskite materials will perform in different climates around the world, helping researchers design solar cells tailored to real operating conditions and paving the way for more efficient solar energy.
"Our successful implementation of machine learning models to analyze experiments is an important demonstration of how AI can be truly helpful."
Co-authors on the paper include Abigail Hering and Mansha Dubey, Ph.D. students in materials science and engineering at UC Davis; materials science and engineering alum Meghna Srivastava; Ph.D. student Elahe Hosseini and Professor Houman Homayoun of electrical and computer engineering at UC Davis; and Yu An and Juan-Pablo Correa-Baena from Georgia Institute of Technology.
The project received financial support from several entities, including the Defense Advanced Research Projects Agency, National Science Foundation and the Department of Energy.
