UC Davis researchers use machine learning to study perovskite solar cells’ properties and advance them
By EKATERINA MEDVEDEVA— science@theaggie.org
Have you ever wondered what solar cells are made out of? Most of the ones that you see nowadays are composed of silicon. They have been the standard of solar cells for over a decade due to being relatively cheap, efficient and durable.
“[Silicon] modules are expected to last for 25 years or more [while] still producing more than 80% of their original power,” an article by the U.S. Department of Energy reads.
However, there is another group of semiconducting materials that have shown remarkable growth in efficiency over the last 10 years and have the potential for low-cost production thanks to requiring simpler processing methods in the manufacturing process compared to silicon — the metal halide perovskites.
This group of crystalline materials has a general composition ABX3, where A stands for an organic cation, inorganic cation or a mixture of the two; B stands for lead or, more rarely, tin cation; and X stands for a halide, usually I−, Br− or Cl−. By adjusting the types of organic/inorganic cations and halides present in the composition of the perovskite, researchers are able to fine-tune their efficiency at absorbing sunlight at different wavelengths, including those not captured by silicon.
UC Davis research group led by Marina S. Leite, a professor of materials science and engineering and a Chancellor’s Fellow, has investigated the properties of different hybrid organic-inorganic perovskites (HOIP) using Artificial Intelligence (AI). This method of studying HOIPs has significantly sped up the research process and is projected to accelerate it even further in the coming years.
“We have an enormous family of chemical compositions, and besides that, this material can change in properties once it is exposed to [different] environmental factors,” Leite said. “So, there are a lot of parameters that can be varied simultaneously, and no one fully knows the influence of each one when they are working individually or in a coupled manner. It would take us an extremely long time as humans to solve this problem [without AI]. It would be just impossible.”
In the experimental process, a sample of perovskite thin film is placed on a substrate and put into small nitrogen-filled environmental chambers “to exclude uncontrolled material changes/degradation.” Then, the sample is loaded into a tightly sealed enclosure with an optical set-up where a laser beam is shot at it, exciting the electrons from the valence band to the conduction band. Shortly after, the electron returns back to the fundamental state, while emitting a photon in the process called photoluminescence. The spectra produced from the photons are the main measurement of interest.
The research team uses a remotely controlled automated system to run these experiments, which they developed during the pandemic when access to the lab was limited.
“These [automated] measurements take less than one second, which enables us to obtain over 20,000 spectra [measurements] per week and consequently have sufficient data to train our machine learning (ML) algorithms,” Leite said. “They are quite informative, too, because we can infer what would be the performance of a full solar cell, even though we just have the material.”
The team’s research considers various types of ML models including linear regression, neural networks and image classification in all steps of designing and testing perovskite for future photovoltaic cells. This is done in order to determine optimal tools for predicting their behavior under the influence of different stressors; for example, finding the hidden correlation between their chemical composition and thermal stability.
So far, this method of studying perovskite solar cells looks very promising. In one of the studies, Leite’s team was able to achieve a 90% accuracy of prediction of HOIPs’ optical behavior over a 50-hour period while exposed to relative humidity cycles, corresponding to those during summer days in Northern California. With more training and testing data collected, the accuracy is projected to grow even further.
As mentioned earlier, the main problem to address with the perovskite solar cells is their susceptibility to degradation under real environmental conditions. However, even when they will become durable enough to be on par with silicon ones, most likely they will not serve as a replacement. Instead, they will complement the current solar cells with their properties, which is enabled by their thin layering and extended spectra of absorption.
“The idea is having a halide perovskite, putting it on top of silicon and having what we call a dual junction,” Leite said. “Because silicon [solar cells are] everywhere, doing this addition is a way that does not increase cost in a concerning way, but increases performance.”
Commenting on the highlights of this work, Leite emphasized the contribution and talent of the students with whom she has worked with. Currently, this project involves two fourth-year Ph.D. students, Mansha Dubey and Abigail Hering, who are working on the different aspects of this research.
The future for the study looks bright as the team collaborates with researchers in and outside of UC Davis, including the UC Davis Electrical and Computer Engineering Department and the Lawrence Berkeley National Lab, integrating their infrastructure and expertise.
Written by: Ekaterina Medvedeva— science@theaggie.org
