Use machine learning to find what you can trust

Researchers at the University of California, Davis School of Engineering are using machine learning to identify new materials for high-efficiency solar cells. Using high-throughput experiments and machine learning-based algorithms, they found that the dynamic behavior of materials could be predicted with very high accuracy without the need to run many experiments.

This is the work that decorated the cover of the April issue. ACS Energy Letter.

Hybrid perovskites are organic-inorganic molecules that have received much attention for their potential use in renewable energy over the past decade, said Associate Professor of Materials Science and Engineering at UC Davis and senior author of the paper. One Marina Leite said: While some solar cells can be manufactured with efficiency comparable to silicon, they are cheaper to manufacture and lighter in weight, potentially enabling a wider range of applications, including light-emitting devices.

A major challenge in this area is that perovskite devices are more prone to degradation than silicon when exposed to moisture, oxygen, light, heat, and voltage. The problem is to find which perovskites combine high efficiency performance with resilience to environmental conditions.

The general structure of perovskites is ABX3, where A is an organic (carbon-based) or inorganic group, B is lead or tin, and X is a halide (based on chlorine, iodine, fluorine, or a combination thereof). Therefore, “the number of possible chemical combinations alone is enormous,” Leite said. Furthermore, they must be evaluated for multiple environmental conditions, either singly or in combination, resulting in a hyperparameter space that cannot be explored by traditional trial-and-error methods.

“The space of chemical parameters is huge,” says Leite. “It would be very time consuming and cumbersome to test them all.”

High-throughput experiments and machine learning

As a first important step in solving these challenges, Leite and graduate students Meghna Srivastava and Abigail Hering found that machine learning algorithms are effective in testing and predicting the effects of moisture on material degradation. I decided to test whether

Srivastava and Hering constructed an automated high-throughput system to measure the photoluminescence efficiency of five different perovskite films for summer day conditions in Sacramento. They allowed him to collect over 7,000 measurements in one week, accumulating enough data for a reliable training set.

We used this data to train three different machine learning algorithms: a linear regression model, a neural network, and a statistical model called SARIMAX. They compared the model’s predictions to physical results measured in the laboratory. The SARIMAX model showed the best performance in 90% agreement with the observed results over 50 hour windows.

“These results demonstrate that machine learning can be used in identifying candidate materials and appropriate conditions to prevent perovskite degradation,” said Leite. The next step is to extend the experiment to quantify combinations of multiple environmental factors.

The perovskite film itself is only part of a complete solar cell, Leite said. The same machine learning approach can also be used to predict device-wide behavior.

“Our paradigm is unique and we look forward to future measurements. Additionally, we are extremely proud of the hard work of our students during the pandemic,” Leyte said.

Srivastava is a 2021 National Science Foundation Fellow. Additional authors of this paper are Yu An and Juan-Pablo Correa-Baena from Georgia Tech. This work was supported by grants from the National Science Foundation and Sandia National Laboratories.

Disclaimer: AAAS and EurekAlert! EurekAlert! is not responsible for the accuracy of news releases posted. Use of information by contributors or via the EurekAlert system.