Training AI to improve film quality in perovskite solar cells

A Swiss-German research team proposed to use artificial intelligence (AI) methods to establish process parameters for low-variance thin film formation and reproducibility to be able to predict the quality of perovskite solar cells.

It required training the project’s neural networks using a supercomputer and labeled datasets based on photoluminescence videos of the thin film vacuum-based quenching step.

“The study employs sophisticated explainable AI methods that render possible to train an AI system for machine learning, specifically deep learning, to detect hidden signs of good or poor coating from the millions of data items on the videos,” Ulrich W. Paetzold, leader of the next-generation PV group at Karlsruhe Institute of Technology (KIT), told pv magazine.

The videos of the formation of the perovskite thin films contain in situ photoluminescence video data of 1,129 perovskite solar cells, including thin film thickness sensor data and resulting solar cell performance indicators, such as power conversion efficiency. All solar cells in the video dataset were fabricated with a double cation perovskite absorber layer based on Cs_0.17 FA_0.83Pb(I_0.91 Br_0.09)³, then a full solar cell device stack was completed.

The efficiency of the perovskite solar cell, as well as the mean thickness of the perovskite thin-film, served as labels for the neural network training, stated the researchers. After data pre-processing, neural network training and testing, the development of a neural network architecture, development of models and attribution methods were carried out.

The resulting models delivered insights about aspects important to PSC process monitoring and optimization that are typically not available in 2D images, such as the temporal dimension, noted the researchers.

“Thanks to the combined use of AI, we have a solid clue and know which parameters need to be changed in the first place to improve production. Now we are able to conduct our experiments in a more targeted way and are no longer forced to look blindfolded for the needle in a haystack,” said Paetzold.

While the researchers note the potential of the data-driven approach to accelerate and facilitate experimentation in materials science, they caution possible limitations need to be considered. For example, production steps can introduce irregularities that adversely affect PCE but are impossible to predict based on videos used for training.

They also note that a data-driven approach “is naturally limited by the dataset used for analysis” and using higher spatial resolution would make defects and crystal structures more visible to widen the range of potential insights.

Furthermore, they warn of confirmation biases and overinterpretation. To mitigate these potential pitfalls, they noted that the team applied a variety XAI methods that tested the conclusions from different perspectives but also performed a large-scale quantitative evaluation.

They also suggest that inference from XAI results be performed by human experts to control against potential confounding, defined as the presence of lurking variables in AI that can go unknown and unmeasured.

The details of the study are in the study “Discovering Process Dynamics for Scalable Perovskite Solar Cell Manufacturing with Explainable AI,” published in Advanced Materials. The research team had members from KIT, DKFZ German Cancer Research Center, Helmholtz Imaging, Helmholtz AI, and the Swiss Federal Institute of Technology Zurich (ETH).

The team foresees further development of the technology for use in perovskite thin film manufacturing to improve the quality and yield of production processes, according to Paetzold. Furthermore, the technology could provide a blueprint for information-driven and AI-driven processes for perovskite thin film manufacturing, as well as for research in other types of material discovery. Based on the results of the study, the team asserted that XAI methods will play a “critical role” in accelerating energy materials science.