Reinforced active learning for CVD-grown two-dimensional materials characterization

AbstractTwo-dimensional (2D) materials are one of the research frontiers in material science due to their promising properties. Chemical Vapor Deposition (CVD) is the most widely used technique to grow large-scale high-quality 2D materials. The CVD-grown 2D materials can be efficiently characterized by an optical microscope. However, annotating microscopy images to distinguish the growth quality from good to bad is time-consuming. In this work, we explore Active Learning (AL), which iteratively acquires quality labels from a human and updates the classifier for microscopy images. As a result, AL only requires a limited amount of labels to achieve a good model performance. However, the existing handcrafted query strategies in AL are not good at dealing with the dynamics during the query process since the rigid handcrafted query strategies may not be able to choose the most informative instances (i.e., images) after each query. We propose a Reinforced Active Learning (RAL) framework that uses reinforcement learning to learn a query strategy for AL. Besides, by introducing the intrinsic motivation into the proposed framework, a unique intrinsic reward is designed to enhance the classification performance. The results show that RAL outperforms AL, and can significantly reduce the annotation efforts for the CVD-grown 2D materials characterization.Keywords: Two-dimensional (2D) materials characterizationactive learning (AL)reinforcement learning (RL)intrinsic reward AcknowledgmentsThe authors thank Jihea Lee for her efforts on data collection and Dr. Alexand er Kuhnle for his kind support.Data availability statementThe data and code that support the findings of this study are available at https://doi.org/10.5281/zenodo.8045494.Additional informationFundingThis work was partially supported by New York State Energy Research and Development Authority (NYSERDA) under Award 138126 and the New York State Center of Excellence in Materials Informatics (CMI). The authors acknowledge the support from the Vice President for Research and Economic Development (VPRED) at the University at Buffalo.Notes on contributorsZebin LiZebin Li received a BE degree in thermal energy and power engineering from Sichuan University, Chengdu, China, in 2016 and M.Sc.Eng. degree in microelectronics and solid-state electronics from University of Chinese Academy of Sciences, Beijing, China, in 2019. He is currently pursuing a PhD degree in the Industrial and Systems Engineering Department, University at Buffalo. His research focuses on data analysis on advanced manufacturing.Fei YaoFei Yao received her dual PhD degree in energy science from Sungkyunkwan University (SKKU), Korea, and in physics from Ecole Polytechnique, France, in 2013. Currently, she is an assistant professor in the Department of Materials Design and Innovation, University at Buffalo. Her research interests include low-dimensional materials synthesis, property engineering, and their applications in electrochemical and electronic devices.Hongyue SunHongyue Sun is an assistant professor in the Department of Industrial and Systems Engineering, University at Buffalo. He has a multidisciplinary background with a BE in mechanical engineering, MS in statistics, and PhD in industrial engineering, respectively. Dr. Hongyue Sun’s research interests are data science for advanced manufacturing, occupational safety, and healthcare systems. His research has been broadly supported by NSF, NIOSH, MxD, etc. His research has been recognized by several best paper awards from INFORMS and IISE. He received the Outstanding Young Manufacturing Engineer Award from the SME and UB’s Exceptional Scholar: Young Investigator Award from University at Buffalo in 2023. He is a member of IISE, INFORMS, IEEE, and ASME.