Recent Achievements of the Team - High-Throughput Material Image Segmentation
Our team, in collaboration with Professor Yang Jiong’s research group from the Institute of Materials Genome Engineering, has recently published a paper titled “Accelerating the Discovery of Cu−Sn−S Thermoelectric Compounds via High-Throughput Synthesis, Characterization, and Machine Learning-Assisted Image Analysis” in the international journal “Chemistry of Materials” (IF: 9.872, Q1 in JCR and Chinese Academy of Sciences SCI journal ranking). “Chemistry of Materials” is one of the most influential top-tier academic journals in the field of engineering, technology, and materials science. The paper is affiliated with Shanghai University as the first institution, with Sheng Ye, a PhD student from the Institute of Materials Genome Engineering, as the first author. Professor Yang Jiong, Professor Xi Jinyang from the Institute of Materials Genome Engineering, and Associate Professor Han Yuexing from the School of Computer Engineering and Science, Shanghai University, are the corresponding authors.
High-throughput (HTP) methods have emerged as powerful approaches for accelerating materials research and development. In this work, we demonstrate the capability of combining machine learning (ML) image segmentation techniques with HTP synthesis and characterization for the discovery of new thermoelectric compounds. First, cylindrical samples with nine different compositions were synthesized using an improved diffusion couple HTP synthesis method. Subsequently, scanning electron microscopy (SEM) characterization was performed on fixed fragments of each composition, resulting in the acquisition of 11 backscattered electron (BSE) images per fragment (99 images in total). To rapidly segment the different phases in the 99 BSE images, we propose two ML image segmentation strategies: a supervised active learning strategy utilizing a fully connected neural network and an unsupervised clustering strategy. The supervised strategy significantly reduces the time cost of image segmentation, achieving a classification accuracy of 0.9. This model is then used for automatic segmentation of a large batch of images. To further reduce the workload of manual annotation, active learning is introduced during the training process. Analyzing the energy-dispersive X-ray spectroscopy (EDS) characterization of the two major phases separated by the first strategy, we discovered a new compound, Cu7Sn3S10, which exhibits promising thermoelectric properties with a zT value exceeding 0.6. In contrast to the supervised strategy, the unsupervised strategy allows for the identification of compounds that may have been overlooked in the BSE images without manual annotation. Based on the results of the unsupervised strategy, we discovered an unreported compound, Cu1.6S. Overall, our work showcases an example of a fully HTP workflow, enabling automated analysis of HTP characterization and new compound identification.