Accelerating the Discovery of Cu-Sn-S Thermoelectric Compounds via High-Throughput Synthesis, Characterization, and Machine Learning-Assisted Image Analysis

High-throughput (HTP) methods have become a powerful method for accelerating the research and development of materials. In this work, we present an example of the whole HTP chain for seeking novel Cu-Sn-S ternary thermoelectric materials including three parts: HTP synthesis, HTP characterization, and HTP analysis. First, the modified diffusion-couple HTP synthesis method is utilized to obtain a bulk sample with nine different raw material (CuS and SnS) ratios. Then, each segment with a fixed raw material ratio is characterized by scanning electron microscopy, and 11 backscattered electron images are taken (99 in total). Finally, we propose two different strategies to classify the different areas in the 99 backscattered electron images and further identify the different phases. The first strategy applies an active learning loop with a fully connected neural network aiming at fast and automated image segmentation. The second strategy employs an unsupervised clustering method to find potentially overlooked compounds without prior knowledge. Two interesting phases Cu7Sn3S10 and Cu1.6S are found and further analyzed, and the former is characterized to be a potential thermoelectric compound with a zT of over 0.6.

Automated summary

This study presents an example of using high-throughput methods to search for novel Cu-Sn-S ternary thermoelectric materials, involving synthesis, characterization, and analysis. Different strategies were proposed to classify different areas in the images and two interesting thermoelectric phases, Cu7Sn3S10 and Cu1.6S, were identified with potential for high thermoelectric efficiency.