Interpreting Computer Vision Models for Inverse Design of Chalcogenides

The use of data-driven approaches for the design and characterization of materials is growing. An important source of data for such design is the experimentally obtained scanning probe microscopy images, which are inadvertently available in limited volume. The characterization of the images provides information on the quality of the grown materials, their properties, and potential applications. Great advances have been made in the deployment of computer vision models for image detection and analysis. Transfer learning, where the knowledge gained in training a model on larger (usually natural) data is used in materials science target data, has been indispensable in mitigating the effect of limited data and improving the generalization of computer vision models in materials science. However, to maximize the potential of the computer vision models for materials design, it is important to go beyond obtaining the predictions provided to explore the basis and the hows of the model's decisions. For instance, images go through different layers of representation in convolutional neural network (CNN) models. Some features of a few dimensions could be obtained from the later layers, usually with fully connected nodes. A major utility is derivable if such features, known as the latent features, could be interrelated with the physical feature of the image. We present CNN transfer learning models trained on five different classes of transition metal dichalcogenides (TMDs) and differently on the materials based on their growth conditions. The materials include MoS₂, WS₂, WSe₂, MoSe₂, and Mo-WSe₂, which were grown by metal-organic chemical vapor deposition (MOCVD) at the 2D Crystal Consortium (2DCC) of the Pennsylvania State University. The trained models, which present excellent accuracy, enable a systematic approach to analyzing the interrelationship among the latent features, TMD classes, their chemical composition, the materials properties, and the growth conditions. This demonstrates the use of computer vision for the inverse and high-throughput design of materials for technological application.