Multiclass Identification of Van der Waals Semiconductors via Attention Neural Networks

Intelligent characterization of van der Waals semiconductors is an essential process for industrial manufacturing and laboratory fabrication. A combination of microscopic images and artificial intelligence models is an efficient way for wafer‐scale layer number identification of van der Waals semiconductors. This methodology overcomes the bottleneck of the conventional manual layer number counting approach, which requires a long period of manual inspection and induces high error rates when distinguishing layers with similar appearance. Here, a convolutional architecture that involves a fused network of ResNet‐Inception with Attention Layer (RIAL) is developed for accurate multiclass classification of randomly distributed layers of chemical vapor deposition (CVD)‐grown van der Waals semiconductors. RIAL model is first validated on the single‐label datasets CIFAR‐10/100, and subsequently fine‐tuned on the custom‐built microscopic image datasets of CVD‐grown MoS 2 . To compare with semantic segmentation, U‐Net with Attention Layer (UNAL) is further implemented for pixel‐wise classification of multiclass semiconductors. The quantitative analysis of RIAL and UNAL illustrates the versatility of attention convolutional network models in the wafer‐scale identification of van der Waals semiconductors.