Microscopic Image Deblurring by a Generative Adversarial Network for 2D Nanomaterials: Implications for Wafer-Scale Semiconductor Characterization

Wafer-scale two-dimensional (2D) semiconductors with atomically thin layers are promising materials for fabricating optic and photonic devices. Bright-field microscopy is a widely utilized method for large-area characterization, layer number identification, and quality assessment of 2D semiconductors based on optical contrast. Out-of-focus microscopic images caused by instrumental focus drifts contained blurred and degraded structural and color information, hindering the reliability of automated layer number identification of 2D nanosheets. To achieve automated restoration and accurate characterization, deep-learning-based microscopic imagery deblurring (MID) was developed. Specifically, a generative adversarial network with an improved loss function was employed to recover both the structural and color information of out-of-focus low-quality images. 2D MoS2 grown by the chemical vapor deposition on a SiO2/Si substrate was characterized. Quantitative indexes including structural similarity (SSIM), peak signal-to-noise ratio, and CIE 1931 color space were studied to evaluate the performance of MID for deblurring of out-of-focus images, with a minimum value of SSIM over 90% of deblurred images. Further, a pre-trained U-Net model with an average accuracy over 80% was implemented to segment and predict the layer number distribution of 2D nanosheet categories (monolayer, bilayer, trilayer, multi-layer, and bulk). The developed automated microscopic image deblurring using MID and the layer number identification by the U-Net model allow for on-site, accurate, and large-area characterization of 2D semiconductors for analyzing local optical properties. This method may be implemented in wafer-scale industrial manufacturing and quality monitoring of 2D photonic devices.