A Multiscale Attention Mechanism Super-Resolution Confocal Microscopy for Wafer Defect Detection

Confocal microscopy is an essential component of wafer defect detection systems. Wafers are raw materials used in the manufacture of semiconductor chips. The semiconductor chip manufacturing process undergoes frequent updates, which cause an increase in the number and types of defects. This leads to lengthy scanning times for large wafers, and warrants the need to enhance the throughput of optical microscopy inspections. To address this issue, we propose the use of the multi-scale residual dilated convolution attention mechanism network (MRDCAN) super-resolution reconstruction algorithm to reproduce high-resolution images from low-magnification objective lens acquired images. The algorithm introduces the attention mechanism to enhance the information richness of wafer images, introduces the multi-scale expansion convolution to expand the convolutional sensor field to eliminate artefacts to enrich the detailed information of wafer image contours, and meets the image quality requirements through the loss calculation method based on the combination of mean-square error (MSE) and structural similarity (SSIM) image evaluation indices. It is shown that the reconstruction of low-resolution wafer images using this algorithm breaks the optical diffraction limit and achieves the purpose of improving the wafer image resolution. Compared with state-of-the-art models, the proposed algorithm can achieve the best performance with an SSIM index of 94.26 percent for the reconstructed super-resolution wafer images. Our algorithm provides fresh insights into the current challenges of confocal microscopy in the field of wafer defect detection. Note to Practitioners—Shrinking semiconductor wafer sizes and increasingly complex inspection steps lead to reduced throughput of optical microscope inspection systems. Current convolutional neural network (CNN) networks cannot solve the problem of super-resolution of complex wafer images well. This seriously affects their application in practical detection. Compared with other algorithms, the super-resolution reconstruction algorithm proposed in this paper has a short training time and a multi-scale structure that effectively prevents the loss function curve from oscillating. And the reconstructed wafer image achieves obvious advantages in terms of visual effect and evaluation indices, with strong robustness to Gaussian noise. In addition, the final discussion shows that high-resolution images can be reproduced through the combination of low-magnification objective lens and deep learning super-resolution algorithm, which can simplify the steps of wafer defect detection and increase the efficiency of the whole wafer defect detection by more than 100%. This study demonstrates the potential of super-resolution confocal microscopy for wafer defect detection.

I. Introduction

Confocal microscopy is a new instrument developed in recent decades with high light sensitivity and high resolution. It has been widely used in biomedical research and surface analysis of materials science [1], [2], [3]. Wafer defect detection is one of the important application directions of confocal microscopy [4]. In recent years, with the growing popularity and demand for smart phones, tablet computers, smart watches and electronic devices, the semiconductor chip industry has developed rapidly. In addition, the development of the global Internet of Things is still in its infancy. Due to the rapid development of 5G communication, there will be a strong demand for semiconductor chips in the future, as will the construction of other modern and intelligent cities [5]. With the development of these industries, the processing costs and detection efficiency of chips are more demanding, so chip processing technologies are constantly updated, chip specifications are continuously reduced, and the demand for chips is constantly increasing [6], [7]. As a result, wafer defect detection is continually facing more difficult challenges, and technical innovation is urgently needed [8], [9].