Mixup-Based Neural Network for Image Restoration and Structure Prediction From SEM Images

Scanning electron microscopy (SEM) has been widely used for the semiconductor industry since it provides high-resolution (HR) details of the semiconductor. However, there is a gap in research for various tasks (i.e., image restoration (IR) and structure prediction) in SEM datasets collected under various conditions. Therefore, we introduce a new SEM dataset with diverse characteristics such as energy, noise, current with various levels for IR, and structure prediction. Furthermore, we propose a new deep-learning-based method for this dataset. The method consists of two stages: IR stage and structure prediction stage. In the IR stage, we design the transformer-based architecture to use pixel information widely. In the structure prediction stage, we introduce a novel training algorithm, SEMixup, and a novel CNN-based network, SEM structure prediction network (SEM-SPNet). Specifically, SEMixup overcomes the generalization and robustness of SEM-SPNet by implicitly interpolating a pair of samples and their labels. Experiments demonstrate that our method achieves state-of-the-art results across all dataset conditions. This work expands the possibilities of SEM image analysis using deep learning, contributing to the semiconductor industry.

I. Introduction

In the semiconductor industry, quantitative analysis of semiconductor structures, including their width, depth, and properties, is increasingly important as the semiconductor structures become more sophisticated and more complex as shown in Fig. 1. Recently, image-based analysis technology has been required by increasing the high-resolution (HR) measurements to obtain the structural information from the semiconductor. Specially, scanning electron microscopy (SEM) image, a 2-D representation of the electron interactions with the surface, has been widely used to observe the detailed surface of a microchip since SEM image provides the HR local geometry information. In practical, the analysis of SEM image is one of the most important tasks to measure the geometric dimensions of the structures on the microchip. However, it is time-consuming and inefficient to manually analyze the SEM images manually. Formerly, the handcraft-based methods [1], [2], [3], i.e., image processing, have been proposed to overcome this issue, but due to the complexity and diversity of nanoparticles of SEM images, there was still a limitation: it is difficult to accurately predict the structure information. Therefore, it is necessary to develop a data-driven method, such as deep learning, to efficiently extract the structural information from the SEM images. Fig. 1.

Development of the semiconductor structure. The pink part composed of Si is the source and drain of the semiconductor. The gray part is a gate to control the current between the source and the drain. The red part is where shallow trench isolation (STI) technology is applied to separate various electronic components within the semiconductor chip. The semiconductor structures are progressively increasing in complexity.