Semiconductor Wafer Particle Defect Classification Based on Deep Learning and Multimodal Feature Fusion

In the semiconductor manufacturing process, the identification and classification of the chemical composition of particle defects on wafers is critical to ensuring product quality. Traditional manual classification methods are affected by the operator's subjective judgement, which affects the accuracy and efficiency of classification. In this study, we propose a deep convolutional neural network (CNN)-based automatic defect classification model that fuses scanning electron microscope (SEM) images and energy dispersive X-ray (EDX) spectral data for more efficient and accurate defect identification.Our model first feeds the image data into a network with Residual Neural Network (ResNet-50) as the model backbone structure, which has learnt to recognise various types of defects through a large number of samples, and introduces a Convolutional Block Attention Module (CBAM) attention mechanism to the network, which aims to improve the network's ability to recognise defects. The attention mechanism is designed to enhance the network's capability to focus on and learn essential features. It employs a Multi-Layer Perceptron (MLP) to extract features from Energy Dispersive X-ray (EDX) spectra. Subsequently, the integration of image and spectral features is achieved through a multi-head self-attention mechanism. Finally, a fully connected layer is utilized to classify eight different types of defects within semiconductor particles.The experimental results show that our system improves the accuracy and processing speed of defect detection compared to traditional manual classification methods, and the comparison of the classification detection accuracy of different models for different input models is investigated.