基於深度學習之遠場強度辨識應用在極紫外光光罩缺陷檢測

Abstract

在半導體製程中，光罩最主要的功能是將電路圖案轉移到晶圓上，因此希望光罩是毫無缺陷的，所以光罩檢測方法在面對越來越精細的製程和微小尺寸的元件時，可能會有檢測方面的問題。因此，希望使用人工智慧成為解決這些問題的方法之一，其能力在於從大量的數據中學習和提取模式，以達到檢測光罩缺陷的任務。 目前光罩上分成吸收層、多層膜和基板缺陷，本論文是專注在吸收層上的缺陷，吸收層上的缺陷有四種凸出、凹陷、大於尺寸和小於尺寸，本研究主要是在利用深度學習從遠場強度回推極紫外光的光罩缺陷種類和相對位置的想法，利用每一種缺陷都會對應到不同的強度資訊，利用強度矩陣轉換成線型圖的方式，用電腦視覺的方法去做深度學習，不藉由反傅立葉轉換和其他種演算法去推回光罩缺陷，因為在同調散射顯微鏡（Coherent Scatterometry Microscope, CSM）架構中，偵測相機只能接收到強度，並不能接收到相位訊號，目前只有用在吸收層的單一缺陷進行時域有限差分模擬，我們希望透過這一種方法，可以知道缺陷的種類和位置，方便進行清除光罩缺陷的後續處理。目前模型可以分辨出三種的缺陷種類，並且可以分類出缺陷的相對位置，對於凹陷、凸出和小於尺寸判斷出缺陷相對位置有較高的準確率，但是大於尺寸目前只有五成的準確率。

In semiconductor manufacturing, the primary function of masks is to transfer circuit patterns onto wafers, thus a defect-free condition of the masks is essential. As processes become more intricate and component sizes decrease, challenges in mask inspection may arise. Consequently, artificial intelligence is hoped to be one of the methods to solve these problems, due to its ability to learn from and extract patterns from vast amounts of data for the task of detecting mask defects. At present, defects on the mask are categorized into those on the absorber layer, multilayer, and substrate, with a focus presently being on defects in the absorber layer. There are four types of defects identified in the absorber layer: extrusions, intrusions, oversize, and undersize. This research primarily revolves around the concept of using deep learning to infer the types and relative positions of EUV mask defects from far-field intensity. By utilizing the specific intensity information corresponding to each defect type, the intensity matrix is transformed into line graphs. Deep learning is then conducted using computer vision techniques, without reliance on inverse Fourier transforms or other algorithms to deduce the mask defects, since only intensity information can be captured by the CCD in the CSM architecture. Currently, simulations are conducted with a single type of defect in the absorber layer using finite-difference time-domain simulations. It is hoped that this method will enable the direct identification of defect types and locations, facilitating subsequent processes for the removal of the defects. Currently, the model can distinguish three types of defects and classify their relative positions. It has a higher accuracy rate in determining the relative positions of defects for intrusions, extrusions, and undersized. However, for oversized defects, the current accuracy rate is only 50%.