Improvement of Spatial Resolution of Long-Working Distance Measurement using Convolutional Neural Network

Abstract

The research in this dissertation aims at improvement of spatial resolution using convolution neural network in critical dimension (CD) measurement. CD measurement is widely used to get characteristic features of inspection targets in manufacturing process. This metrology step examines whether or not patterns are fabricated as designed. The results of such implementation are important for manufacturing process to produce high-quality products reliably.<br/> In recent years, organic light-emitting diodes (OLEDs) display continues to be a major trend, improving high pixel density. Because most OLED display are made using fully automated encapsulation system in order to prevent organic materials from oxidation, metrology equipment locates out of system, for example, vacuum chamber. This prevalence have demanded long working distance (W.D) measurement of small features, satisfying the measurement performance of industry–level at the same time.<br/> Long W.D measurement result in reducing numerical aperture (NA) of optical system. It is desirable to use a larger lens in order to increase the NA, however, telephoto lens are expensive and heavier. Imaging from large stand-off distance typically suffer from low spatial resolution. Edge detection senses the intensity change of image and separates the object from the background by finding a boundary line. Therefore, blurry image drops the edge detection performance and it is hard to attain high-accuracy and high-repeatability. <br/> In this thesis, convolutional neural network (CNN) was suggested to improve spatial resolution without any changes in optics. The proposed convolutional neural network use a single image acquired from long working distance measurement system as input, and rapidly outputs an image that having an improved spatial resolution. The convolution neural network was used to connect two different optical systems. We had learned the network so that the long-distance measured image was transformed like an image measured at a very close distance. In order to improve the performance and speed of learning, neurons, layers, and input-output data were newly constructed. In addition, it was configured to learn and correctly recognize dead neurons that were not learning properly in the network<br/> Under 195 mm W.D condition, the measurement accuracy and 3σ repeatability is 3.0%, 60 nm (20 repeat measurements), respectively. The measurable minimum size is 0.5 ㎛ and the measurement time is only under 0.2 s. The proposed method not only showed sub-micro level resolution, but also successfully industry-level measurement accuracy and repeatability in real time by implementing CNN. These results highlight the promise of the proposed method as a long-working distance measurement system of industry field. <br/>