|
Continuous reduction in pattern size, the primary path of advancement for the semiconductor industry, has greatly increased resolution and throughput demands for defect inspection and metrology, where Electronic-beam (E-beam) wafer inspection equipment has been commonly used for both purposes. High resolution is specifically needed in order to inspect or measure these smaller patterns and is accomplished by either decreasing pixel size or increasing frame averages. Both of these adjustments come with a big penalty of throughput, which is an extremely important metric as large areas of the wafer must be inspected in a reasonable time to meet semiconductor development, yield ramp and high volume manufacturing process control requirements. A slow inspection means more inspection tools are required and lots are delayed by the longer process time. In order to regain throughput, it is common to try to back off on the frame averages, but this often results in low quality images with noise, blurring effects, and distortions. The end result is less defect sensitivity for inspections, lower CD measurement accuracy and precision for metrology. Image quality enhancement (IQE) algorithms can compensate for this and thereby play a significant role in achieving higher throughput while keeping sufficient sensitivity. In recent years, deep learning methods have demonstrated superior performance to traditional algorithms for IQE. However, these methods often require clean ground truth data for supervised training purposes, which is extremely difficult and expensive to achieve. For example, ground truth images with lower noise levels can be obtained by averaging hundreds of frames at the same location, but, in addition to taking a very long time, can cause permanent physical damage to the wafer due to the E-beam wafer imaging process, and unexpected artifacts or shadowing effects. In order to alleviate these issues, we propose an unsupervised machine learning- based image quality enhancement framework (uMLIQE) using deep learning methods, which does not require clean target images for the training process. In fact, only one or a few images are required since the required information can be extracted by segmenting the available image. The performance of this system was compared both via simulation and experimentally to a comprehensive list of alternate IQE approaches. The wafer we used for data collection was generated with standard semiconductor processing representative of CMOS processing across the industry. The unsupervised approach is clearly superior to all alternatives both qualitatively and quantitatively. Our proposed unsupervised deep learning IQE framework for SEM images has proven superior for throughput enhancement for high resolution E-beam wafer imaging.
|
