Model-based contour extraction: an enabler for very low-frame SEM images metrology

Among metrology tools in the semi-conductor manufacturing, critical dimension scanning electron microscopes (CD-SEM) are the most broadly used, especially due to their high resolution, low destructivity, and high throughput. Contour metrology on CD-SEM images has become essential for characterization, modelling, and control of advanced lithography processes. In particular, OPC model's accuracy can be highly improved using contours metrology. One of the issues when dealing with CD-SEM metrology is that the results are noise sensitive. Moreover, diminishing noise in CD-SEM acquisition leads to resist shrinkage due to exposure time increase. In addition, post-treatment of these shrinkage effects requires compensation algorithms such as artificial intelligence (AI)- driven algorithms, that are another contributor to the error budget of metrology systems. There is thus a need for an accurate, robust to noise, and purely deterministic edge detection algorithm. In this article, we evaluate the benefits of relying on a model-based contour extraction approach for performing measurements. This approach is applied onto both synthetic and experimental CD-SEM images with various patterns (mostly 2D) and noise levels to assess the influence of image integration (frame number) on the contour detection and CD measurement. We demonstrate that a model-based contour extraction algorithm is able to precisely characterize SEM-induced 2D resist shrinkage. We observe that this model-based approach is more robust to noise than standard algorithms by 21% on synthetic data and by 36% on experimental data. Another way of seeing it is, while keeping the same precision, a model-based contour extraction approach can significantly reduce the requested image frame number. The benefits of adopting this approach range from reducing the shrinkage effects to improving SEM image acquisition time. Eventually, no step of shrinkage modelling calibration nor AI-driven image post processing are needed which implies a gain on simplicity and avoids modelling errors.