Image Quality Measurement: a novel approach to control metrology uncertainty and process quality of in-line CDSEM

As CD size shrinks, process margin and its correspondent metrology margin are getting tightened, which means metrology induced measurement uncertainty should also be reduced. To proceed diagnostics of process control, so far, CD trend was managed with dedicated standard layer/wafer for quality control. But tighter process/metrology margin makes it hard to maintain and control only with CD trend management. Top tier semiconductor industry is trying to establish additional key index for monitoring quality control and preventing any abnormal trend issues, including several hardware (HW) parameters, such as probe current, beam sharpness (resolution), abrupt changes of beam alignment/astigmatism coil current, and so on. In addition to the HW parameters, several other approaches are also being considered for capturing abnormality of data trend. Among those, utilizing image itself might be best candidate as images containing plenty of information on its own. There could be always doubt whether measured CD values representing true process (i.e, process capability index) or not. What if tool stability was getting worse suddenly but no CD metrics and HW index could capture those abnormality? If we can capture some of the abnormality from raw images themselves, it will be a good indicator for judging whether measured CD is reliable or not. In this work, we propose image quality measurement (IQM) associated with image sharpness and contrast to noise ratio (CNR) for tracking image health. By applying unique algorithm on each image of SEM metrology on any pattern, image sharpness and CNR could be outputted with quantities. We performed several designs of experiment (DOE) whether IQM is well representative per each application and each layer by utilizing recent cutting edge DRAM layers with various layer types (DI/FI) and applications including CD and overlay. In order to imitate various instable SEM tool situation, we verified both with recipe and HW knob DOE. Regarding recipe DOE, we controlled spot size for imitating defocus situation, the number of frames for imitating shot noise impact and low CNR situation, and gaussian filter adoption for imitating image processing impact. Regarding HW knob DOE, we manipulated knobs to imitate variations in electric field, and dosage. By integrating all these results, we have verified the performance and applicability of IQM. In addition, we conducted an evaluation to establish criteria for identifying abnormal CD measurement results by adapting machine learning model based on IQM. This enabled us to develop standards for effectively eliminating inaccurate CD results caused by tool-related issues, which is required for tighter robust process controls.