Deep-learning-based SEM image denoiser
Dorin Cerbu1, Sandip Halder1, Philippe Leray1
1IMEC, Kapeldreef 75, B-3001 Leuven, Belgium
We report the development of a new method to denoise SEM images with the help of artificial neural networks. Upon using a preprocessing and training scheme tailored for SEM images of structures, most often encountered in semiconductor manufacturing, we can efficiently denoise images affected with varying degrees of noise severity and origin. In the figure below, we show an example of how we can use this filter efficiently to treat noisy images and improve the image quality. This can help in acquisition of more stable and better metrology data.
Fig1(a) original image (b) Image which has been denoised using deep-learning based algorithms
This development is of utmost importance for the case of post-litho processing step where resist nanostructures when SEM inspected are usually impacted by the electron beam and shrink, hence skewing critical dimension measurements. This is especially true as we push towards sub N-10 nm nodes. Application of our deep-learning processing scheme allows efficient noise reduction on SEM inspection images and helps us discern minor details previously shadowed by noise. This is extremely important as we move towards using EUV in high volume manufacturing. Small details can be crucial to understand the root-cause of stochastic and process defects. In previous work, we have already shown different approaches to understand stochastic defects [1-2]. The goal of this work is to enhance the image quality as much as possible to gain further fundamental understanding on nano-defects.
 S. Halder et. al., ‘Using machine learning techniques to understand EUV stochastics, SPIE Photomask Technology + Extreme Ultraviolet Lithography, 2018
 K. Sah et.al., ‘EUV stochastic defect monitoring with advanced Broadband optical wafer inspection and e-Beam review systems’, SPIE Photomask Technology + Extreme Ultraviolet Lithography, 2018
CD-based process windows have been an analysis workhorse for estimating and comparing the robustness of semiconductor microlithography processes for more than 30 years. While tolerances for variation of CD are decreasing in step with the target CD size, the acceptable number of printed defects has remained flat (Hint: Zero) as the number of features increases quadratically. This disconnect between two key process estimators, CD variability and defect rate, must be addressed. At nodes that require EUV lithography, estimating the printed defects based solely on a Mean CD (“Critical Dimension”) process window is no longer predictive. The variability / distribution of the printed CDs must be engineered so that there are no failures amongst the billions of instances, rendering the Mean CD, often measured on just hundreds or thousands of instances, a poor predictor for outliers. A “defect-aware” process window, where the count of printed defects is considered in combination with more advanced statistical analysis of measured CD distributions can provide the needed predictability to determine whether a process is capable of sufficient robustness. Determining process robustness where stochastics and defects are taken into account can be simplified by determining the CD process margin. In this work we study dense contact hole arrays exposed with 0.33NA single exposure EUV lithography after both the lithography and etch steps. We describe a methodology for expanding the analysis of process windows to include more than the mean and 3σ of the data. We consider the skew and kurtosis of the distribution of measured CD results per focus-exposure condition and compare / correlate the measured CD process window results to the CD process margin.