Applying AI for Advanced Semiconductor Node Challenges

Background:

As we are moving to smaller and smaller dimensions in semiconductor manufacturing in accordance with Moore’s Law, process control as well as metrology and defect inspection are becoming more and more challenging. Poor inspection capabilities may potentially cause yield loss due to missing defects at (early) stages of the semiconductor manufacturing process, which in turn, cause an increase in overall production cost. The most viable imaging method at nanometer scale is scanning electron microscopy (SEM), which is high resolution but inherently noisy. Conventional inspection techniques (e.g., canny edge detection for contour extraction) rely on careful selection of certain parameters which are sensitive to noise as well as other imaging anomalies such as contrast change. Modern deep Learning (DL) based frameworks have emerged as a next-generation inspection tool which requires minimal manual tuning and is robust to artefacts such as noise.

Aim:

To propose a deep learning denoiser-assisted framework for the extraction and analysis of SEM contours with main contributions as:

(a) a novel and effective edge extraction technique,

(b) with minimum/no requirement of external user input or metadata (like GDSII / OASIS data, CSV-like meta-data set, etc.) to extract and analyze information from noisy SEM images,

(c) an improved contour extraction algorithm capable to extract contours on the body of noisy raw image itself with a posteriori knowledge derived from its denoised twins.

The proposed contour extraction algorithm, aided by deep learning denoiser, specifically eliminates the presence of any outlier patterns by analyzing the shape properties such as #sides, perimeter, area, etc., then differentiating from the target patterns. We have analyzed, compared, and validated our contour extraction results for each noisy/denoised image pair for categorically different geometrical patterns such as L/S (line-space), C/H (contact-hole), pillars with different scan types, SRAM structures, Logic structures and DRAM (SNLP+BLP) 2D-structures, respectively. We have demonstrated that our proposed method is capable to extract contours on the body of the noisy SEM images with accuracy in close proximity to design data. The proposed framework is shown in Fig. 1.

Results: The figure below gives preliminary results of experiments of contour extraction and defect detection.

Tools used:

• OpenCV:

> Reading and writing images

> Making binary images (for masks)

> Visualization

> Contour extraction and analysis

• PyTorch: > Implementation of neural networks

References:

1. Bappaditya Dey, Sandip Halder, Kasem Khalil, Gian Lorusso, Joren Severi, Philippe Leray, Magdy A. Bayoumi, "SEM image denoising with unsupervised machine learning for better defect inspection and metrology, " Proc. SPIE 11611, Metrology, Inspection, and Process Control for Semiconductor Manufacturing XXXV, 1161115 (22 February 2021); https://doi.org/10.1117/12.2584803

2. Bappaditya Dey, Stewart Wu, Sayantan Das, Kasem Khalil, Sandip Halder, Philippe Leray, Samir Bhamidipati, Kiarash Ahi, Mark Pereira, Germain Fenger, Magdy A. Bayoumi, "Unsupervised machine learning based SEM image denoising for robust contour detection, " Proc. SPIE 11854, International Conference on Extreme Ultraviolet Lithography 2021, 1185411 (30 September 2021); https://doi.org/10.1117/12.2600945

3. Bappaditya Dey, Sandip Halder, Argho Das, Sayantan Das, Stewart Wu, Germain Fenger, "Deep-learning denoiser-assisted framework for robust SEM contour extraction and analysis for advanced semiconductor node, " Proc. SPIE PC12292, International Conference on Extreme Ultraviolet Lithography 2022, PC122920Z (31 October 2022); https://doi.org/10.1117/12.264540