CDSEM Pattern ROI Denoising with 1D AI Scanlines
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing CDSEM denoising techniques, including 2D AI-based solutions, are inefficient in terms of time and accuracy, failing to meet semiconductor industry demands for high throughput and precise measurement of critical dimensions.
Innovation Solution
A 1D AI denoising method using gaussian weighted averaging and deep learning neural networks to process scanlines, involving augmentation and noise addition, followed by iterative parameter updating to create a noise discrimination function.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If 2D AI-based denoising techniques are used, then denoising capability is improved, but processing time increases significantly
Solution Approach 1:
The patent segments the 2D image denoising problem into multiple 1D scanline denoising operations. Each scanline is processed independently through the trained 1D AI model, avoiding the computational burden of 2D processing while maintaining effective denoising. This segmentation approach reduces processing time significantly while preserving the essential denoising functionality.
Solution Approach 2:
The patent transitions from 2D denoising to 1D denoising by changing the dimensionality of the processing approach. Instead of applying AI models to two-dimensional image data, the patent processes one-dimensional scanline waveforms, reducing computational complexity and processing time while achieving comparable or superior denoising results for CD measurement applications.
2Measurement precision
If 2D AI denoising models are trained, then denoising performance is improved, but training time increases to ten or more minutes
Solution Approach 1:
The patent segments the training process by creating 1D AI models that process individual scanlines rather than training complex 2D models on entire images. This segmentation allows for faster model creation and training while maintaining effective denoising performance for the specific application of CD measurement on scanline data.
Solution Approach 2:
The patent changes the parameter dimensionality from 2D to 1D, fundamentally altering the model architecture and training requirements. This parameter change reduces the complexity of the training process, enabling rapid model creation that takes seconds rather than minutes, while still achieving superior denoising results for the target application.
3Productivity
If classical 2D denoising techniques are used, then denoising is achieved, but precision and accuracy are insufficient for CD measurement
Solution Approach 1:
The patent replaces classical mechanical filtering techniques with AI-based denoising. The trained 1D AI model learns optimal denoising parameters and patterns from training data, substituting fixed mechanical filtering algorithms with adaptive intelligent processing that achieves superior precision and accuracy for CD measurement applications.
Solution Approach 2:
The patent performs preliminary training of the 1D AI model using augmented scanline data before actual denoising operations. This preliminary action of training the model with diverse synthetic data prepares it to handle various noise conditions, ensuring high precision and accuracy in subsequent measurements without requiring complex real-time adjustments.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
Significantly reduces denoising time and improves precision and accuracy in CDSEM measurements by leveraging 1D AI denoising, enabling faster and more accurate critical dimension metrology.
Implementation Method 1
scanning an electron beam across a pattern of interest on a substrate; detecting a single scan line waveform of the scanned electron beam
Data Source
AI summary
A system and method for denoising a grey scale image of a pattern on a substrate, including: scanning an electron beam along a pattern, detecting a scan line waveform from the scanned electron beam, obtaining a model scan line waveform from the detected scan line waveform, augmenting the model scan line waveform and adding noise to the augmented model scan line waveform, inputting the noisy augmented model scan line through a deep learning neural network (DNN) process, comparing the output of the DNN process to the augmented model scan line waveform before noise is added, and backpropagating the compared results into the DNN process to obtain a noise discrimination function.