Boundary-Aware Semantic Segmentation for Semiconductor Metrology
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing semiconductor metrology methods face challenges in accurately measuring nano-scale semiconductor devices due to noise in TEM and SEM images, which blur object boundaries and degrade image quality, and the scarcity and ambiguity of training data for semantic segmentation models.
Innovation Solution
A learning-based semantic segmentation method using a pre-training stage with encoder-decoder architectures and a boundary-focused loss function (LBF) to enhance the recognition of object boundaries in TEM and SEM images, employing transfer learning and morphological edge detection to improve boundary clarity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional semantic segmentation is used on noisy TEM/SEM images, then processing speed is maintained, but measurement precision deteriorates due to blurred object boundaries
Solution Approach 1:
The patent applies pre-training on large-scale datasets before fine-tuning on specific noisy TEM/SEM images. This preliminary action on clean data establishes robust boundary detection capabilities that can later withstand the noise and blur present in actual metrology images, thereby improving measurement precision without requiring the model to learn from noisy data directly.
Solution Approach 2:
The patent employs different loss functions for different regions: Dice loss for general segmentation and Boundary Loss specifically for boundary pixels. This local differentiation allows the model to prioritize accurate boundary detection in critical regions while maintaining overall segmentation quality, directly addressing the blurred boundary problem in noisy images.
2Productivity
If manual inspection is used for nano-scale measurements, then measurement precision can be maintained by experienced inspectors, but productivity deteriorates due to time-consuming processes
Solution Approach 1:
The patent implements automated semantic segmentation that performs its own optimization through the proposed loss function that automatically identifies and weights boundary pixels. The model self-adjusts to focus on boundary regions without manual intervention, enabling high-speed automated processing that maintains precision comparable to expert manual inspection.
Solution Approach 2:
The patent changes the optimization parameter by introducing a boundary-aware loss function that modifies how pixels are weighted during training. By assigning higher weights to boundary pixels through the Dice loss and Boundary Loss combination, the model learns to prioritize boundary accuracy, achieving automated measurements with precision that would otherwise require manual expert inspection.
3Measurement precision
If standard loss functions are used for training, then training simplicity is maintained, but boundary detection accuracy deteriorates in noisy images
Solution Approach 1:
The patent applies different loss functions to different spatial regions: Dice loss for the entire image and Boundary Loss specifically for boundary pixels. This local differentiation ensures that boundary regions receive specialized attention during training, improving boundary recognition accuracy in noisy images without requiring complete redesign of the entire loss function.
Solution Approach 2:
The patent combines multiple loss functions (Dice loss and Boundary Loss) into a composite loss function. This composite approach leverages the strengths of each individual loss function: Dice loss for overall segmentation quality and Boundary Loss for precise boundary detection, achieving superior boundary recognition accuracy while managing complexity through modular composition.
Data Source
AI summary
A learning-based semantic segmentation method and apparatus for semiconductor metrology are disclosed. The method includes performing, using a processor, a pre-training stage to determine initial weights among nodes within a neural network model by pre-training the neural network for process-specific semantic segmentation; and performing, using a processor, a fine-tuning stage to classify an input wafer TEM or SEM image into at least one object of interest based on pre-trained weights, and to assign a weight (α) greater than one ( ) to pixels corresponding to boundaries separating the objects of interest and a weight one (1) to other pixels corresponding to regions distinct from the boundaries using a loss function (LBF).