SEM Denoising Model Training With Simulated Pattern Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing denoising models for SEM images require extensive training data from patterned substrates, leading to impractical metrology time and computing resources, and necessitate frequent retraining due to limited pattern coverage.
Innovation Solution
A method involving converting design patterns into simulated images using a generator model, adding noise to create training data, and training a denoising model with these simulated images, along with limited SEM images, to enhance pattern coverage and reduce retraining needs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing denoising models are trained using real SEM images of patterned substrates, then the model can achieve reasonable denoising performance, but the metrology time and computing resources required become impractically large
Solution Approach 1:
The patent creates simulated SEM images as copies of real SEM images through a generator model trained on real substrate images. These simulated images serve as training data for the denoising model, replacing the need to use extensive real SEM images for training. The generator model produces synthetic images that replicate the statistical properties and noise characteristics of real SEM images, enabling model training without consuming valuable metrology time on actual substrate measurement.
Solution Approach 2:
The generator model is trained in advance on real SEM images to learn the underlying distribution of substrate patterns and noise characteristics. This preliminary training phase captures the essential features of real SEM data, which are then used to generate unlimited training samples for the denoising model. By performing this action beforehand, the system eliminates the need for time-consuming real-time or near-real-time data collection during model development.
2Productivity
If existing denoising models are trained with limited real SEM images, then training can be completed with reasonable resources, but the pattern coverage is limited requiring frequent retraining
Solution Approach 1:
The generator model produces diverse simulated images that replicate various pattern types and noise conditions observed in real SEM data. By synthesizing images across different pattern categories (e.g., different device structures, layout densities, and manufacturing variations), the training dataset achieves comprehensive pattern coverage without requiring collection of extensive real SEM images across all possible patterns.
Solution Approach 2:
The generator model serves multiple functions: it generates training data for the denoising model, simulates various noise conditions, and creates diverse pattern representations all from a single training process on real SEM images. This multi-functional approach enables the denoising model to be trained on a universally applicable dataset that covers broad pattern variations, reducing the need for frequent retraining when encountering different pattern types.
3Adaptability or versatility
If more real SEM images are collected to improve pattern coverage during training, then the denoising model becomes more versatile, but the metrology time and computational resources increase significantly
Solution Approach 1:
Instead of collecting and processing large quantities of real SEM images, the system uses the generator model to synthesize unlimited training images computationally. These simulated images replicate the essential characteristics of real SEM data including noise patterns, contrast properties, and structural features, providing sufficient training material without the resource-intensive process of acquiring, storing, and processing extensive real image datasets.
Solution Approach 2:
The generator model learns the statistical parameters and distribution characteristics of real SEM images during its training phase. Once trained, it can generate diverse training samples by varying latent space parameters and noise configurations, effectively creating unlimited training data with different pattern variations, noise levels, and image conditions without requiring additional real SEM image acquisitions or substantial computational storage resources.
Data Source
AI summary
Described herein is a method for training a denoising model. The method includes obtaining a first set of simulated images based on design patterns. The simulated images may be clean and can be added with noise to generate noisy simulated images. The simulated clean and noisy images are used as training data to generate a denoising model.