Deep Learning Model for Semiconductor Pattern Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current device manufacturing processes, particularly in semiconductor production, face challenges in accurately transferring patterns onto substrates due to limitations in pattern recognition and process control, leading to defects such as edge placement errors and line edge roughness.
Innovation Solution
A deep learning model is trained using convolution operations with dilation kernels to segment images of substrates, allowing for precise feature recognition and parameter determination, such as edge placement error and line edge roughness, enabling adjustments to the patterning process to improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional pattern recognition methods are used in lithography, then the device complexity remains low, but the manufacturing precision deteriorates due to edge placement errors and line edge roughness
Solution Approach 1:
The patent replaces conventional mechanical/optical pattern recognition methods with a deep learning-based computational system. A neural network model processes aerial images to identify and classify pattern features, substituting traditional image processing algorithms with machine learning-based feature extraction and classification to achieve higher measurement precision for edge placement error and line edge roughness
Solution Approach 2:
The patent introduces an intermediary deep learning system between the lithographic exposure process and the measurement/inspection stage. The neural network acts as a mediator that processes aerial images and extracts precise feature measurements, enabling improved pattern transfer accuracy without directly modifying the lithography hardware
2Measurement precision
If deep learning models with dilation kernels are implemented, then the measurement precision improves for feature recognition, but the device complexity increases due to computational requirements
Solution Approach 1:
The patent applies segmentation by dividing the pattern recognition task into distinct stages: aerial image generation, neural network processing, feature identification, and parameter measurement. The dilation kernels in the neural network further segment the feature extraction process, allowing parallel processing of different spatial scales and improving measurement precision for edge placement error and line edge roughness independently
Solution Approach 2:
The patent transitions from conventional 2D image processing to multi-dimensional feature space analysis using deep learning. The neural network with dilation kernels operates in expanded dimensional space, considering multiple spatial scales and feature hierarchies simultaneously, which enhances measurement precision while the computational complexity is managed through efficient network architecture design
3Manufacturing precision
If real-time process adjustments are made based on deep learning analysis, then the manufacturing precision improves, but the loss of time increases due to additional processing steps
Solution Approach 1:
The patent implements preliminary action by performing deep learning-based feature recognition and parameter measurement during or immediately after the aerial image generation phase, before the actual lithographic exposure. This allows process adjustments to be prepared in advance, reducing the time penalty of real-time corrections and improving manufacturing precision through proactive optimization
Solution Approach 2:
The patent establishes a feedback loop where deep learning analysis of aerial images provides real-time measurements of edge placement error and line edge roughness, which are then fed back to adjust lithographic process parameters. This closed-loop control system continuously improves manufacturing precision while minimizing time loss through efficient iterative optimization
Data Source
AI summary
A method for training a deep learning model of a patterning process. The method includes obtaining (i) training data comprising an input image of at least a part of a substrate having a plurality of features and a truth image, (ii) a set of classes, each class corresponding to a feature of the plurality of features of the substrate within the input image, and (iii) a deep learning model configured to receive the training data and the set of classes, generating a predicted image, by modeling and/or simulation of the deep learning model using the input image, assigning a class of the set of classes to a feature within the predicted image based on matching of the feature with a corresponding feature within the truth image, and generating, by modeling and/or simulation, a trained deep learning model by iteratively assigning weights using a loss function.