Deep Learning Overlay Key Centering for Semiconductor Alignment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The semiconductor process faces limitations in accurately centering the overlay key due to variations in overlay key images caused by process effects, leading to potential Pattern Recognition (PR) failures and misalignment in the alignment state of fine patterns on semiconductor substrates.
Innovation Solution
A deep learning-based overlay key centering system that collects input data sets from overlay measurement devices, trains a model using deep learning algorithms to optimize weight modifications, and selects the best model based on precision, recall, and other metrics to precisely center the overlay key and prevent PR failures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional Pattern Recognition (PR) action is used to center the overlay key, then the system can operate with simple hardware, but measurement precision deteriorates due to severe variations in overlay key images caused by process effects
Solution Approach 1:
The patent transforms the overlay key image processing from traditional PR action to deep learning-based parameter extraction. The system uses a trained neural network model to extract center coordinates and bounding box parameters from overlay key images, replacing the 1:1 image matching approach. This parameter-based approach maintains reliability under severe process-induced variations while achieving precise centering.
Solution Approach 2:
The patent implements preliminary training of a deep learning model using a dataset of overlay key images with varying process conditions. This pre-trained model is then deployed for production use, allowing the system to handle variations without requiring real-time adaptation. The preliminary action of training ensures reliable performance across different process conditions.
2Manufacturing precision
If 1:1 image matching is used for Pattern Recognition, then the system can maintain simplicity, but manufacturing precision deteriorates when overlay key image variations are severe
Solution Approach 1:
The patent replaces the mechanical/image-matching-based PR action with a deep learning-based parameter extraction system. Instead of performing 1:1 image matching, the system uses a neural network model to directly extract center coordinates and bounding box parameters from images, achieving higher precision while managing complexity through efficient model architecture and training.
3Measurement precision
If deep learning model training is performed with extensive data, then measurement precision improves, but loss of time increases during the training phase
Solution Approach 1:
The patent employs data augmentation techniques to generate additional training samples from limited real data, applying transformations such as rotation, scaling, and noise addition. This partial action of synthesizing data achieves robust model training without requiring extensive manual data collection, reducing time loss while maintaining high measurement precision.
Data Source
AI summary
There is provided a deep learning-based overlay key centering system and a method thereof that may precisely measure and examine an alignment state of fine patterns of a semiconductor substrate. The method includes collecting an input data set from at least one device, the input data set comprising measurement image data of an overlay key and label data including information on a position and bounding box size of the overlay; and training the model by inputting the input data set to a model for deep learning. The step of training the model may include a step of calculating a loss function by comparing result data predicted by the model with the label data; and a step of optimizing an algorithm of the model by modifying a weight of the model so that a loss value calculated with the loss function may become smaller than a reference value.