Semiconductor Process Simulation With Augmented Wafer Pattern Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing process simulation models for semiconductor manufacturing suffer from overfitting and lack robustness due to differences between design and actual patterns on wafers, necessitating improved simulation accuracy.
Innovation Solution
An electronic device transforms input images of design patterns into augmented images using factors that do not affect the manufacturing process, such as translation, rotation, and flipping, and trains a process simulation model using pairs of these images to reduce loss functions, thereby enhancing model robustness and preventing overfitting.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a process simulation model is trained using only the original design pattern images, then the model may achieve high accuracy on training data, but it suffers from overfitting and lacks robustness when predicting actual patterns on wafers
Solution Approach 1:
The patent applies data augmentation transformations (rotation, translation, flipping, scaling) to the training input images before model training. This preliminary action creates diverse training samples that simulate various possible actual patterns on wafers, enabling the model to learn more robust features and generalize better to unseen data, thereby preventing overfitting while maintaining accuracy
Solution Approach 2:
The patent changes the parameters of the input images by applying geometric transformations (rotation angles, translation distances, scaling factors) to generate augmented training data. These parameter variations expose the model to different pattern configurations, improving its robustness and ability to handle real-world variations in semiconductor manufacturing patterns
2Measurement precision
If the model is trained to minimize only the loss function based on difference from actual patterns, then prediction accuracy improves, but the model becomes sensitive to variations and loses stability
Solution Approach 1:
The patent merges two loss functions into a combined training objective: (1) the loss based on difference between predicted and actual patterns, and (2) the loss based on difference between transformed and original input images. This combination stabilizes training by providing an additional regularization term that penalizes excessive sensitivity to input variations, thereby improving model stability while maintaining prediction accuracy
Data Source
AI summary
Provided are an electronic device and a method of an operation thereof. The electronic device includes a memory and processing circuitry configured to execute the instructions stored in the memory to cause the electronic device to acquire a first input image based on an image of a design pattern to be formed on a wafer, transform the first input image into a second input image based on a factor not affecting a process of manufacturing a semiconductor device, and train a process simulation model by using a pair of the first input image and the second input image.