Physics-Informed Kernel Learning for Lithographic Aerial Image Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional optical and resist modeling in lithography is time-intensive and computationally demanding, with artificial intelligence mechanisms for aerial image prediction exhibiting performance and accuracy limits that constrain their applicability to state-of-the-art lithography due to improper biasing and configuration.
Innovation Solution
A physics-informed kernel learning approach that transforms a mask image into light intensity feature maps, applies lithographic optical physics bias, and utilizes neural networks to generate high-quality aerial images, incorporating convolutional and transpose convolutional neural networks with a cost function to optimize the model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional optical and resist modeling is used for aerial image prediction, then manufacturing precision is maintained, but productivity deteriorates due to time-intensive and computationally demanding design optimization loops
Solution Approach 1:
The patent creates a trained neural network model that copies the predictive capabilities of conventional optical and resist modeling. The model is trained using aerial image data generated through rigorous optical modeling, then deployed as a lightweight surrogate that reproduces prediction accuracy without requiring repeated computational optimization loops, thus resolving the contradiction between precision and productivity
Solution Approach 2:
The patent performs preliminary training of the neural network model using extensive optical modeling data before deployment. This preliminary action pre-computes the complex optical and resist effects, allowing the model to make rapid predictions during actual lithography design without repeating the computationally intensive optimization processes, thereby improving productivity while maintaining precision
2Productivity
If conventional artificial intelligence mechanisms are used for aerial image prediction, then productivity is improved through faster computation, but manufacturing precision deteriorates due to performance and accuracy limits
Solution Approach 1:
The patent introduces lithographic optical physics bias as an intermediary component between the neural network and the aerial image prediction task. This bias encodes domain-specific optical and resist knowledge into the model, enabling it to achieve state-of-the-art prediction accuracy while maintaining the computational efficiency of neural network inference, thus resolving the contradiction between productivity and precision
Solution Approach 2:
The patent modifies the neural network by incorporating lithographic optical physics bias through parameter adjustments in the model architecture and training process. This changes the model's parameters to reflect physical realities of lithography, improving prediction accuracy without sacrificing the computational speed advantages of AI mechanisms
3Manufacturing precision
If state-of-the-art lithography processes are pursued, then manufacturing precision is improved through smaller dimensions and higher densities, but device complexity increases due to improper biasing and configuration requirements
Solution Approach 1:
The patent creates a universal neural network model with lithographic optical physics bias that can handle multiple lithography prediction tasks through a single unified architecture. This multi-functional model eliminates the need for separate configurations for different prediction scenarios, reducing device complexity while maintaining the precision required for state-of-the-art lithography scaling
Data Source
AI summary
A lithographic aerial image generator configured to transform a mask image of a circuit into light intensity feature maps, to bias the light intensity feature maps with lithographic optical physics weights to generate a coarse aerial image, and to transform the coarse aerial image through a first neural network to generate a lithographic-quality areal image.