Multi-task Neural Network for Lithography and Etch CD Modeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional semiconductor manufacturing processes face challenges in accurately modeling post-lithography and post-etch critical dimensions, which are crucial for integrated circuit design, as existing methods lack the precision and efficiency in simulating these dimensions simultaneously.
Innovation Solution
The use of multi-task neural networks that generate lithographic aerial images as training data, constructing a shared and customized neural network architecture to output simulated critical dimension values for both lithography and etch processes, allowing for concurrent modeling of post-lithography and post-etch critical dimensions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If separate modeling methods are used for post-lithography and post-etch critical dimensions, then each process can be modeled independently, but the overall simulation accuracy and efficiency deteriorate due to lack of coordination between the two processes
Solution Approach 1:
The patent combines separate post-lithography and post-etch critical dimension modeling into a unified multi-task neural network framework. The shared backbone network processes input data once and generates features that are then used by both lithography and etch prediction heads, eliminating redundant computations while maintaining accurate predictions for both processes simultaneously.
Solution Approach 2:
The neural network employs a universal shared backbone that serves multiple functions: extracting features for both lithography CD prediction and etch CD prediction. This multi-functional architecture allows a single model to handle both critical dimension modeling tasks, improving efficiency while maintaining reliability through coordinated learning.
2Ease of manufacture
If conventional separate modeling approaches are used for lithography and etch processes, then implementation simplicity is maintained, but simulation accuracy deteriorates due to inability to capture process interactions
Solution Approach 1:
The unified model is segmented into distinct functional components: a shared backbone network for feature extraction, a lithography prediction head, and an etch prediction head. This segmentation allows the complex multi-task model to be implemented systematically while capturing process interactions, balancing implementation feasibility with prediction accuracy.
3Productivity
If multi-task neural networks with shared architecture are used, then simulation efficiency and accuracy are improved, but model complexity increases
Solution Approach 1:
The patent merges the computational pathways for lithography and etch modeling into a single shared backbone network, reducing overall model complexity compared to maintaining entirely separate models. The shared features and parameters are reused by both prediction tasks, improving efficiency while managing complexity through resource sharing.
Data Source
AI summary
A method is presented for constructing a deep neural network based model to concurrently simulate post-lithography critical dimensions (CDs) and post-etch critical dimensions (CDs) and to improve the modeling accuracy of each process respectively. The method includes generating lithographic aerial images of physical design layout patterns, constructing a multi-task neural network including two output channels, training the multi-task neural network with the training data of the lithographic aerial images, and outputting simulated critical dimension values pertaining to lithography and etch processes.