Self-Supervised Neural Networks for OCD Data Interpretation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current optical and machine learning-based methods for measuring optical critical dimension (OCD) parameters in integrated circuit wafer patterns face challenges such as time-consuming and error-prone optical modeling, and the need for large datasets and expensive equipment in ML modeling, especially as pattern dimensions shrink.
Innovation Solution
The development of a self-supervised representation learning system that trains auto-encoder neural networks and transfer neural networks to estimate OCD parameters from scatterometric data, using a processor with instructions to divide scatterometric data into sub-vectors and train networks to map these sub-vectors, enabling accurate prediction of wafer pattern parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If optical modeling methods are used to measure OCD parameters, then measurement capability is provided, but the process becomes time-consuming and error-prone
Solution Approach 1:
The system performs preliminary action by pre-training auto-encoder neural networks on scatterometric data to learn optimal feature representations and patterns. This pre-processing and training phase enables the model to make rapid, accurate predictions during actual OCD measurements without requiring time-consuming optical modeling computations at measurement time.
Solution Approach 2:
The invention replaces the mechanical/optical modeling system with a machine learning-based neural network system. Instead of using traditional optical models that require complex computations to simulate light scattering, the system uses trained neural networks that have learned the relationships between scatterometric data and OCD parameters, significantly reducing computation time while maintaining or improving measurement accuracy.
2Productivity
If machine learning modeling is used to estimate OCD parameters, then measurement speed improves, but large datasets and expensive equipment are required
Solution Approach 1:
The scatterometric data is divided into multiple sub-vectors, and separate auto-encoder neural networks are trained for each sub-vector. This segmentation allows the system to process and learn from smaller, more manageable data portions independently, reducing the overall data requirement while maintaining comprehensive feature extraction capability across the complete dataset.
Solution Approach 2:
The auto-encoder neural networks perform self-supervised learning on the scatterometric data, meaning they learn to compress and reconstruct the data without requiring external supervision or large labeled datasets. The networks automatically learn optimal feature representations from the data structure itself, reducing dependency on expensive labeled training data and external equipment.
3Measurement precision
If scatterometric data is divided into sub-vectors and multiple auto-encoder networks are trained, then measurement accuracy improves, but system complexity increases
Solution Approach 1:
Multiple auto-encoder neural networks that process different sub-vectors of scatterometric data are merged into a single transfer neural network. This consolidation integrates the capabilities of multiple specialized networks into one unified system, maintaining the accuracy benefits of sub-vector processing while simplifying the overall system architecture and reducing operational complexity during actual measurements.
Data Source
AI summary
A system and methods for OCD metrology are provided including receiving multiple first sets of scatterometric data, dividing each set into k sub-vectors, and training, in a self-supervised manner, k2 auto-encoder neural networks that map each of the k sub-vectors to each other. Subsequently multiple respective sets of reference parameters and multiple corresponding second sets of scatterometric data are received and a transfer neural network (NN) is trained. Initial layers include a parallel arrangement of the k2 encoder neural networks. Target output of the transfer NN training is set to the multiple sets of reference parameters and feature input is set to the multiple corresponding second sets of scatterometric data, such that the transfer NN is trained to estimate new wafer pattern parameters from subsequently measured sets of scatterometric data.