Parallel Metrology Models Using Mueller Signals for Process Robustness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current optical metrology techniques for integrated circuit manufacturing face challenges with complex spectroscopic analysis, requiring extensive computational resources and lacking robustness to process variations, and single-machine learning models suffer from limited accuracy and robustness due to undesirable correlations and a limited set of analyzed signals.
Innovation Solution
A metrology system utilizing multiple machine learning models, each trained on different measurement configurations, generates intermediate metrology measurements from measurement datasets, which are combined using a weighting model to produce a final measurement, allowing for efficient and robust metrology.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional RCWA is used for optical metrology, then computational solutions can be obtained, but the time to solution is long and computational resources are extensive
Solution Approach 1:
The patent pre-computes and stores calibration data for multiple process conditions before actual metrology measurements. This preliminary action creates a lookup table of pre-analyzed spectra that can be quickly compared against measured data, eliminating the need for time-consuming RCWA computations during actual measurements while maintaining measurement accuracy.
Solution Approach 2:
The patent creates simplified models or surrogate models that replicate the behavior of complex RCWA simulations. These copied models are trained on pre-computed RCWA data and can predict measurement outcomes much faster, providing a computationally efficient alternative to running full RCWA simulations for every measurement.
2Measurement precision
If conventional RCWA is used for optical metrology, then computational solutions can be obtained, but extensive computational resources are required
Solution Approach 1:
The patent performs computationally intensive RCWA simulations and data analysis in advance, storing the results in calibration datasets. This shifts the computational burden from the measurement phase to the calibration phase, allowing actual measurements to be processed with minimal computational resources while maintaining high accuracy through comparison with pre-computed reference data.
Solution Approach 2:
The patent creates lightweight surrogate models that copy the essential predictive capabilities of full RCWA simulations. These simplified models require fraction of the computational resources to run, enabling rapid analysis while preserving the accuracy benefits of rigorous electromagnetic modeling.
3Productivity
If a single machine learning model is used to generate metrology measurements, then the model can be trained on measurement signals, but accuracy and robustness to process variations are limited
Solution Approach 1:
The patent divides the single machine learning model into multiple specialized models, each trained on specific subsets of measurement signals or specific process conditions. This segmentation allows each model to specialize in particular aspects of the measurement space, improving overall accuracy and robustness by combining the strengths of multiple specialized models rather than relying on one general-purpose model.
Solution Approach 2:
The patent trains different machine learning models on different subsets of training data corresponding to specific process conditions or signal characteristics. Each model develops local expertise in its designated domain, and the system selectively applies the appropriate model based on the measurement conditions, thereby improving accuracy across diverse process variations.
4Productivity
If a single machine learning model is used, then the model can process measurement signals, but robustness to process variations is limited
Solution Approach 1:
The patent segments the processing system into multiple machine learning models, each trained on specific process condition subsets. This segmentation enables the system to maintain specialized knowledge for different process scenarios, improving robustness to variations by selecting the appropriate pre-trained model for each measurement rather than relying on a single general model.
Solution Approach 2:
The patent varies training parameters such as the subsets of measurement signals, process conditions, and data distributions used to train different models. By changing these parameters across multiple models, the system captures a broader range of process variations, thereby improving robustness while maintaining efficient processing through model specialization.
Data Source
AI summary
A metrology system may receive two or more measurement datasets associated with a test feature on a sample from one or more measurement sub-systems operable under two or more measurement configurations, where a respective one the two or more measurement datasets is generated with a respective one of the two or more measurement configurations. A metrology system may generate two or more intermediate metrology measurements of the test feature using two or more machine learning models, where a respective one of the two or more intermediate metrology measurements is generated using at least a portion of a respective one of the two or more measurement datasets as an input to a respective one of the two or more machine learning models. A metrology system may determine a final metrology measurement of the test feature using a weighting model based on the two or more intermediate metrology measurements.