Composite Metrology Data for Robust ML Device Parameter Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning-based metrology solutions for characterizing device structures face challenges in robustness due to limited reference data, especially when process variations occur, leading to inefficient re-training processes that are costly and time-consuming, and struggle with measuring key parameters like overlay with low sensitivity.
Innovation Solution
The use of composite metrology data, generated by merging measured data from a reference device with synthetic data calculated from models of modified reference devices, expands the training parameter space and improves robustness by incorporating variations and correlations between key and non-key parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If machine learning models are trained using only measured reference metrology data, then the models can accurately characterize device structures under normal process conditions, but the models lack robustness when process variations occur and require costly and time-consuming re-training
Solution Approach 1:
The patent applies preliminary action by generating synthetic metrology data that anticipates potential process variations before they occur in actual manufacturing. This synthetic data is created in advance and merged with measured reference data to pre-train the machine learning model, enabling it to handle process variations without requiring re-training when deviations occur.
Solution Approach 2:
The patent employs parameter changes by systematically varying model parameters to generate diverse synthetic metrology data representing different process conditions. By training the machine learning model on this varied synthetic data alongside measured data, the model learns to tolerate a broader range of process variations, improving robustness without additional re-training costs.
2Adaptability or versatility
If additional reference metrology data is obtained and the machine learning model is re-trained in response to process revisions, then the model can adapt to new process conditions, but this process is inefficient and prohibitively expensive
Solution Approach 1:
The patent applies copying by creating synthetic metrology data that replicates the characteristics of actual measured data under various process conditions. Instead of obtaining additional expensive measured reference data for each process revision, the system generates synthetic copies that capture the essential variations, enabling model adaptation at minimal cost.
Solution Approach 2:
The patent implements universality by creating a synthetic data generation system that can produce training data for multiple different process conditions and revisions simultaneously. This universal approach allows a single trained model to handle various process variations without requiring separate re-training campaigns for each specific revision.
3Productivity
If machine learning models are trained with limited reference data, then the training process is faster and cheaper, but the models struggle to accurately measure key parameters with low sensitivity such as overlay
Solution Approach 1:
The patent applies merging by combining measured reference metrology data with synthetic metrology data to create a comprehensive training dataset. This merged dataset retains the authenticity of measured data while incorporating the diversity and volume of synthetic data, enabling the model to achieve high measurement precision for low-sensitivity parameters without sacrificing training efficiency.
Data Source
AI summary
A machine learning model that uses composite metrology data determines at least one parameter of a device under test using measured metrology data from the device. The composite metrology data is generated by merging measured metrology data from a reference device with synthetic metrology data calculated from a model of the reference device. The composite metrology data may be generated further based on a synthetic metrology data calculated from a model for a modified reference device. The modified reference device may be generated using variations of at least one parameter of the model to expand the parameter space of the training range.