Metrology Apparatus Calibration Using Adversarial Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current metrology apparatuses face challenges in achieving consistent tool-to-tool matching, especially for sophisticated measurement modes and small targets, due to the difficulty in distinguishing between target and apparatus influences on optical signals.
Innovation Solution
A method involving machine learning processes to calibrate multiple metrology apparatuses using training data, where encoders and decoders are trained to minimize differences between detected and synthetic representations, and a classifier is trained to identify the originating apparatus, optimizing fidelity and confusion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional calibration methods are used for metrology apparatuses, then the measurement process is simple, but tool-to-tool matching precision deteriorates due to inability to distinguish target and apparatus influences
Solution Approach 1:
The patent segments the calibration process into two distinct machine learning tasks: an encoder-decoder model that learns to remove apparatus-specific influences from detected representations, and a classifier that identifies the originating apparatus. This segmentation allows the system to separately optimize for measurement consistency across tools while maintaining the ability to identify tool-specific characteristics when needed.
Solution Approach 2:
The patent introduces an intermediate encoded representation that serves as a mediator between the raw detected representation and the final measurement parameters. This encoded representation contains the target information while having apparatus-specific influences minimized, effectively acting as a bridge that separates target properties from measurement system characteristics.
2Measurement precision
If sophisticated measurement modes are used with small targets, then measurement capability is improved, but tool-to-tool matching becomes increasingly difficult due to low sensitivity and correlated responses
Solution Approach 1:
The patent applies preliminary action by training the encoder-decoder model on diverse training data that includes various targets, measurement modes, and apparatus configurations before actual measurements are taken. This pre-training enables the model to learn robust representations that are insensitive to apparatus-specific variations, preparing the system in advance for challenging measurement scenarios with small targets and low sensitivity.
Solution Approach 2:
The patent utilizes parameter changes by transforming the detected representation into an encoded representation through the encoder, which changes the parameter space from raw optical signals to a transformed domain where apparatus-specific influences are minimized. This parameter transformation enables better separation of target properties from measurement system characteristics.
3Measurement precision
If the encoder and decoder are trained to minimize differences between detected and synthetic representations, then measurement fidelity is improved, but the ability to identify apparatus-specific characteristics deteriorates
Solution Approach 1:
The patent implements feedback by using the classifier's apparatus identification predictions as a guiding signal during encoder-decoder training. The classifier provides feedback about which apparatus-specific characteristics remain in the encoded representation, allowing the encoder to selectively minimize only those influences that harm measurement fidelity while preserving characteristics that are useful for apparatus identification when needed.
Solution Approach 2:
The patent applies local quality by making different parts of the encoded representation serve different functions: some components are optimized to minimize apparatus-specific influences for high-fidelity measurements, while other components retain apparatus-specific characteristics that the classifier can use for identification. This local differentiation allows simultaneous optimization of both measurement fidelity and apparatus identifiability.
Data Source
AI summary
Methods for calibrating metrology apparatuses and determining a parameter of interest are disclosed. In one arrangement, training data is provided that comprises detected representations of scattered radiation detected by each of plural metrology apparatuses. An encoder encodes each detected representation to provide an encoded representation, and a decoder generates a synthetic detected representation from the respective encoded representation. A classifier estimates from which metrology apparatus originates each encoded representation or each synthetic detected representation. The training data is used to simultaneously perform, in an adversarial relationship relative to each other, a first machine learning process involving the encoder or decoder and a second machine learning process involving the classifier.