Metrology Signal Reliability Monitoring via Inference Model Feedback
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing metrology methods in lithographic processes face challenges in reliably monitoring and assessing the accuracy of overlay measurements due to process variations and drifts, leading to potential inaccuracies in inferring critical parameters like overlay values.
Innovation Solution
A method and system that determine a reliability metric for metrology signals and parameter values by using a trained inference model to assess the similarity of measurement signals to those in a training data set, incorporating dual pupil techniques to account for process variations and optimize weighting for accurate overlay inference.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a trained inference model is used to infer parameter values from metrology signals, then measurement precision is improved, but reliability deteriorates due to process variations and drifts outside the training data space
Solution Approach 1:
The system implements feedback by calculating a reliability metric that compares the measurement signal against the training data distribution, and uses this feedback to determine whether the inference result is reliable or if the model needs retraining. This closed-loop approach ensures that the system adapts to process variations by triggering retraining when reliability drops below a threshold.
Solution Approach 2:
The system performs preliminary action by pre-calculating the distribution characteristics of the training data (mean and covariance matrix) and storing them for later comparison. This preparation enables rapid reliability assessment during actual measurements without requiring access to the entire training dataset, allowing quick detection of out-of-distribution signals.
2Reliability
If metrology recipes are frequently updated to adapt to process variations, then reliability is improved, but productivity deteriorates due to additional measurement and retraining time
Solution Approach 1:
The system applies partial action by performing reliability monitoring on a subset of measurements rather than requiring full retraining after every process variation. The reliability metric provides a continuous indication of model performance, allowing the system to maintain operational status with partial confidence rather than requiring complete retraining cycles, thus balancing reliability with productivity.
3Measurement precision
If the measurement signal is heavily weighted to emphasize certain features, then measurement precision is improved, but adaptability deteriorates when process conditions change from training conditions
Solution Approach 1:
The system implements dynamics by making the weighting adaptive rather than static. The reliability metric dynamically adjusts the effective weighting based on how well the current measurement signal matches the training data distribution. When signals are within the training space, the original weights apply; when they fall outside, the system reduces confidence and triggers retraining, effectively adapting the weighting strategy to current process conditions.
Data Source
AI summary
Disclosed is a method of determining a reliability metric describing a reliability of metrology signal and/or a parameter of interest value derived therefrom and associated apparatuses. The method comprises obtaining a trained inference model for inferring a value for a parameter of interest from a measurement signal and one or more measurement signals and/or respective one or more values of a parameter of interest derived therefrom using said trained inference model. At least one reliability metric value is determined for the one or more measurement signals and/or respective one or more values of a parameter of interest, the reliability metric describing a reliability of one or more measurement signals and/or respective one or more values of a parameter of interest with respect to an accurate prediction space associated with the trained inference model.