Metrology Model Updating for Process Drift and Inference Consistency
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Lithographic processes face challenges in accurately measuring and maintaining overlay accuracy due to process drift, leading to inconsistent model performance in inferring parameters of interest, which compromises Advanced Process Correction strategies.
Innovation Solution
A method for training a second model by optimizing a cost function while constraining it to maintain consistency with a first model, using a second set of measurement signals from a drifted process range, ensuring that inferred values for a parameter of interest do not differ by more than a threshold margin from the first model's values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If models are periodically retrained to adapt to process drift, then model accuracy for drifted processes is improved, but consistency with historical model inferences deteriorates
Solution Approach 1:
The patent implements a dynamic model training approach where the second model is trained on extended process ranges including drifted conditions, while applying a consistency constraint that adapts the model's behavior. This allows the model to dynamically adjust to process drift while maintaining stability for normal operating conditions through the threshold-based constraint mechanism.
Solution Approach 2:
The patent changes the training parameters by extending the process range from the first process range to a second process range that includes drifted conditions.同时, it introduces a new parameter - the threshold margin - that controls the degree of consistency required between first and second model inferences, allowing flexible adjustment between adaptability and consistency.
2Adaptability or versatility
If the second model is trained on extended process range data, then adaptability to drifted processes is improved, but deviation from first model inferences increases
Solution Approach 1:
The patent applies preliminary action by pre-defining the threshold margin before training the second model. This threshold serves as a prior constraint that guides the training process to maintain consistency with the first model's inferences while adapting to extended process ranges, preventing excessive deviation before it occurs.
Solution Approach 2:
The patent implements feedback through the consistency constraint mechanism that compares second model inferences against first model inferences for overlapping process conditions. The threshold margin acts as a feedback boundary that guides the training process to adjust the second model's parameters when deviations exceed the acceptable margin.
3Stability of the object's composition
If constraint on inference difference is applied, then backward consistency is improved, but model flexibility to capture process drift is reduced
Solution Approach 1:
The patent applies partial action by implementing the consistency constraint only for process conditions that fall within both the first and second process ranges. For conditions unique to the extended second process range, the model has full flexibility to adapt without constraint, thus maintaining both consistency where applicable and adaptability where needed.
Data Source
AI summary
Disclosed is a method of updating of a first model by training a second model, the first model relating to a first process range and trained using a first set of measurement signals relating to a first set of structures. The method comprises: obtaining a second set of measurement signals, the second set of measurement signals relating to a second set of structures comprising said first set of structures or a subset thereof; and training the second model using said second set of measurement signals and corresponding reference values for the parameter of interest as training data. The training comprises optimizing a cost function in terms of the second model while constraining the second model to infer values for the parameter of interest from the first set of measurement signals.