Real-Time Regularization Control for ML Measurement Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning based measurement models for semiconductor metrology are suboptimal due to lack of regularization, leading to large modeling errors and lack of robustness to changing process conditions, requiring costly and time-consuming retraining.
Innovation Solution
Training a machine learning based measurement model with an ensemble of regularization control values to adapt to varying conditions, allowing the model to capture a broader range of process variations and automatically update the regularization control value based on measurement performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If machine learning based measurement models are trained without regularization, then training is simpler and faster, but modeling errors become large and model robustness deteriorates
Solution Approach 1:
The patent applies parameter changes by introducing regularization control values as hyperparameters that modify the training objective. Instead of using a fixed loss function, the system varies the regularization strength through control values to balance between training speed and model robustness, resolving the contradiction between fast training and reliable predictions.
Solution Approach 2:
The system dynamically adjusts regularization control values during the training process rather than using a static approach. This dynamic adaptation allows the model to learn from data efficiently while maintaining robustness, addressing both the time loss and reliability concerns simultaneously.
2Reliability
If multiple machine learning based measurement models are trained with different regularization control values, then model robustness improves, but computational effort and training time increase
Solution Approach 1:
The patent segments the regularization control values into an ensemble of discrete values rather than training completely separate models. This segmentation approach allows the system to explore multiple regularization strengths without duplicating the entire model training process, reducing computational effort while maintaining robustness.
Solution Approach 2:
A single measurement model is designed to handle multiple regularization conditions through a universal training approach that incorporates ensemble regularization control values. This multi-functional design eliminates the need for separate models for different regularization scenarios, significantly reducing training time while preserving model robustness across varying conditions.
3Measurement precision
If measurement models are retrained when process conditions change, then measurement accuracy is maintained, but productivity decreases due to repeated training
Solution Approach 1:
The system performs preliminary action by pre-computing and storing an ensemble of regularization control values during the initial training phase. When process conditions change, the system can quickly select appropriate pre-computed regularization values without retraining, maintaining measurement accuracy while avoiding the productivity loss associated with repeated training.
Solution Approach 2:
The patent implements feedback mechanisms that monitor measurement performance and process conditions. When drift is detected, the system uses feedback to select appropriate regularization control values from the pre-computed ensemble, maintaining accuracy without triggering costly retraining cycles, thus preserving productivity.
4Device complexity
If total measurement uncertainty is minimized without regularization, then training objective is simpler, but measurement performance becomes suboptimal due to large modeling errors
Solution Approach 1:
The patent changes the training objective by introducing regularization control values as additional parameters. This modification allows the system to balance training complexity with measurement performance, achieving suboptimal results without the need for overly complex training procedures by using controlled regularization instead.
Data Source
AI summary
Methods and systems for training a machine learning based measurement model conditioned by at least one regularization control parameter are described herein. A ML based measurement model conditioned by at least one regularization control parameter is trained for different control parameter values. A regularization control value provided as input to the trained ML based measurement model defines the regularization condition at inference. In a further aspect, an optimal value of a regularization control parameter is selected based on measurement performance on a set of measurement data. As measurement conditions change, the optimal value is reevaluated based on measurement performance on an updated set of measurement data that reflects the changing measurement conditions. In another further aspect, changes in measurement conditions and reevaluation of a regularization control value are performed automatically as measurement data is collected by a measurement system without interruption of the measurement process.