Automatic ML Retraining for Wafer Metrology Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning (ML) models in semiconductor metrology suffer from decreased accuracy and reliability due to complex process variations and measurement tool states, leading to costly manual retraining and potential missed detection of flier wafers.
Innovation Solution
Implement an automated machine learning model retraining mechanism using statistical and unsupervised algorithms to detect shifts in measurement distributions, initiating data collection and model updates only when necessary, thereby reducing unnecessary retraining and improving throughput.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual retraining of ML models is implemented when predictions are unsatisfactory, then model accuracy is maintained, but human and tool time costs increase significantly
Solution Approach 1:
The system implements automated model validation and retraining triggering without human intervention. The validation module automatically assesses model performance using measurement data, and when degradation is detected, the system autonomously initiates retraining with updated data, eliminating the need for manual monitoring and intervention while maintaining high prediction accuracy
Solution Approach 2:
The system establishes a closed-loop feedback mechanism where model predictions are continuously validated against new measurement data. When validation metrics indicate model degradation, the system automatically triggers retraining. This feedback-driven approach ensures model accuracy is maintained while minimizing unnecessary retraining operations that waste time and resources
2Adaptability or versatility
If time-based periodic triggering of ML retraining is implemented, then model predictions adapt to slowly varying process shifts, but unnecessary data collection and retraining occur even when not needed
Solution Approach 1:
The system transitions from static periodic retraining to dynamic condition-based retraining. Instead of retraining at fixed intervals, the system continuously monitors model validation metrics and dynamically triggers retraining only when actual model degradation is detected. This dynamic approach maintains model adaptability to process shifts while avoiding unnecessary retraining operations that reduce tool throughput
Solution Approach 2:
The system changes the triggering parameter from fixed time intervals to variable validation-based conditions. Model retraining is triggered based on changes in validation metrics (such as prediction residuals or statistical tests) rather than elapsed time. This parameter change enables the system to adapt to process shifts when needed while maintaining high productivity by avoiding retraining when the model remains valid
3Reliability
If time-based periodic triggering of ML retraining is implemented, then model predictions are updated regularly, but flier wafers may be missed causing false negatives
Solution Approach 1:
The system implements continuous model validation between retraining cycles by monitoring measurement data against the current model predictions. This continuous validation ensures that flier wafers are detected immediately when they occur, rather than waiting for the next periodic retraining cycle. The useful action of model validation continues uninterrupted, maintaining high detection reliability without requiring continuous full retraining cycles
Solution Approach 2:
The system performs preliminary validation checks on incoming measurement data before full retraining is initiated. When anomalies or flier wafers are detected in these preliminary checks, the system can immediately flag them for review or trigger targeted retraining. This preliminary action enables early detection of invalid predictions while minimizing the time loss associated with complete retraining cycles
Data Source
AI summary
A metrology method with automated triggering of retraining of a machine learning model (MLM) is disclosed. The method may acquire metrology measurement data from a plurality of sites of a wafer. The method may apply a MLM to the measurement data to predict a metrology metric. The method may apply a triggering algorithm to monitor the effectiveness of the MLM, wherein the triggering algorithm determines a dissimilarity between the measurement data and a training data set of the MLM. The triggering algorithm may identify a failed MLM state when the distance between the measurement data and the training data set exceeds one or more thresholds. The method may retrain the MLM using an adjusted training data set. The adjusted training data set may be generated by adding the measurement data from the wafer to the training data set. The method may apply the retrained MLM to a subsequent wafer.