Virtual Metrology Control for Drift-Adaptive Process Monitoring
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current virtual metrology systems struggle to accurately model data drifts, data shifts, and hierarchical data structures in manufacturing processes, leading to decreased prediction accuracy and model failures, especially in semiconductor manufacturing where equipment aging and maintenance cause gradual and abrupt changes in process dynamics.
Innovation Solution
The system employs a data processing module that generates training data from diverse sources, a training and optimization module using machine learning pipelines with feature engineering, time-aware data normalization, and adaptive learning algorithms, and an inference module for real-time prediction and process control, capable of detecting and mitigating drifts and shifts in process equipment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional virtual metrology systems use historical data and static models, then initial prediction accuracy is achieved, but prediction accuracy deteriorates over time due to data drifts and equipment aging
Solution Approach 1:
The system transitions from static historical models to dynamic adaptive models that continuously learn from incoming data streams. The machine learning models are updated in real-time to adapt to changing process conditions, equipment aging, and data drifts, ensuring sustained prediction accuracy without requiring complete model retraining.
Solution Approach 2:
The system implements continuous feedback loops where prediction residuals and actual measurements are fed back into the model training process. This feedback mechanism enables the system to detect drifts and automatically adjust model parameters, maintaining reliability over time by correcting deviations from expected behavior.
2Reliability
If comprehensive data from multiple sources is collected for training, then model robustness improves, but data processing complexity increases
Solution Approach 1:
The system segments the complex multi-source data into structured hierarchical components (process parameters, equipment data, environmental conditions). This segmentation allows modular processing where each data type is handled by specialized preprocessing routines, reducing overall complexity while maintaining comprehensive data utilization for robust model training.
Solution Approach 2:
The system implements a universal data processing framework that handles multiple data sources through standardized interfaces and common preprocessing pipelines. This multi-functional approach consolidates what would otherwise require separate processing systems, reducing complexity while maintaining the ability to ingest and process diverse data types uniformly.
3Measurement precision
If continuous real-time model updating is implemented, then prediction performance is maintained, but computational resource requirements increase
Solution Approach 1:
The system applies partial updating strategies where only critical model parameters are continuously updated in real-time, while less sensitive parameters are updated periodically or offline. This selective approach maintains prediction performance for critical process variables while reducing overall computational burden compared to complete model retraining at every update cycle.
Solution Approach 2:
The system performs preliminary data preprocessing, feature extraction, and candidate model selection during off-peak periods or in parallel processing streams before real-time prediction is needed. This preliminary action reduces the computational load during critical real-time updating, allowing sustained prediction performance with lower instantaneous resource requirements.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Described are systems and methods for advanced process control and monitoring. Systems and methods may be associated with a data processing module configured to receive and process a plurality of data types and datasets from a plurality of different sources for generating training data; a training and optimization module configured to provide the training data to a machine learning pipeline for training and optimizing a model; and an inference module configured to use the model for generating one or more predicted metrics substantially in real-time, wherein the one or more predicted metrics are useable to characterize an output of a process performed by a process equipment.