Virtual Metrology Activation Using Tool Event Prediction Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional virtual metrology systems fail to accurately predict tool processing events in real time, leading to user misjudgments and inefficient operations, as they often incorrectly generate predicted values for events like wafer chipping or tool processing actions.
Innovation Solution
A method involving historical process data analysis using classification algorithms to build models that predict tool processing events, including decision trees and reliance index models, which determine the reliability of predictions and activate virtual metrology only when events occur, thereby avoiding misjudgments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If conventional virtual metrology is applied to predict tool processing events, then real-time prediction capability is provided, but prediction accuracy deteriorates leading to user misjudgments
Solution Approach 1:
The patent segments the prediction system into multiple specialized classification models (e.g., Random Forest, XGBoost, LightGBM) that divide the prediction task into separate decision trees. Each tree handles specific aspects of the prediction, improving overall accuracy while maintaining real-time performance through parallel processing of segmented data features.
Solution Approach 2:
The patent transforms raw process data into optimized feature parameters through data preprocessing and feature engineering. By changing the representation of input data into meaningful parameters that better reflect underlying patterns, the system achieves higher prediction accuracy without sacrificing real-time processing capability.
2Reliability
If virtual metrology is activated for all workpieces, then comprehensive monitoring is achieved, but operational efficiency deteriorates due to unnecessary processing
Solution Approach 1:
The patent applies partial action by activating virtual metrology only for specific workpieces where prediction confidence exceeds a threshold or where processing events are detected. Instead of universally applying monitoring to all workpieces, the system selectively engages metrology functions based on real-time prediction outcomes, eliminating unnecessary processing while maintaining comprehensive monitoring coverage for critical cases.
Solution Approach 2:
The system implements feedback mechanisms where prediction results from classification models feed back into the decision-making process for activating virtual metrology. This closed-loop approach allows the system to learn from past predictions and adjust activation thresholds, ensuring reliable monitoring is applied only when truly necessary rather than universally.
3Measurement precision
If multiple classification models are used to improve prediction accuracy, then prediction reliability is improved, but system complexity increases
Solution Approach 1:
The patent merges multiple classification models (Random Forest, XGBoost, LightGBM) into a unified ensemble system that processes data through parallel pipelines. By combining the strengths of different algorithms and integrating their outputs through a coordinated architecture, the system achieves superior prediction accuracy while managing complexity through modular design and shared infrastructure.
Solution Approach 2:
The patent implements a universal prediction platform that can accommodate multiple classification models and algorithms within a single system framework. This multi-functional architecture allows different models to coexist and be selected or combined based on specific prediction needs, reducing overall system complexity compared to maintaining separate standalone systems for each model type.
Data Source
AI summary
Embodiments of the present disclosure provide a method for predicting an occurrence of a tool processing event, thereby determining whether to activate a virtual metrology. In a model-building stage, plural sets of model-building data are used to create at least one classification model in accordance with at least one classification algorithm, in which each classification model includes plural decision trees. Then, probabilities of the decision trees are used to create at least one reliance index model, and the sets of model-building data are used to create at least one similarity index model in accordance with a statistical distance algorithm. In a conjecture stage, a set of processing data of a workpiece is inputted into each classification model, each reliance index model and each similarity index model to determine whether to activate (start) virtual metrology.