Learning-Based Defect Classifier for Semiconductor Wafer Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
As semiconductor device dimensions shrink, identifying nuisances and actual defects on wafers becomes increasingly challenging due to higher nuisance rates, which can lead to ineffective inspection processes and reduced yield, as existing methods struggle to accurately differentiate between nuisances and defects of interest, especially when the nuisance filter is not adequately representative of the defects encountered.
Innovation Solution
A system and method for training a learning-based defect classifier using a combination of inspection wafers with known high and normal defectivity, where defects of interest and nuisances are identified and combined to generate a training set, allowing the classifier to distinguish between them effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If inspection sensitivity is increased to detect smaller defects, then defect detection capability is improved, but nuisance rates increase dramatically
Solution Approach 1:
The system performs preliminary classification of detected defects into DOIs and nuisances using a learning-based classifier before final inspection decisions are made. This preliminary action allows the system to maintain high inspection sensitivity while filtering out nuisance signals that would otherwise contaminate the results.
Solution Approach 2:
A learning-based defect classifier is introduced as an intermediary component between the inspection system and the final defect determination. This classifier acts as a mediator that processes detected defects and distinguishes between DOIs and nuisances, allowing the inspection system to operate at high sensitivity without being overwhelmed by nuisance rates.
2Object-generated harmful factors
If nuisance filter parameters are adjusted to suppress nuisances, then nuisance rates are reduced, but detection of actual defects may be compromised
Solution Approach 1:
The system uses feedback from the learning-based classifier to continuously refine nuisance filter parameters. By analyzing classified results and adjusting filter parameters accordingly, the system can suppress nuisances while maintaining reliable detection of actual defects, as the classifier provides feedback on what constitutes a true defect versus a nuisance.
Solution Approach 2:
The system dynamically changes filter parameters based on the output from the learning-based classifier. Rather than using fixed parameters, the system adjusts filter characteristics according to the classified defect types, allowing optimal suppression of nuisances while preserving detection of actual defects across different inspection conditions.
3Device complexity
If training data is limited to single-wafer characteristics, then training process is simplified, but generalization to wafer-to-wafer variations is poor
Solution Approach 1:
The training system is designed to handle multiple wafer types simultaneously, making the training process universal rather than wafer-specific. By processing both high-defectivity and normal-defectivity wafers through the same training pipeline, the system generates a classifier that generalizes across wafer variations without requiring separate training processes for each wafer type.
Solution Approach 2:
The training dataset is constructed as a composite of defects from multiple wafer types with different defectivity characteristics. By combining high-defectivity wafer data and normal-defectivity wafer data into a single training set, the system creates a robust classifier that can handle wafer-to-wafer variations, analogous to creating composite materials with enhanced properties.
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach improves the accuracy of defect classification by considering wafer-to-wafer variations and provides a sufficient number of defect examples for training, reducing the adverse effects of nuisance filters and enhancing defect-related functions such as sampling and suppression.
Implementation Method 1
The energy source is configured to generate energy that is directed to wafers
Implementation Method 2
The detector is configured to detect energy from the wafers and to generate output responsive to the detected energy
Data Source
AI summary
Methods and systems for training a learning based defect classifier are provided. One method includes training a learning based defect classifier with a training set of defects that includes identified defects of interest (DOIs) and identified nuisances. The DOIs and nuisances in the training set include DOIs and nuisances identified on at least one training wafer and at least one inspection wafer. The at least one training wafer is known to have an abnormally high defectivity and the at least one inspection wafer is expected to have normal defectivity.