Defect Location Prediction With Active Learning for Wafer Inspection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing inspection methods for semiconductor substrates are inaccurate and resource-intensive, often missing defective locations due to low confidence predictions, and lack self-repairing capabilities, leading to inefficient defect detection in integrated circuits.
Innovation Solution
A particle beam inspection apparatus using a defect location prediction model that assigns confidence scores to inspection locations and incrementally trains on actual inspection results to improve prediction accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional inspection methods are used to detect defects on semiconductor substrates, then comprehensive coverage can be achieved, but resource consumption increases and accuracy decreases
Solution Approach 1:
The defect location prediction model performs preliminary analysis on the substrate to identify high-probability defect locations before actual inspection. By predicting likely defect positions in advance and ranking them by confidence score, the system directs inspection resources only to the most promising locations, avoiding exhaustive inspection of the entire substrate and thereby reducing resource consumption while maintaining or improving detection accuracy.
2Measurement precision
If inspection coverage is increased to improve defect detection, then accuracy improves, but throughput decreases
Solution Approach 1:
The system applies local quality by concentrating inspection resources on specific high-probability defect locations identified by the prediction model, rather than applying uniform inspection across the entire substrate. The confidence score ranking enables the system to prioritize inspection of locations with highest defect probability, achieving high detection accuracy for critical areas while maintaining overall throughput by skipping low-probability regions.
3Measurement precision
If the prediction model is trained more extensively to improve accuracy, then prediction precision improves, but training time and computational resources increase
Solution Approach 1:
The system employs partial action by training the prediction model on a strategically selected subset of training data that is most relevant to the specific substrate and defect types being inspected. Rather than exhaustively training on all possible data, the model focuses on pertinent patterns and features, achieving sufficient prediction accuracy for practical inspection purposes while minimizing training time and computational resource requirements.
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
The method effectively identifies high-confidence defect locations, reducing resource consumption and improving yield by refining predictions through incremental training with actual inspection data.
Implementation Method 1
a particle beam inspection apparatus using a plurality of charged particle beams
Data Source
AI summary
A method and apparatus for identifying locations to be inspected on a substrate is disclosed. A defect location prediction model is trained using a training dataset associated with other substrates to generate a prediction of defect or non-defect and a confidence score associated with the prediction for each of the locations based on process-related data associated with the substrates. Those of the locations determined by the defect location prediction model as having confidences scores satisfying a confidence threshold are added to a set of locations to be inspected by an inspection system. After the set of locations are inspected, the inspection results data is obtained, and the defect location prediction model is incrementally trained by using the inspection results data and process-related data for the set of locations as training data.