Semiconductor Defect Correlation Algorithm
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for correlating inspection layer defects with die faults in semiconductor manufacturing fail to utilize statistics and confidence analysis, leading to ineffective identification of defects causing die failure, as they do not distinguish between random and correlated defect distributions.
Innovation Solution
An inspection layer correlation algorithm that classifies die based on defects and testing results, using a hypergeometric distribution to calculate probabilities and confidence factors, adjusts for random defect occurrences, and identifies defects causing die failure with high confidence, thereby improving defect analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional defect analysis techniques are used to identify defects causing die failure, then defect identification is performed, but statistical confidence is not established to distinguish between random and correlated defect distributions
Solution Approach 1:
The patent implements feedback by using die test results to inform and refine defect analysis. The system correlates defect data with pass/fail test data, using the test outcomes as feedback to identify which defects are truly dispositive. This closed-loop approach allows the system to continuously improve defect identification accuracy by learning from actual die performance data.
Solution Approach 2:
The patent introduces statistical confidence metrics as an intermediary between raw defect data and defect classification. Rather than directly classifying defects as causing or non-causing, the system uses statistical confidence levels as a mediating layer to evaluate the strength of correlation between defects and die failures. This intermediary enables nuanced differentiation between random and correlated defect distributions.
2Quantity of substance
If all identified defects are treated as potential causes of die failure, then comprehensive defect coverage is achieved, but resource-intensive analysis of non-dispositive defects increases processing time and cost
Solution Approach 1:
The patent extracts and isolates only the dispositive defects from the complete set of identified defects. By using statistical confidence metrics and correlation analysis with die test results, the system separates the small subset of defects that actually cause die failures from the larger population of benign defects. This extraction focuses analytical resources on the critical minority of defects that matter.
Solution Approach 2:
The patent applies local quality by treating different defects with different levels of analytical scrutiny based on their statistical confidence scores. Rather than uniformly analyzing all defects, the system concentrates detailed analysis on defects with high confidence of being dispositive, while accepting lower-confidence defects with less intensive analysis. This differentiated approach optimizes resource allocation across the defect population.
3Measurement precision
If statistical methods are implemented to calculate confidence factors for defect correlation, then accurate identification of dispositive defects is achieved, but computational complexity and algorithm sophistication increase
Solution Approach 1:
The patent employs computationally efficient statistical methods that provide sufficient confidence metrics without requiring complex, resource-intensive algorithms. The system uses straightforward statistical calculations based on defect frequency and die failure correlation rather than sophisticated machine learning models. This approach achieves adequate measurement precision with minimal computational overhead, treating complex analysis as a disposable resource used only when necessary.
Data Source
AI summary
The defect-containing die identified from an inspection layer analysis subsequent to a manufacturing step for a wafer including a plurality of die and as well as the faulty die identified from a fault testing of the wafer are processed to identify a subset of the die that both contain a defect and are faulty. A probability analysis is performed to determine a confidence level of whether the die in the subset are faulty due to their defects.