Inspection Result Analysis Using Non-Negative Matrix Factorization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In the semiconductor industry, inspecting substrates for large-area failures is challenging due to the complexity of identifying the cause of failures and classifying substrates with similar failure causes.
Innovation Solution
An inspection result analysis device that uses non-negative matrix factorization (NMF) to factorize inspection data from substrates into feature maps and weight groups, allowing for accurate identification and classification of features associated with substrate failures.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If substrates are subjected to inspection and classification by features, then failure causes can be identified and substrates with similar failure causes can be grouped, but the complexity of identifying and classifying failure features increases
Solution Approach 1:
The inspection result analysis device segments the complex task of failure analysis into distinct functional modules: an extraction unit that identifies candidate features from inspection results, a selection unit that chooses representative features, and a classification unit that groups substrates. This segmentation reduces the overall complexity by breaking down the monolithic analysis process into manageable, specialized components.
Solution Approach 2:
The device introduces an intermediary feature extraction and selection mechanism between the raw inspection results and the final classification. The extraction unit identifies candidate features, and the selection unit filters these candidates to select representative features, acting as intermediaries that simplify the relationship between complex inspection data and classification outcomes.
2Measurement precision
If multiple features are extracted from inspection data to improve classification accuracy, then the precision of failure identification increases, but the computational complexity and processing time increase
Solution Approach 1:
The selection unit implements partial action by not processing all extracted candidate features, but only selecting a representative subset. This partial processing approach maintains sufficient classification precision while significantly reducing the computational burden and processing time compared to analyzing all possible features.
Solution Approach 2:
The extraction unit selectively extracts only the necessary candidate features from the full inspection data set, and the selection unit further extracts representative features from these candidates. This extraction approach isolates the essential information needed for accurate classification while discarding redundant data, thereby reducing processing time without sacrificing precision.
3Loss of information
If comprehensive feature extraction is performed on all inspection data, then the completeness of failure analysis improves, but the data processing load and system resource consumption increase
Solution Approach 1:
The system performs partial feature extraction by identifying candidate features and then selecting only representative features from this candidate set. This partial processing approach maintains the completeness of essential failure information while avoiding the excessive computational resource consumption that would result from processing all possible features comprehensively.
Solution Approach 2:
The extraction unit takes out candidate features from the complete inspection data, and the selection unit further extracts representative features from the candidates. This two-stage extraction process isolates the critical information needed for complete failure analysis while minimizing the data volume and computational resources required.
Data Source
AI summary
An inspection result analysis device includes a processor. The processor factorizes each first distribution of a first distribution group into a second distribution group and a weight group corresponding to the second distribution group by non-negative matrix factorization of the first distribution group. Each first distribution is a distribution of numerical data on a surface of a substrate, the numerical data resulting from an inspection of the substrate. The processor generates a fourth distribution by multiplying a third distribution by a weight corresponding to the third distribution to acquire a representative value of a numerical data group constituting the fourth distribution. The processor associates a fifth distribution with the third distribution when the representative value exceeds a threshold. The fifth distribution is one first distribution of the first distribution group. The third distribution is one second distribution of the second distribution group which is factorized from the fifth distribution.