Real-Time Defect Classification in 3D IC Subcomponents
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The high-density and three-dimensional nature of advanced integrated circuit packaging technologies makes it challenging to detect and classify defects in connections and subcomponents, such as through-silicon vias and solder joints, in real-time.
Innovation Solution
The technology employs a method for rapid defect classification using a synthetic training set generated by Monte Carlo simulation, feature extraction with techniques like SURF, Haar filters, or Hu moments, and principal component analysis (PCA) to transform feature vectors into an uncorrelated space, enabling real-time classification with artificial neural networks or other models in a high-speed x-ray imaging system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional defect detection methods are used in advanced 3D integrated circuit packaging, then defect detection may be achieved, but real-time classification becomes infeasible due to computational complexity
Solution Approach 1:
The system performs preliminary actions by pre-computing feature extraction parameters and classification models during system initialization or offline training phases. The neural network is trained in advance with synthetic defect data, and feature extraction parameters are pre-calculated, enabling rapid real-time classification without performing heavy computations during actual defect detection
Solution Approach 2:
The system creates synthetic copies of defect data through Monte Carlo simulations to build training datasets. These synthetic defect images and their corresponding feature vectors serve as proxies for actual defects, enabling the neural network to learn classification patterns without requiring extensive real defect samples or complex real-time analysis
2Measurement precision
If detailed feature extraction is performed on all subcomponents, then classification accuracy improves, but processing time and computational resources increase
Solution Approach 1:
The system extracts only the most relevant and discriminative features from subcomponent images using pre-determined feature extraction parameters. By selecting specific features (such as shape, texture, or structural characteristics) that are most indicative of defect types, the system avoids processing all possible image features, thereby maintaining accuracy while reducing computational burden
Solution Approach 2:
The system transforms the feature space by applying parameter changes through principal component analysis or similar dimensionality reduction techniques. This transforms the original feature vectors into a reduced set of uncorrelated components that capture the essential variance, enabling faster processing while preserving classification accuracy
3Productivity
If multiple subcomponents are inspected simultaneously, then productivity increases, but measurement precision and classification accuracy decrease
Solution Approach 1:
The inspection system segments the overall inspection task by processing different subcomponents through specialized classification pathways. Each subcomponent type (e.g., TSVs, solder joints, bump bonds) can be handled by dedicated neural network classifiers or feature extraction parameters optimized for that specific subcomponent, allowing parallel processing while maintaining high accuracy for each type
Solution Approach 2:
The system employs a universal neural network architecture and feature extraction framework that can handle multiple subcomponent types simultaneously. The same core classification engine processes different subcomponent types by adapting to their specific characteristics through the pre-computed feature parameters, enabling multi-functional inspection without sacrificing precision
Data Source
AI summary
The present disclosure provides methods and apparatus for rapidly classifying detected defects in subcomponents of a manufactured component or device. The defect classification may occur after defect detection or, because the classification may be sufficiently rapid to be performed in real-time, during defect detection, as part of the defect detection process. In an exemplary implementation, the presently-disclosed technology may be utilized to enable real-time classification of detected defects in multiple subcomponents of the component in parallel. The component may be, for example, a multi-chip package with silicon interposers, and the subcomponents may include, for example, through-silicon vias and solder joints. Defects in subcomponents of other types of components may be also be classified. One embodiment relates to a method of classifying detected defects in subcomponents of a manufactured component. Another embodiment relates to a product manufactured using a disclosed method of inspecting multiple subcomponents of a component for defects.