Wafer Defect Map Classification Using UMAP and Density Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification systems struggle to accurately and efficiently classify wafer defect maps, particularly those with additional or unclassified defect patterns, due to their inability to identify these patterns and the computational expense of existing methods, leading to inaccuracies and impracticality in real-time manufacturing processes.
Innovation Solution
A system utilizing unsupervised classification circuitry with non-linear dimensionality reduction and density-based clustering models, such as Uniform Manifold Approximation and Projection (UMAP) and Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), to generate wafer defect pattern clusters and classification predictions, along with supervised classification using convolutional neural networks, to update and refine defect pattern classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing classification systems are used to classify wafer defect maps, then classification can be performed, but accuracy is poor especially for additional or unclassified defect patterns
Solution Approach 1:
The patent inverts the traditional supervised classification approach by implementing unsupervised classification. Instead of requiring pre-labeled training data and fixed defect pattern categories, the system allows defect patterns to self-organize into clusters without prior knowledge of their identities. This inversion enables the discovery of additional or previously unclassified defect patterns while maintaining high classification accuracy.
Solution Approach 2:
The patent transforms the classification problem by changing the parameter space through non-linear dimensionality reduction. By projecting high-dimensional defect map data into a lower-dimensional manifold space, the system preserves critical defect pattern characteristics while removing redundant information. This parameter transformation enables more accurate clustering and identification of defect patterns that were previously indistinguishable.
2Measurement precision
If existing classification methods are applied to wafer defect maps, then classification results can be obtained, but computational expense is high making real-time implementation impractical
Solution Approach 1:
The patent extracts and applies pre-trained deep learning feature extractors to wafer defect maps, obtaining rich feature representations without performing expensive end-to-end training. By separating feature extraction from clustering, the system achieves high classification accuracy while significantly reducing computational requirements for real-time deployment.
Solution Approach 2:
The patent applies partial action by using only the feature extraction capability of pre-trained models rather than full training. This selective application of deep learning provides sufficient feature quality for accurate clustering while avoiding the excessive computational cost of training models from scratch, enabling real-time processing.
3Adaptability or versatility
If traditional classification approaches are used, then existing defect patterns can be categorized, but additional or unclassified patterns cannot be identified
Solution Approach 1:
The patent implements self-service through unsupervised clustering, where the system automatically discovers and organizes defect patterns without human intervention or pre-defined categories. The clustering algorithm autonomously identifies meaningful groupings in the data, enabling both classification of known patterns and discovery of new patterns simultaneously.
Solution Approach 2:
The patent performs preliminary dimensionality reduction and feature extraction to transform raw defect map data into a form suitable for clustering. This preliminary processing preserves essential pattern information while reducing data complexity, enabling effective unsupervised learning that can identify both existing and new defect patterns.
Data Source
AI summary
Various embodiments of the present disclosure provide methods, apparatus, systems, computing devices, computing entities for generating classification predictions for wafer defect maps. Reduced feature data may be generated, using a non-linear dimensionality reduction machine learning model and based at least in part on vector representations for a set of wafer defect map images. One or more wafer defect pattern clusters may be generated, using a density-based clustering machine learning model and based at least in part on the reduced feature data. Each wafer defect map image may be associated with a particular wafer defect pattern cluster of the one or more wafer defect pattern clusters. A classification prediction may be generated for each wafer defect map image based at least in part on the particular wafer defect pattern cluster associated with the respective wafer defect map image.