Defect Data Embedding for Wafer Inspection Classification

The patent introduces machine learning models as intermediary tools between the defect images and human analysts. These models automatically categorize defects using learned patterns from training data, providing preliminary classification results that human experts can then verify or adjust. This intermediary processing layer significantly reduces the time required for initial defect analysis while maintaining or improving accuracy through consistent, data-driven classification.

The system enables self-service defect classification through machine learning models that autonomously categorize defects without requiring immediate human intervention. The models are trained on historical defect data and can independently perform initial classification, filtering, and prioritization of defects. This self-service capability allows the system to handle large volumes of defect images efficiently, freeing human analysts to focus on complex cases that require expert judgment.

The patent applies preliminary action by implementing comprehensive image preprocessing steps before defect classification. This includes noise reduction, normalization, alignment, and enhancement techniques that prepare images for more accurate machine learning analysis. By performing these preliminary actions, the system ensures that the machine learning models receive high-quality input data, which significantly improves their ability to distinguish subtle but significant differences between defect images and normal variations.

The system employs parameter changes by adjusting and optimizing multiple parameters throughout the defect analysis pipeline. This includes modifying image processing parameters (such as contrast enhancement, threshold values), tuning machine learning model parameters (such as learning rates, feature extraction parameters), and adjusting classification thresholds. These parameter optimizations enable the automated system to achieve high precision in distinguishing significant defect differences while maintaining efficient automation.

The patent applies dimensionality reduction techniques such as Principal Component Analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE) to transform high-dimensional defect data into lower-dimensional representations. This dimensionality change enables effective visualization of defect patterns in 2D or 3D spaces, making it easier for analysts to identify clusters, outliers, and relationships among different defect types. The techniques preserve the most significant variance and structural relationships in the data, maintaining analytical utility while reducing complexity.

The system manages information loss during dimensionality reduction by carefully selecting and tuning reduction parameters. This includes choosing appropriate reduction techniques based on the specific defect data characteristics, adjusting the number of retained dimensions to balance visualization quality and information preservation, and applying preprocessing steps that enhance the signal-to-noise ratio before reduction. These parameter optimizations ensure that the reduced-dimensional data retains sufficient information for accurate defect analysis and classification.