Random Forest Classifier for Semiconductor Defect Inspection

The patent segments the defect inspection process into multiple stages: initial high-sensitivity optical scanning to capture all potential defects, followed by systematic filtering through multiple criteria (image quality assessment, defect characteristic analysis, spatial relationship evaluation). This segmentation allows the system to maintain high detection sensitivity while systematically eliminating nuisances at each stage.

The patent applies different evaluation criteria and filtering thresholds to different types of defects and different regions of the wafer. By analyzing local characteristics such as defect morphology, size, location, and contextual features, the system can distinguish between real defects and nuisances more effectively, reducing false positives while maintaining detection sensitivity.

The patent implements an automated classification system that performs defect verification and nuisance filtering without continuous human intervention. The system uses machine learning algorithms and automated decision-making protocols to independently evaluate candidate defects, apply classification rules, and filter nuisances, thereby eliminating the time-consuming manual review process while maintaining high classification accuracy.

The patent replaces the mechanical human-operated filtering process with an automated computational system. Instead of relying on operators to manually examine and classify each defect candidate, the system uses computer vision algorithms, pattern recognition, and automated decision-making to perform the same function, dramatically reducing time consumption and labor requirements.

The patent implements automated feature selection and parameter optimization that performs extensive exploration of the feature space without human intervention. The system automatically identifies relevant features, optimizes decision thresholds, and fine-tunes classification parameters through self-learning and adaptive algorithms, thereby achieving effective filtering without requiring manual labor for feature exploration and parameter tuning.

The patent employs automated parameter optimization techniques that dynamically adjust filtering thresholds, decision tree cut values, and classification parameters based on the specific characteristics of the inspection data. This automated parameter tuning achieves effective filtering adaptively without requiring manual fine-tuning, reducing labor requirements while maintaining or improving filtering effectiveness.