Semiconductor Defect Inspection Using Etch-Labeled Image Pairs

The patent applies preliminary action by training a machine learning model in advance using paired images (etched and unetched specimens) to learn the correspondence between defect appearances in etched and unetched states. This pre-trained model can then rapidly classify defects in unetched specimens without requiring actual destructive etching during inspection, thus maintaining high classification accuracy while eliminating the time-consuming destructive process.

The patent uses copying by creating synthetic training data through image processing. It generates synthetic etched images from unetched images (and vice versa) using image processing techniques, thereby creating large datasets for training the machine learning model without requiring physical destructive etching of numerous specimens. This virtual copying approach enables the model to learn from abundant data while preserving actual specimens.

The patent replaces the mechanical/chemical destructive etching process with a computational machine learning system. Instead of physically etching specimens to reveal defect characteristics, the system uses a trained neural network model that analyzes optical images and predicts defect types based on learned patterns from training data, thereby substituting a non-destructive computational approach for a destructive physical process.

The system performs preliminary training of the machine learning model using paired etched and unetched specimen images before actual inspection. This pre-training phase enables the model to internalize the relationship between defect appearances in etched and unetched states, so that during production inspection, defects can be classified rapidly without performing actual destructive etching, thus maintaining high reliability while improving productivity.

The patent applies copying by generating synthetic paired images through image processing. It creates synthetic etched images from unetched images (or vice versa) using image processing algorithms, thereby generating large volumes of labeled training data without requiring physical etching of specimens. This virtual data generation approach reduces the complexity of data acquisition while providing sufficient training data for accurate model training.

The patent introduces an image processing module as an intermediary that transforms unetched specimen images into synthetic etched images (or vice versa). This intermediary process automatically generates the paired labeled data needed for training, eliminating the need for manual data collection through destructive etching and significantly reducing data acquisition complexity while maintaining data quality for accurate model training.