Wafer Fault Reference Map Denoising for Accurate Classification

The patent applies preliminary action by pre-processing reference maps before they are used for fault classification. Auto-encoders are trained in advance to learn noise patterns specific to each fault type, and these pre-trained models are then used to remove noise from reference maps. This preliminary noise removal and map generation actions improve the quality of reference data before actual fault classification occurs.

The patent implements local quality by creating fault-type-specific auto-encoders that are tailored to the characteristics of each particular fault type. Instead of using a single generic noise removal model, separate auto-encoders are trained for different fault types (e.g., scratch, particle, dishing), allowing each model to learn and remove noise patterns specific to its designated fault type, thereby improving local classification accuracy.

The patent applies copying by generating synthetic reference maps through the auto-encoder framework. Instead of collecting and storing numerous real reference maps for each fault type, the system learns the underlying patterns from limited real data and generates additional synthetic reference maps that replicate these patterns. This copying approach expands the effective size of the reference database without proportionally increasing the complexity of data collection and management.

The patent utilizes parameter changes by transforming the reference maps through the auto-encoder process. The auto-encoders learn optimal parameter transformations that remove noise while preserving fault characteristics. By changing the parameters (noise levels, map characteristics) through the learned transformation, the system generates enhanced reference maps that improve classification accuracy without requiring proportional increases in database complexity.

The patent applies preliminary action by pre-training the auto-encoders offline before actual fault classification operations. The noise removal process is performed in advance during the model training phase, and once trained, the auto-encoders can rapidly apply the learned noise removal patterns to new reference maps during operational use. This preliminary training action separates the computationally intensive learning phase from the faster application phase.

The patent implements self-service by using the auto-encoder framework to automatically learn and apply noise removal patterns without requiring manual intervention for each noise removal operation. The system trains the models to self-adapt to different fault types and noise patterns, and once trained, the models autonomously perform noise removal on reference maps, reducing the need for continuous human oversight and manual processing.