Reclaim wafer defect classification can help semiconductor manufacturers make more consistent decisions about whether used test wafers are suitable for repolishing. Monitoring and dummy wafers support process control and equipment maintenance, but they are not automatically reusable after use. Their surface condition must be inspected, and wafers with unacceptable defects need to be screened out before reclamation continues.
One proposed approach applies deep learning to defect maps generated during wafer inspection. Rather than relying only on an inspector to interpret each map, a convolutional neural network (CNN) learns visual patterns associated with wafers classified as Good (G) or No-Go (NG). The method illustrates how image-based machine learning can support a specific inspection task in semiconductor manufacturing.
Why reclaim wafer inspection matters
Silicon wafers are a critical manufacturing resource. In addition to prime wafers used to produce integrated circuits, semiconductor facilities use test wafers, including monitoring and dummy wafers, to check process parameters and maintain equipment. The study focuses on reclaiming these test wafers; prime wafers that have already been processed into chips are outside the described reuse process.
Reclamation involves multiple operations, including sorting, stripping, lapping and grinding, polishing, cleaning, and inspection. A wafer may be considered for repolishing when its defects are not disqualifying and enough material remains for the process. The inspection decision is therefore important: a wafer judged reusable may return to manufacturing, while an NG wafer is discarded.
In the approach described, inspection maps represent surface information such as defect position and distribution. Defects may include particles, contamination, voids, and other surface features. The maps can also reflect variations in defect size, shape, location, and orientation, making classification difficult. The source notes that experienced inspectors have traditionally interpreted these patterns, a task that can be tedious and dependent on human judgment.
How deep learning classifies wafer defect maps
The proposed system uses automated visual inspection data from a G-sorter instrument. The instrument captures three-dimensional surface information, including geometry, flatness, and particle position. This numerical information is transformed into graphical defect maps, where color represents height at different locations on the wafer surface.
Before classification, the images are preprocessed to remove information considered unrelated to the defect pattern and to emphasize the relevant area. The reported workflow removes image edges and boundaries, retains a three-channel color image, and resizes the original maps—approximately 1,000 by 1,000 pixels—to 128 by 128 pixels. The intention is to reduce image size while retaining visually useful texture.
Learning features from the full image
Instead of first defining a set of defect measurements by hand, the CNN receives the resized image and learns visual features during training. Its convolutional layers detect patterns at different levels, while pooling reduces the representation passed between layers. Fully connected layers then support the final classification, with a softmax output used to produce class predictions.
The study describes ten typical defect-map types grouped into nine Good categories and one NG category. This distinction matters: the network is trained to recognize more detailed patterns, while the practical screening decision is whether a wafer is acceptable for reclamation. Dropout was used as a measure against overtraining, and the Adam algorithm was used to minimize cross-entropy during training.
Training and reported results
The system was implemented in Python using Keras with TensorFlow as the backend. The paper reports a dataset of approximately 5,000 images, divided into training and testing sets at an 8:2 ratio. It also describes 3,640 reclaimable samples for training, including 646 NG images, and reports separate counts for Good and NG test images. The training configuration used batches of 100 images and allowed up to 300 epochs, with training set to stop automatically when improvement ceased.
During the reported experiments, the learning process was described as converging after about 30 epochs. The authors state that the CNN identified Good and NG wafers effectively and performed better than a back-propagation neural network approach. The paper does not provide a specific classification-accuracy figure for this reclaim-wafer experiment, so its results should be understood as a reported comparison rather than a quantified performance guarantee.
Practical considerations for automation teams
For plant and engineering teams, the main takeaway is that a defect map can serve as an input to a repeatable image-classification workflow. A system of this kind could support inspectors by sorting maps into decision categories, but adopting it requires attention to the data and inspection process behind the model. The reported results depend on the images, preprocessing, class definitions, and training examples used in the study.
- Define the decision clearly: Establish how Good and NG labels correspond to the facility’s reclamation criteria.
- Review image preparation: Resizing and removal of image regions should preserve the surface patterns relevant to wafer disposition.
- Maintain representative training data: The model learns from examples, so the quality and coverage of labeled defect maps matter.
- Evaluate results in context: A reported comparison does not replace validation against the inspection requirements and data of a particular operation.
Conclusion
Deep learning offers a way to automate pattern recognition in reclaim wafer inspection by training a CNN on defect maps rather than relying entirely on manually defined visual features. The described approach transforms surface measurements into images, preprocesses them, and classifies patterns associated with Good and NG wafers. For semiconductor manufacturers, it points to a potential tool for supporting test-wafer reuse decisions. Its practical value depends on suitable inspection data, clear classification criteria, and validation for the intended manufacturing process.