Access Official Technical PDF

Share Summary:LinkedInEmail

Semiconductor defect classification is a demanding part of wafer failure analysis. Particle defects can arise as manufacturing processes, equipment, and handling practices change, while manual review of scanning electron microscopy (SEM) images and energy-dispersive X-ray (EDX) spectra takes time and can be affected by inconsistent interpretation. A deep-learning approach that combines both data types offers a way to make chemical-composition classification more systematic and support the investigation of defect sources.

The central idea is straightforward: an SEM image describes a particle’s visible characteristics, while its EDX spectrum provides evidence about chemical composition. Because either source can be ambiguous on its own, a convolutional neural network (CNN) can use both when estimating the likely defect class.

Why semiconductor defect classification is challenging

Defect analysis commonly proceeds from broad inspection to more detailed characterization. Wafer-level inspection identifies defect locations and distributions; SEM review helps assess individual particle morphology; and EDX analysis helps determine chemical composition. Each stage contributes useful information, but none guarantees a definitive identification in every case.

Wafer maps may suggest spatial patterns associated with a process step or equipment condition. SEM images can reveal shape, size, and surface features, yet particles from different classes may look similar. EDX spectra add chemical evidence, but small particles can produce weak signals, and overlapping peaks can make elements difficult to distinguish. Automatically labeled peaks can also be affected by noise or overlap.

These limitations matter because chemical composition can help narrow the possible source of a particle. Equipment materials and process conditions provide context for interpreting what a defect contains. Automating classification may therefore help engineers review more defects consistently, although classification is one part of root-cause analysis rather than a replacement for process knowledge.

How the CNN combines SEM images and EDX spectroscopy

The evaluated system takes a centered, top-view SEM image of a wafer section containing a particle and processes it through convolutional and pooling layers. Convolutional layers learn image features; pooling reduces the spatial detail passed forward. The resulting visual representation captures characteristics such as shape and topography.

The network then merges that visual representation with the corresponding EDX data in its fully connected layers. The spectral inputs include raw intensity values and automatically labeled peaks. Using raw values alongside peak labels gives the model additional spectral information, including where automated peak labeling may be imperfect. The network produces probabilities for the candidate defect classes rather than only a single unqualified label.

This fusion is important: the image and spectrum provide different, complementary evidence. The study’s comparisons found that the combined SEM-and-EDX model performed better than models using either image data or spectral data alone. On the reported test set, the combined model achieved 82.1% Top-1 accuracy and 99.2% Top-3 accuracy. Top-1 measures whether the highest-probability class is correct; Top-3 measures whether the correct class appears among the three most likely results. The authors considered the Top-3 result above their stated 95% baseline for industrial utility.

Training data and class balance affect results

The model was evaluated using 5,761 defects across eight chemical-composition classes. The class counts were uneven, meaning some types were represented by substantially more examples than others. The data was split into training, validation, and test sets, with five-fold cross-validation used during training and model selection.

That imbalance showed up in class-specific results. Some classes were classified accurately, while others were more often confused with chemically or visually similar classes. For example, SiOxCy defects were frequently classified as SiO2, and CuO/S defects were frequently classified as Fe-Ni/O. The study notes that similarities in the available evidence and differences in the number of training examples can contribute to these errors.

In a comparison using equal numbers of examples per class, classification of CuO/S and SiOxCy improved, while accuracy for some more highly represented classes decreased. This illustrates an operational trade-off: overall performance can conceal weaker results on underrepresented defect types. For deployment, teams should examine confusion matrices and class-level performance, not just a single aggregate accuracy measure.

Data augmentation was also investigated. Because the defects were treated as orientation-independent, training and validation images were rotated to expand the available examples; test images were not rotated. The work also explored how training-set size and augmentation affected performance. These techniques can help make more use of limited labeled data, but they do not eliminate the need for representative examples and reliable labels.

Plan for defect-free images and model updates

The classification network was designed for data containing a defect. The researchers separately addressed SEM images of wafer areas with no particle, using image-based Canny edge detection to distinguish defect-free images from images containing a visible defect before applying the SEM-and-EDX classifier. This separation was important because the EDX spectrum from a defect-free wafer area could contain silicon and oxygen signals that interfere with classification of silicon-based particles.

Transfer learning was explored as a potential way to reduce the effort of retraining when defect classes change. The study retrained models using a subset of classes and then tested approaches that reused learned network weights. This is relevant to manufacturing environments where new defect types may appear, but the results should be interpreted as an investigation of the approach, not a guarantee that transfer learning will improve every update.

Practical implications for engineering teams

For process engineers, equipment teams, and integrators, the main lesson is that multimodal data can make automated defect classification more useful than relying on morphology or spectra alone. A practical evaluation should consider:

  • Data coverage: Check whether each important defect class has enough representative, correctly labeled examples.
  • Class-specific performance: Use confusion matrices to identify where similar classes are being mixed up.
  • Workflow boundaries: Decide how defect-free images are identified before chemical classification.
  • Model maintenance: Plan for retraining and validation as processes and defect classes change.

Combining SEM image analysis with EDX spectroscopy can provide a more consistent basis for classifying wafer particle defects and supporting root-cause investigations. The reported results are promising, particularly for Top-3 classification, but performance depends on the training data and the defect classes being assessed. For industrial use, validation should reflect the application’s data, class balance, and review requirements.

Access Official Technical PDF

About The Author