Access Official Technical PDF

Share Summary:LinkedInEmail

In semiconductor manufacturing, pass/fail test results show whether devices meet specification limits, but they do not reveal how close passing devices may be to those limits. Analog wafermap analysis examines the original measurement values and their spatial relationships across a wafer. This can help identify developing patterns that warrant investigation before they result in specification violations.

Automating this analysis is challenging: wafermaps contain noise, missing measurements and patterns that may vary in intensity, size, position or orientation. Two feature-extraction strategies offer different ways to represent those maps for pattern discovery: classical image processing with engineered features, and a deep generative model based on a convolutional variational autoencoder.

Why analyze analog wafer test data?

During electrical wafer testing, measurements are collected for devices across a wafer and compared with predefined limits. Conventional pattern-recognition methods often focus on pass/fail or bin data. Those representations are useful for examining known failures, but they discard detail about the actual measurements of devices that still pass.

Analog data retains that detail. Spatial regularities—such as gradients or concentrated regions of higher or lower measurements—may be present even when all devices remain within specification. Detecting such patterns can support earlier process monitoring and give engineers a basis for investigating potential deviations. It does not, by itself, establish a root cause; it helps organize and surface patterns for further analysis.

Manual wafermap screening can demand substantial engineering effort and may lead to subjective judgments. Feature extraction provides a way to summarize maps consistently so that similar patterns can be grouped, inspected and, where suitable, used in later classification workflows.

Two approaches to wafermap pattern recognition

Classical image processing and engineered features

The classical method first denoises measurements using a Markov random field (MRF). This model uses the wafer’s spatial neighborhood: nearby device measurements are treated as related, while observed values are understood to include noise. The resulting estimate smooths local variation while retaining a representation of the underlying wafermap.

Next, local binary patterns (LBPs) describe how each device’s value compares with those of its neighbors. The comparisons form a local texture code, and a histogram of those codes summarizes patterns across selected wafer regions. Rotated local binary patterns (RLBPs) orient the local description around a dominant direction, helping account for rotations. The method uses automated thresholding to select regions of interest and principal component analysis (PCA) to reduce the combined feature representation to three dimensions.

This pipeline is built from defined processing steps and features. That makes the representation more directly connected to image characteristics that engineers can inspect. It also means the approach depends on how effectively those engineered features capture the patterns that matter.

Deep feature learning with a variational autoencoder

The alternative uses a convolutional variational autoencoder (CVAE), a type of generative neural network. An encoder compresses a wafermap into a compact latent representation; a decoder attempts to reconstruct the input from that representation. Training balances compact encoding with accurate reconstruction, encouraging the model to learn useful structure in the data.

In the evaluated setup, wafer measurements are scaled and converted into images, then processed through convolutional layers. The two-dimensional latent representation can be used to compare or visualize wafermaps. Unlike the classical pipeline, this method learns its representation from examples rather than relying on a predefined set of texture features.

What the comparison found

The approaches were evaluated by clustering extracted features on a synthetic dataset with five prototype pattern types and on a real-world dataset. The simulated patterns included a border ring, a spot, a gradient, two edge spots and a crescent-shaped region, with variations in characteristics such as size and intensity. Clustering results were assessed using internal measures of cluster structure and, for the synthetic data with known pattern labels, an external measure of agreement with those labels.

Overall, the methods produced similar evaluation results on the synthetic data. Both could distinguish important pattern types, although some patterns were grouped together or separated into subclusters. For example, different gradient directions could form separate groups even when they represented variations of the same general pattern. The findings underline that a cluster is a useful way to organize wafermaps, not automatically a definitive process diagnosis.

On real-world data, both methods produced plausible groupings, including distinctions between weaker or noisier maps and maps with more pronounced spatial structure. Because the real dataset did not have ground-truth pattern labels, the results could not be evaluated against a known correct classification. Expert review remains important when interpreting such clusters.

Choosing an approach for production use

The best fit depends on the data, computing resources and how the results will be used. The comparison points to several practical considerations:

  • Limited training data or changing product conditions: The classical method requires less training data; estimating the PCA parameters is its main training step. This can be helpful when a product is new or a process step changes.
  • Interpretability: Engineered features such as local texture patterns are more readily connected to recognizable image properties, which can make them easier for production experts to examine.
  • More complex patterns: A learned representation may be preferable when maps contain complex arrangements or combinations of spots and lines that are difficult to capture with a fixed feature-engineering pipeline.
  • Data and computing availability: The deep learning method can benefit from larger training datasets, but it has a greater training-time and computing-resource requirement. Evaluation after training is less demanding than model training.

Neither approach removes the need for process knowledge. Clustering can reveal similarities and support screening, while engineers must assess whether a pattern is meaningful, whether it corresponds to a known condition and what follow-up is appropriate.

Practical implications

Analog wafermap analysis extends process monitoring beyond pass/fail outcomes by preserving measurement detail that may signal change before specification limits are crossed. Classical image processing offers a more interpretable, data-efficient starting point; a convolutional variational autoencoder offers a more flexible alternative when sufficient data and computing resources are available and patterns are more complex.

For manufacturing teams, the practical next step is to match the method to the available data and operational context, then validate the resulting groups with domain expertise. Both approaches can help turn large collections of wafer test maps into structured evidence for investigation, but neither should be treated as a substitute for engineering judgment or root-cause analysis.

Access Official Technical PDF

About The Author