Semiconductor defect detection is a practical application of machine learning in wafer manufacturing: image recognition models can assess inspection images and help identify patterns that may otherwise require close, repetitive human review. At TowerJazz, the goal was to explore whether a model could assist operators examining wafers for defects, while improving identification and control and saving operator time.
The initiative focused on a specific visual inspection process, not on replacing engineering judgment across the fab. Its experience highlights the steps that matter when moving from an inspection challenge to a machine-learning workflow: defining the use case, preparing representative images, evaluating model decisions, and planning how results will fit into operations.
Why wafer inspection is a machine-learning opportunity
In the Fab2 process described by TowerJazz, Viper metrology tools capture dark-field and bright-field wafer images. Operators review those images for issues such as scratches, spots, splashes, and bubbles. If they find no defects, the lot proceeds to the next station. If a defect is identified, the lot is held for process or engineering inspection.
The process ran across five Viper tools, with one or two operators covering inspection around the clock. This makes image review a focused candidate for automation support: the task involves interpreting visual data against defect categories and deciding whether material can move forward or needs further attention. The intended role for machine learning was to help assess images, not to assume that every inspection decision could be automated immediately.
Defining a manageable inspection use case
TowerJazz began with a deliberately limited scope: one technology and one wafer layer. Rather than trying to model every product and inspection condition at once, the team assembled classified images for that area. The initial categories comprised four defect classes and one clean class.
This bounded approach connects the model to a defined manufacturing decision. It also gives the team a practical way to examine whether the training examples represent the images operators encounter and whether model outputs are useful for the process. The project brought together process expertise, IT programming, and data science support from an AWS specialist and AWS partner AllCloud.
For industrial teams considering a similar application, this division of responsibilities is important. Process specialists understand what constitutes a meaningful defect and how inspection decisions affect production. IT and data science contributors can support the data and modeling work, while operations personnel help assess how results could be used in the inspection workflow.
Preparing images and evaluating model performance
Image recognition depends on the examples provided for model development. TowerJazz’s early evaluation identified a specific issue: 21% of inspection events categorized as “spot stops” were defined as good. This indicated that the model’s handling of that category needed improvement before its output could be trusted as a decision input.
The team worked with AllCloud on two actions: adjusting images to create more examples for algorithm learning and changing thresholds. These steps reflect an iterative development process. Model performance is not determined solely by selecting an image-recognition approach; the quality and coverage of labeled data, the distinction between classes, and the decision thresholds all influence how outputs behave.
In a production setting, false defect indications and missed defects can have different operational consequences. The source material does not provide a final accuracy figure or establish that the model was ready to control production. Instead, the reported work describes improvement efforts and identifies further validation as part of the path forward.
Integrating model output into the operator workflow
Even a technically promising model needs an operational interface and a defined response to its predictions. TowerJazz identified an improved graphical user interface for operators as a next step, along with comparing operator decisions against model decisions. That comparison can help teams understand where the model agrees with human review and where additional investigation or training data may be needed.
The planned work also included collecting more classified images and extending the effort to additional layers. These steps matter because a model developed for one technology and layer should not be assumed to apply unchanged to other inspection conditions. Each expansion requires relevant examples and evaluation in the context where the model will be used.
TowerJazz’s longer-term objective was to deploy a model as the decision maker for each modeled layer and technology. That was a next step, not a reported completed deployment. For manufacturers, the distinction is useful: model development, operator decision support, and autonomous disposition are different stages, each requiring appropriate validation and operational controls.
What the use case means for semiconductor manufacturers
Machine learning can also address manufacturing problems beyond visual inspection. The broader semiconductor applications described in the source include defect classification, yield diagnostics, failure prediction, lithography optimization, predictive maintenance, and excursion prevention. Across these use cases, the common foundation is the ability to collect and use engineering and operational data effectively.
For wafer image inspection, a practical evaluation plan should include:
- A specific scope: Start with a defined technology, layer, and inspection decision.
- Useful labels: Work with process experts to classify clean images and relevant defect types.
- Representative data: Review whether the image set adequately reflects the cases the model will encounter.
- Operational comparison: Compare predictions with operator decisions and investigate disagreements.
- Controlled expansion: Gather and evaluate additional data before extending the model to other layers or technologies.
Conclusion
TowerJazz’s work shows how semiconductor defect detection can begin with a narrow, well-defined inspection task and develop through repeated evaluation of image data and model decisions. The initial findings exposed a classification problem, prompting image and threshold adjustments, while the next steps centered on operator interfaces, decision comparison, and broader training data. For manufacturers, the practical lesson is to treat machine learning as a process improvement effort involving data, domain expertise, and workflow design—not simply as a model-selection exercise.