Wafermap Pattern Recognition

Introduction

Trying to catch and understand defect/failure patterns can be quite challenging. Our solution brings together different aspects of the Spotfire analytics stack to solve this tricky problem. Available for download on the Spotfire Exchange, the Wafer Pattern Detection Dashboard for Spotfire® can get you started. This dashboard helps you identify wafer map patterns that are interesting, helps to find more examples of those patterns, and gives you, regardless of your level of data science expertise, the ability to train accurate machine learning models to detect those patterns in the future data.  

• Try out a cloud version of this demo in the Spotfire Demo Gallery.

Summary of Capabilities

The dashboard is split into the following sections:

• Upload Data: On this page, there are instructions for uploading your own data (bin or parametric) into the dashboard.
• Generate Patterns: Here, you will be able to explore the data and categorize wafers you find interesting into patterns. Through our data pipeline, the wafers will be pre-clustered into similarly looking groups of wafers. This will give you a good starting point for finding interesting wafers/patterns.
• Create Models: Now that some patterns have been created, you can create powerful machine learning models that will recognize a specific pattern in the future. Through an iterative process, you can train and correct models until the model adequately recognizes your selected pattern. After, there is an option to save your model.
• Upload New Data: On this page, you will take similar steps to the Uploading Data page for bringing in a new batch of data.
• Score New Data: Models are ready to go, so now you can deploy those models on unseen data. By selecting from your library of created models, you can run pattern recognition to detect selected patterns in new data.

Overview of Methods

Our solution utilizes a sophisticated data pipeline to transform and process wafer data.  It efficiently reduces dimensionality and handles missing values. The final step is clustering the wafers using self-organizing maps (SOM). This method is an artificial neural network that not only groups wafers into similar looking clusters, but also organizes them spatially. This means that clusters next to each other contain wafers that might look similar. SOM makes it easier for the user to identify patterns in their wafer data.

To Learn More

TIBCO Analytics Forum | Watch the end-to-end demonstration of our wafermap pattern recognition work:  

Dr. Data Science | This is the first of a series of Dr. Data Science episodes where we detail this project from start to finish: 

TIBCO Analytics Meetup | At our TIBCO Analytics Meetup, we give a demo of this project and field Q&A about our techniques/approach: