Manufacturing Automation with Artificial Intelligence
Our Client is an established equipment manufacturer supplying to the semiconductor flat panel display industry.They also provide equipment that goes into the manufacture and inspection of photomasks for flat panel displays(FPD) , which serve as the original template for chip manufacture.
Semiconductor manufacturing fabs depend on the photomask which is the master template of an original IC Design that will be printed onto millions of wafers that then go into our everyday electronic devices like our flat panel televisions or mobile phones. These photomasks are produced at a photomask facility.
It is imperative that the photomask that leaves the mask facility is free of defects so that it does not travel into production fabs downstream and we are not left with defective devices.The chip manufacturing industry, specially for electronic goods, is highly competitive with an intense pressure to deliver on time and lower costs to stay competitive.
Manufacturing Automation Functions:
Inspection, Quality Control, Defect Detection, Classification, Artificial Intelligence, and Computer Vision.
Results:
​
-
TIME: 3000 images in under 6 minutes from 7-10 hours earlier.
-
AUTOMATED: 24/7 availability
-
ACCURATE - saves on human effort and errors on judgement. 100% improvements in identifying and classifying false defects.
-
ANALYTICS: Immediate batch and source analysis for actionable insights.
Challenge
Defects, if present, are classified as hard or soft defect and curated by their size and type at the mask facility. There are a lot of false defects flagged. There were as many as 1000 false defects out of every 2000 images checked. That is a 50% False Defect Rate.
​
To identify the False Defects, an operator would manually inspect the masks peering over a high resolution microscope. For defects smaller than 4 pixels, this was an effort-intensive task taking upwards of 7-10 hours, causing great eye-strain and prone to human-errors in judgement. Moreover, there was no classification done at this level with defects being classified as 'Defect/ No Defect'. This put an ambiguity on the subsequent repair operation with a heavy reliance on the skill of the operator.
Solution
The solution was designed in three modules:
Pre-processing :
The pre-processing module takes input images from the scan.
It suppresses noise to remove or isolate unwanted parameters. It
processes them at pixel level, detects defects, and generates a defect map
Defect Classification : This takes inputs from the pre-processing module to classify and output the defect label in real-time.
Dashboard: This offers the user interface and provides the option to input data, view preprocessed output as well as defect label in real-time.
