Vlado Menkovski bio photo

Vlado Menkovski

Assistant Professor at Eindhoven University of Technology.

Email Google Scholar

Project Description: Our ADC (automated defect classification) team has worked for the past 2 years on developing and optimizing the classification of SEM images using convolutional neural networks (CNNs). In 2019, we will move to a new state-of-the-art SEM inspection tool. Images created by this tool can differ significantly from our existing data pool, and our trained CNN-models will have difficulty classifying them.

Your goal will be to enable the continued use ADC with the new inspection tool. Some possibilities that can be explored are:

(1) Zero-shot learning: by taking SEM images of the same defects with both the old and new inspection tool, teach the computer how to extract the same information from the new images. That way, the old models can be reused without the need to relabel any data.

(2) One-shot / few-shot learning: make use of prior knowledge of learnt categories to quickly retrain the model with few examples.

(3) Semi-supervised learning: the model provides suggestions to the classification expert after a few examples have been manually classified, and the expert can continuously revise the model.

Company Description: About ASML: ASML has 19,000 employees worldwide. We are headquartered in Veldhoven (The Netherlands) and have over 60 offices in 16 countries. ASML provides chipmakers with everything they need - hardware, software and services - to mass produce patterns on silicon, helping to increase the value and lower the cost of a chip. Our key technology is the lithography system, which brings together high-tech hardware and advanced software to control the chip manufacturing process down to the nanometer.

About the group: This graduation project is positioned within the group responsible for qualifying the defect performance of ASML machines. Part of this qualification process involves exposing wafers and inspecting them with a scanning electron microscope (SEM). This creates large quantities of SEM images that need to be classified into different defect and non-defect classes (mostly by hand). Making this process more efficient by automation and better classification schemes is essential for faster and more robust system qualifications.