Meteoriet Project: AI-Based MEMS Defect Detection
We are looking for a graduation or internship student in the area of Computer Science, Applied Physics or Electrical Engineering for an assignment within the EFRO research project “METEORIET”. Here is a link to a popular publication about this project and its partners.
Many micro-scale devices are widely used in today’s high-tech society: microprocessors, integrated electronics, accelerometers and other sensors in for example consumer products such as phones, computers, EV’s or in smart industrial systems such as autonomous vehicles, robots or even in healthcare. The manufacturing of these so-called MEMS (Micro Electro Mechanical Systems) requires many fabrication steps in a cleanroom environment, each of which adds to the physical structure and thus functionality of the dies (die = individual chips on a silicon wafer, see Figure 1). This also includes errors/defects and hence variations in each die from its nominal geometry or material properties.
Figure 1 A wafer with MEMS dies during production
Within the “Meteoriet” project, the project partners develop MEMS test & inspection technologies so that their fabrication can be scaled up in a cost-effective and environment friendly way.
The prospective candidate will be working on the research and development of an automatic visual inspection method for MEMS production. The focus is on using Convolutional Neural Networks (CNNs) together with computer vision in order to detect and localize defects on a MEMS die. We train the CNN using artificially generated faulty images or so-called synthetic training datasets. Since real defects are rare among the wafer dies, we effectively mimic them instead. Every defect has a specific visual appearance and we generate them artificially on a drawing of an “ideal” die that we know beforehand. The trained CNN is then able to detect the defect class and localize it on a real image so that an operator can inspect the results faster and without human errors.
Previous work has allowed us to gather a database of wafer die images and develop a prototype tool that can generate some classes of synthetic defects from an ideal die drawing.
The current tasks are to estimate how the accuracy of the CNN trades-off with the amount of detail in the artificial training defects as well as add to the functionality of the artificial defect generation tool. We work with real practice on images of existing MEMS products.
What tasks do we expect from the graduate to encounter during the project?
Analyse, characterize, describe and capture in code various defect classes from available real die images
Refine and add new classes to the synthetic defect generation (SDG) tool
Generate and use synthetic training datasets to optimize the AI’s object detection model and evaluate its accuracy
Depending on the study direction, background (University or University of Applied Sciences) as well as the nature of the stay (internship or graduation) there are different levels of depth that can be achieved in the above tasks. These are to be discussed with the student’s study advisor before the beginning of the project so as to fit any requirements.
Study directions: Computer Science, Applied Physics, Electrical Engineering, Mechantronics
The student will be working within the “Meteoriet” project and have regular meetings with the project partners. The assignment will be carried out at the Saxion Research Group Applied Nanotechnology in Enschede (NL) and will be supported by the Dutch Space Organisation SRON.
Are you interested? Send an email to Aleksandar Andreski (email@example.com).