BE Semiconductor Industries (BESI) produces die bonders that place chips with micrometer precision at speeds of several thousand units per hour. As a computer science student at the University of Zurich, I joined BESI's R&D team to integrate machine learning into quality control, developing a computer vision system for automated defect detection in memory chips.
I initiated and managed the full data science lifecycle to replace a traditional computer vision algorithm with a TensorFlow-based deep learning model. I implemented a CNN architecture achieving 99% accuracy in automated defect detection at production speeds incompatible with manual inspection. The deep learning framework integrated into existing production processes, including parametrization and visualization tools with TensorBoard for production team operation.
Through workshops and documentation, I trained the team in system operation and underlying ML concepts. The interactive training materials became the foundation for subsequent ML projects at BESI. The system remains in production use and established the pattern for AI integration in quality control processes.
The main objective was developing an ML system for automated defective memory chip detection using computer vision in semiconductor production. The system had to achieve 99% accuracy at high production speeds and integrate seamlessly into existing processes.
The project combined deep learning and CNN architecture with industrial semiconductor production requirements. Modern frameworks like TensorFlow were used, and tools for parametrization and visualization like TensorBoard were implemented to enable effective use by the production team.
Knowledge transfer was ensured through regular workshops, detailed documentation, and interactive training materials. These materials conveyed not only system usage but also deep understanding of underlying ML concepts and served as foundation for further ML projects at BESI.
The system had to control chips with micrometer precision at speeds of several thousand units per hour. The implemented CNN architecture achieved 99% detection accuracy under real production conditions.
Integration followed a systematic approach from theoretical conception to practical implementation. Special tools for parametrization and visualization were developed, enabling the production team to effectively integrate the system into existing production processes.
The developed system continues to be used at BESI and served as starting point for further AI projects in quality control. It established new standards for ML integration into industrial processes and contributed to quality control modernization in semiconductor production.
Academic background in machine learning was crucial for the systematic approach from theoretical conception to practical implementation. Scientific methodology enabled successful transfer of complex ML concepts into industrial applications.
Main challenges included connecting deep learning with strict industrial requirements, ensuring high precision at high speeds, sustainable anchoring in the company through knowledge transfer, and seamless integration into existing production processes.
.
Copyright 2026 - Joel P. Barmettler