果冻传媒

Hi, my name is Maurice Poot, and I enable semiconductor manufacturing machines to learn from their mistakes. During my PhD research in the Control Systems Technology group within the Department of Mechanical Engineering, I focused on improving the performance of mechatronic machines by developing new learning-based motion control methods. My research was directly funded by , a leading global supplier of machines for semiconductor assembly & packaging, making application-driven research a central focus from the outset. This meant understanding ASMPT's specific challenges, developing new methods to address them, and demonstrating these advancements on industrial-scale machines.

The mechatronic machines used to produce computer chips, such as wire bonder machines that make micrometer-scale connections on chips, need to achieve higher accuracy and throughput to meet the demands on more powerful, smaller and cheaper computer chips. This poses technological challenges for the control software that enables the operation of these machines, as it needs to cope with higher forces, vibrations and numerous unwanted effects that are difficult to model. To address these challenges, learning-based motion control provides major opportunities, as it uses data to fine-tune the machine for higher accuracy.

In my research, I collaborated closely with Guido Knippels, Dragan Kosti膰 and the rest of their team at the ASMPT Center of Competency in Beuningen to tackle their specific challenges. Even before starting my PhD, I had the honor of meeting Wong YM, the CTO, and learning about ASMPT's commitment to bridging the gap between academia and industry. The support I received from them was exemplary, providing access to commercial machines and the freedom to explore new methods for their challenges. To gain a deeper understanding of their problems, ASMPT organized a visit to their facilities in Hong Kong, allowing me to see the production of the industrial-scale machines. This understanding led to the development of new frameworks that integrate control engineering and machine learning. Specifically, machine learning methods like Gaussian processes provide accurate data-based learning without the need to specify detailed machine models to deal with difficult-to-model effects that limit accuracy. These frameworks facilitate high throughput and accuracy on an industrial scale through automated learning.

At the university, I was supervised by Tom Oomen from the Control Systems Technology group and Jim Portegies from the Department of Mathematics and Computer Science. Tom and Jim, who initially connected with each other through the Eindhoven Young Academy of Engineering, played important roles in guiding the research. Our work showcased the value of multidepartmental collaboration, blending the expertise of control engineering and mathematics. This combination proved to be beneficial not only for ASMPT but also for academia. We shared our findings at international congresses all over the world and at local events like the Precision Fair, alongside Jan-Jaap Koning who is now Program Officer at the High Tech Systems Center. Not only did Tom and Jim play an important role in the success of this research, but numerous outstanding MSc students also made significant theoretical and practical contributions. A few of these students went on to become PhD students and my colleagues.

Working on real-world problems presented by ASMPT was a unique experience for me and the students I supervised. Unlike other high-tech companies, ASMPT allowed us direct access to the commercial machines. Initially, this was challenging, as we had to manually transfer the control signals in txt-files from our Matlab environment towards the commercial machines to perform an experiment. For our learning methods, we sometimes needed 30 consecutive experiments to observe the results of our developed algorithms. Adjusting settings and parameters for each iteration was time-consuming. Later, Kelvin Kai Wa Yan from ASMPT developed a server running Matlab, reducing the 20-minute cycle to a few seconds and fully automating the process, which significantly saved time and facilitated easier development on commercial machines.

The ability to experiment easily on commercial machines had significant benefits for both our research and its deployment. The learning algorithms we developed were directly applicable to commercial machines. ASMPT recognized this, and when we demonstrated performance improvements with new algorithms, they quickly deployed them on their commercial machines, sometimes even before we finished writing the publications. This short cycle of feedback for application-driven research on commercial machines exemplifies the benefits of collaborations between academia and industry. It highlights how application-driven research can drive innovation that contributes to advancements in technology that impact our daily lives.