THIS STORY HAS BEEN APPROVED BY THE CUSTOMER FOR EXTERNAL PUBLICATION. PLEASE FEEL FREE TO SHARE!
Read it here: jmp.com/KAIST The recording is also available on JMP's YouTube.
Drawing from real industry experience, Associate Professor Lee Jong-Seok at KAIST shows why AI must respect basic physical principles to be trusted on the factory floor.
Rethinking AI for real manufacturing problems
Short summary:
Associate Professor Lee Jong-Seok at KAIST shares the challenges of applying AI in real-world manufacturing. While AI models are growing larger, they often lack transparency and fail to respect fundamental physics, making human expertise essential. His research focuses on bridging academia and industry by building simple, interpretable models that remain accurate and powerful. By incorporating domain knowledge and statistical constraints, he solved real manufacturing problems, such as predicting steel coating weight, demonstrating that AI alone isn’t enough without deep expertise and understanding of the problem context.
Key takeaways:
- Simpler models drive real-world impact: While large AI models dominate research, Prof. Lee emphasizes that simpler, more interpretable models often perform better in manufacturing. Models must align with real physical principles to be trusted and deployed on the production floor.
- Physics-Constrained AI Improves Reliability: In a steel coating application, a standard neural network produced predictions that violated basic physics. By embedding statistical constraints into the model to enforce a positive relationship between process variables, the team ensured predictions remained both accurate and physically consistent.
- Bridging Academia and Industry: Academic methodologies do not always translate directly to factory environments. Prof. Lee positions his work between theory and practice—adapting advanced machine learning methods, so they solve real industrial problems
- Interpretability Enables Action: Manufacturing decisions require clear cause-and-effect relationships. By prioritizing transparency over model size, engineers can confidently adjust process parameters without risking unintended consequences.
- Statistical Foundations Still Matter: Advanced AI alone is not enough. Deep knowledge of statistics and optimization is critical for designing robust, explainable, and safe models for real-world applications.
Quotes from Associate Professor Lee, KAIST
“AI at this moment is not transparent enough. People just apply AI to their own problems, even though they don’t know if it will actually make processes better.”
“There are some challenges when applying AI to real problems because of the limited access to data. AI only learns from publicly available data, so it doesn’t know specific areas well.”
“If you go back to the basics, for example, in manufacturing, the human expert can do much better.”
“I position myself in the bridging between academia and industry because methodologies in research articles are often not directly applicable to real problems.”
“A big trend of research is making models as large as possible, but it’s getting more difficult to understand the model. With my students, we try to make the model as simple as possible.”
“To apply AI in manufacturing, we should be able to interpret the models. The easiest way to do that is to make the model simple, but still accurate and powerful.”
