PRISM-AI

We’re pleased to share a presentation and perspective introducing the vision and strategy of the PRISM-AI InnoCORE Research Center, which aims to advance AI-driven innovation in smart manufacturing. 

- Presentation slide pdf (in Korean): https://lnkd.in/grDaBT7b 

- Perspective preprint available on arXiv: https://lnkd.in/gXH4wwzB

Summary)

[1] AI models can be viewed as functions mapping from the design space (domain) to the performance space (codomain), and under the Universal Approximation Theorem (UAT) and Gaussian Process Regression (GPR), sufficient predictive accuracy can be achieved with well-trained surrogate models.

[2] Predictive performance depends heavily on how uniformly and strategically the training data covers the design space, and addressing poor generalization outside the training distribution remains a key challenge.

[3] Therefore, problems in material, structural, and process design can all be addressed under a unified mathematical framework—yet under limited time and budget, expert-guided sampling strategies are essential.

[4] Our current research focuses on data-efficient physics-informed machine learning (PIML) to address limited resources and improve extrapolation performance beyond the training distribution, as well as on developing user-friendly LLM-based tools for domain experts unfamiliar with AI. These approaches are being applied to real-world use cases such as injection molding control and nuclear reactor operation.

[5] In the mid-term, our core vision is to build a next-generation design-to-manufacturing platform that integrates PIML, multi-agent AI LLM systems, and intuitive graphical user interfaces (GUIs).