Generalized graph foundation models as versatile data-driven digital twins for complex technological systems

Area of Science:

• Engineering
• Computer Science
• Data Science

Background:

• Traditional physics-based digital twins (pbDT) are complex to build and may not capture real-world system degradation.
• pbDTs often require detailed physics knowledge and struggle with systems operating off-specification.
• Existing methods lack flexibility for diverse systems and tasks.

Purpose of the Study:

• To present a unified pipeline for constructing data-driven Foundation Models (ddDTs).
• To demonstrate the adaptability of this approach across diverse systems like solar fleets and additive manufacturing.
• To offer a flexible and agile alternative to pbDTs for real-world system assessment.

Main Methods:

• Utilized spatiotemporal graph neural networks (st-GNNs) for a flexible, data-driven modeling architecture.
• Employed a self-supervised learning approach with a reconstruction objective to train encoder modules.
• Developed a unified pipeline applicable to solar-photovoltaic fleets, direct-ink-write, and laser-powder-bed-fusion manufacturing.

Main Results:

• Successfully constructed data-driven Foundation Models for three distinct systems using a single methodology.
• Demonstrated the capability of ddDTs to implicitly capture complex physical phenomena from operational data.
• Showcased the modularity and reusability of the trained encoder as a Foundation Model.

Conclusions:

• Data-driven Foundation Models provide an agile, flexible, and unified approach to digital twin development.
• This methodology reduces the reliance on detailed physics-based modeling and facilitates faster system assessment.
• The proposed pipeline enables scientists to focus on scientific objectives rather than complex modeling intricacies.