Multi-Fidelity Digital Twin Framework for Dynamic Data-Driven and Adaptive Intelligence in Cyber-Physical Systems

Abstract

Digital Twin (DT) technology has emerged as a promising paradigm for cyber-physical systems, yet a substantial gap remains between its conceptual vision and its practical realization in data-constrained, evolving, and security-sensitive industrial environments. In many formulations, the DT is implicitly treated as a high-fidelity virtual counterpart that should replicate the dynamics of the physical twin (PT) with maximal precision. However, such an assumption is neither universally practical nor necessary, since achieving extremely high fidelity often requires dense sensing, fine-grained spatiotemporal resolution, and computationally intensive modeling, resulting in substantial costs in computation, energy, and system complexity that may exceed operational requirements and hinder scalability and real-time responsiveness. This dissertation addresses that limitation by developing a multi-fidelity DT framework for cyber-physical systems that treats fidelity as a design dimension rather than a fixed property, and systematically organizes DT representations across multiple levels of fidelity according to task requirements, available data, and computational constraints. In doing so, the proposed framework enables principled and scalable support for data-driven modeling, adaptive system understanding, security-aware monitoring, and downstream human-centered applications under realistic Industry 4.0 conditions. The dissertation first establishes a DT reference model and a multi-fidelity DT framework that organizes the relationships among physical space, virtual space, and system-level indicators, and positions different levels of fidelity as complementary forms of system representation rather than competing alternatives. Building on this foundation, the dissertation develops two distinct DT methodologies with different operational roles. The low-fidelity DT is designed for manufacturing security and operational monitoring, where MLOps supports model construction, deployment, and lifecycle maintenance for scalable anomaly-detection services in cyber-manufacturing environments. In contrast, the medium-fidelity DT is designed for adaptive behavior modeling under limited data and evolving observations. To support this objective, the dissertation introduces the Physical Twin Observation Graph (PTOG), a structured representation that provides the DT with context-aware capabilities for state organization and analysis. Building on this representation, the framework leverages Generative AI to perform data-driven and zero-shot behavior prediction, enabling the DT to reason over physical-system observations, model evolving process dynamics, and maintain alignment with the changing physical system over time. The dissertation then extends the framework beyond DT-centered operations into education and workforce development through a multi-fidelity DT-for-education model, immersive DT-based learning environments, large language model(LLM)-based zero-shot and probabilistic affective analytics, and LLM-based personalized guidance. In parallel, it investigates auxiliary knowledge discovery to enhance reliable DT systems through LLM-assisted mining of large-scale cybersecurity corpora, identifying 1,742 hardware-related CVEs, 411 of which contributed to the set of 1,026 CVEs used in the MITRE Most Important Hardware Weaknesses (MIHW) 2025 workflow. Finally, the dissertation discusses how these low- and medium-fidelity DT methodologies provide a structured pathway toward high-fidelity digital twins, where richer physical-system integration, stronger adaptive capabilities, and more comprehensive cross-layer intelligence can be progressively achieved as sensing, modeling, and computational resources increase. In summary, this dissertation advances DT methodology from static digital mirroring toward a structured, multi-fidelity, data-driven, and extensible foundation for intelligent cyber-physical system support across monitoring, prediction, security, and human-centered applications.