Beyond Ontology Standards: A LLM-Based Approach to Dynamic Semantic Interoperability in Federated Digital Twin Systems

Beyond Ontology Standards: A LLM-Based Approach to Dynamic Semantic Interoperability in Federated Digital Twin Systems

Author: Sebastiano Martignano, research strategy advisor -CRF Italy

Abstract

Semantic interoperability remains one of the principal challenges in federated Digital Twin (DT) environments. While reference ontologies such as SAREF, QUDT, and SSN/SOSA provide structured vocabularies for data exchange, their top-down design and static nature render them insufficient for handling the heterogeneity and dynamism of real-world distributed systems. This paper analyzes the limitations of current ontology-based approaches, reviews recent research on Large Language Models (LLMs) for ontology extraction, alignment, and transformation, and proposes a novel framework. In this framework, each node in a DT federation operates with its own local ontology—structured as a hierarchical object graph—while semantic interoperability is achieved via LLM-mediated dynamic mapping. This bottom-up, adaptive strategy avoids centralized standardization and enables scalable, evolutive, and decentralized digital infrastructures

Introduction

Digital Twins (DTs) have emerged as a transformative paradigm for monitoring, simulating, and optimizing physical systems in real time. Initially developed for isolated assets in manufacturing, DTs are now increasingly envisioned as part of federated networks representing distributed infrastructures—such as energy grids, water networks, logistics corridors, or agro-industrial systems. In such federated architectures, DTs interact as autonomous nodes while collectively modeling interdependent processes across multiple domains and territories.

This evolution, however, introduces a fundamental requirement: semantic interoperability. As nodes are independently developed and reflect divergent modeling conventions, system architectures, and vocabularies, the ability to share not only data but also meaning becomes critical. Traditional approaches to this problem rely on pre-defined ontologies and semantic standards, assuming that interoperability can be enforced by design. In practice, however, the real-world deployment of DTs rarely conforms to such top-down prescriptions.

This paper argues for an alternative strategy: allowing each node to maintain its own localized semantic structure and leveraging LLMs as mediators capable of dynamically translating and aligning ontologies at runtime. This opens the possibility of achieving interoperability by construction—through adaptive translation—rather than by standardization.

The Limits of Current Ontology-Based Approaches

Ontologies are widely used to define shared vocabularies and formalize knowledge structures in semantic systems. Initiatives such as SAREF (ETSI), QUDT (Quantities, Units, Dimensions and Types), and SSN/SOSA (W3C) offer well-structured frameworks for describing concepts, relationships, and units of measurement in domains ranging from smart homes to industrial automation.

These ontologies are typically expressed in RDF, OWL, or JSON-LD and rely on hierarchical class structures, inheritance, and logical constraints to support machine reasoning. They enable the semantic annotation of data streams, facilitate resource discovery, and underpin linked data architectures.

However, when applied to federated DT systems—composed of heterogeneous, evolving, and loosely coupled components—these ontologies face four critical limitations:

1. Top-down rigidity: Centralized, expert-driven design makes it difficult to adapt ontologies to local innovations, emergent behaviors, or domain-specific constraints.
2. Limited structural expressiveness: While most ontologies capture naming conventions and type hierarchies, they often lack mechanisms for modeling the internal structure and dynamic behavior of complex objects (e.g., actuators, process units, composite devices).
3. Semantic mismatch between nodes: Nodes may represent the same physical object with different naming conventions, nesting logics, or aggregation levels. For instance, one node might express humidity data under irrigationModule.sensorData.humidity, while another uses climate.h_rel.
4. Manual integration burden: Aligning ontologies across systems currently requires human experts, extensive schema mapping, and context-specific rule definition. This process is not scalable for large, dynamic networks of DTs.

These challenges highlight the need for new strategies that support semantic integration in flexible, distributed, and evolving environments.

LLMs and the Shift Toward Dynamic Ontology Mediation

Recent advances in natural language processing, particularly the development of Large Language Models (LLMs), offer a compelling alternative to static, rule-based semantic integration. Unlike traditional tools that depend on hardcoded mappings or predefined ontological rules, LLMs can interpret contextual meaning, identify semantic similarities, and generate structured outputs from natural language inputs.

Research from Google Cloud (2023) demonstrates that LLMs can be used to extract domain-specific concepts, relations, and properties from textual documentation and convert them into OWL or RDF-compatible class hierarchies. This approach enables the semi-automatic construction of ontologies from unstructured sources, significantly reducing the reliance on manual taxonomy engineering.

In another study, Xu et al. (2023) explore the use of LLMs in multi-agent DT systems for collaborative logistics. They show that inconsistencies between heterogeneous agents—each modeled independently—can be resolved through semantic translation mediated by LLMs, enabling inter-agent cooperation without centralized schema enforcement.

Furthermore, Zhou et al. (2024) propose embedding LLMs into federated learning architectures to enable dynamic semantic adaptation in edge computing contexts. Their work suggests that LLMs can continuously adjust semantic representations in response to context shifts, such as changes in mobility patterns, data availability, or operational constraints.

These developments collectively suggest that LLMs are not only capable of interpreting and generating ontologies, but can also act as runtime semantic mediators in distributed systems—mapping concepts, restructuring data objects, and aligning knowledge graphs dynamically.

Proposed Architecture: Local Ontologies and LLM-Based Semantic Mapping

This paper proposes an architecture in which each DT node maintains its own local ontology, developed autonomously in coherence with its specific physical configuration and control logic. These ontologies are represented as object-oriented graphs, where classes, attributes, and nested structures reflect the physical and operational reality of the system.

Rather than enforcing global standardization, interoperability is achieved through the mediation of an LLM that performs the following tasks:

• Conceptual alignment: Identifies semantic equivalences across node-specific terminologies, e.g., mapping umid1 to relativeHumidity or flowrate to volumetricFlow.
• Structural transformation: Translates nested or composite objects between different formats, enabling interoperability even when data schemas differ substantially.
• Context generation: Dynamically produces JSON-LD contexts or RDF graphs that map local terms to shared URIs based on public ontologies or inferred equivalence.
• Adaptive learning: Continuously refines mappings based on usage patterns, error signals, or user corrections, enabling ontological structures to evolve in practice.

This architecture allows nodes to remain semantically autonomous while enabling them to interoperate as part of a larger federated infrastructure. Interoperability becomes an emergent property of the system—constructed through interaction and mediated by intelligent translation—rather than imposed a priori.

Research Agenda and Open Challenges

The integration of LLMs into semantic interoperability frameworks introduces several open questions and defines a new research agenda for federated DT systems:

• Ontology graph interpretation: What is the optimal way to represent hierarchical object models such that LLMs can reason over their semantics?
• Robustness and verification: How can we ensure the accuracy, stability, and safety of LLM-generated mappings in critical infrastructure applications?
• Cross-domain generalization: To what extent can mappings be transferred or adapted across application domains (e.g., from agriculture to logistics)?
• Negotiation protocols: Can LLMs be embedded in multi-agent systems as active participants in semantic negotiation, contributing to the formation of emergent shared vocabularies?
• Human-in-the-loop design: How can expert knowledge and LLM inference be combined effectively to guide and validate semantic translations?

Addressing these challenges will require interdisciplinary collaboration across AI, knowledge representation, systems engineering, and domain-specific modeling.

Conclusion

Semantic interoperability is a foundational enabler for federated Digital Twin ecosystems. As these systems evolve from isolated platforms to distributed infrastructures, traditional ontology-based approaches prove increasingly inadequate. This paper has proposed a paradigm shift: enabling each node to operate with its own local ontology and leveraging LLMs as dynamic, context-aware mediators.

By doing so, we move from rigid standardization to adaptive construction—where meaning is not prescribed but negotiated through interaction. This opens the way for scalable, resilient, and intelligent infrastructures capable of collective modeling, simulation, and decision-making in complex socio-technical environments.

References

1. Ziemann, J., & Weidlich, M. (2022). Semantic Interoperability for Digital Twins: Challenges and Research Directions. ACM Computing Surveys, 55(10), Article 202.
2. San, O., Rasheed, A., & Kvamsdal, T. (2021). Hybrid analysis and modeling, eclecticism, and multifidelity computing toward digital twin revolution. arXiv preprint, arXiv:2103.14629.
3. Google Cloud AI. (2023). Large Language Models for Ontology Construction. Retrieved from https://medium.com/google-cloud/large-language-models-for-ontology-construction-fb2f751e3226
4. Xu, L., Mak, S., Schoepf, S., Ostroumov, M., & Brintrup, A. (2023). Multi-Agent Digital Twinning for Collaborative Logistics: Framework and Implementation. arXiv preprint, arXiv:2309.12781.
5. Zhou, Y., Fu, Y., Shi, Z., Yang, H. H., Hung, K., & Zhang, Y. (2024). Energy-Efficient Federated Learning and Migration in Digital Twin Edge Networks. arXiv preprint, arXiv:2503.15822.