Heterogeneous Model Alignment in Digital Twin

Digital twin (DT) technology integrates heterogeneous data and models, along with semantic technologies to create multi-layered digital representation of physical systems. DTs enable monitoring, simulation, prediction, and optimization to enhance decision making and operational efficiency. A key challenge in multi-layered, model-driven DTs is aligning heterogeneous models across abstraction layers, which can lead to semantic mismatches, inconsistencies, and synchronization issues. Existing methods, relying on static mappings and manual updates, are often inflexible, error-prone, and risk compromising data integrity. To address these limitations, we present a heterogeneous model alignment approach for multi-layered, model-driven DTs. The framework incorporates a flexibility mechanism that allows metamodels to adapt and interconnect seamlessly while maintaining semantic coherence across abstraction layers. It integrates: (i) adaptive conformance mechanisms that link metamodels with evolving models and (ii) a large language model (LLM) validated alignment process that grounds metamodels in domain knowledge, ensuring structural fidelity and conceptual consistency throughout the DT lifecycle. This approach automates semantic correspondences discovery, minimizes manual mapping, and enhances scalability across diverse model types. We illustrate the approach using air quality use case and validate its performance using different test cases from Ontology Alignment Evaluation Initiative (OAEI) tracks.

💡 Research Summary

The paper addresses a fundamental challenge in multi‑layer, model‑driven Digital Twins (DTs): aligning heterogeneous artifacts—data, operational models, metamodels, and domain ontologies—across abstraction levels. Existing approaches rely on static, manually maintained mappings that are brittle and error‑prone, especially when models evolve during the DT lifecycle. To overcome these limitations, the authors propose a two‑pronged framework that combines (i) a flexible conformance mechanism for dynamic metamodel adaptation and (ii) a Large Language Model (LLM)‑validated alignment process that grounds metamodel elements in domain knowledge.

The flexible conformance mechanism is built on the JavaScript Modeling Framework (JSMF), a lightweight analogue of Eclipse Modeling Framework (EMF) that relaxes strict metamodel conformance through partial adherence, dynamic typing, and modular evolution. JSMF supports “proto‑metamodels,” on‑the‑fly customization, and posterior inference, allowing metamodels to evolve as new sensors, attributes, or relationships appear in the physical system. This enables bottom‑up change management: when a data schema changes, corresponding metamodel entities can be re‑instantiated without breaking existing models.

The second pillar leverages an LLM to automatically discover semantic correspondences between metamodel entities and concepts in a domain ontology. Using prompts that describe both sides, the LLM proposes candidate mappings, which are then filtered through structural checks (cardinality, attribute types) and semantic validation (e.g., SKOS exactMatch). The authors call this approach Semantics‑and‑Structure‑aware Metamodel‑Ontology Matching (SSM‑OM). By grounding metamodels in ontologies, the DT maintains both structural fidelity and conceptual consistency throughout its lifecycle.

The framework is demonstrated with an indoor air‑quality monitoring use case. A metamodel describing buildings, rooms, controllers, and various sensors (temperature, humidity, CO₂, pressure) is defined in JSMF. Real‑time data from ten sensors (2.5 million one‑minute samples) are streamed, parsed, and instantiated as model objects that conform to the metamodel. New sensor types can be added on the fly, and the LLM automatically aligns the corresponding metamodel classes with ontology concepts such as “AirQuality,” “Comfort,” and “Safety.” The system automatically triggers alarms when thresholds are exceeded, illustrating end‑to‑end DT operation.

To evaluate alignment quality, the authors benchmarked their approach against state‑of‑the‑art ontology alignment tools on several OAEI tracks (Anatomy, Conference, Knowledge Graph). Their method achieved F1 scores above 0.87, outperforming traditional tools especially on complex, multi‑inheritance ontologies. The results show that LLM‑driven matching can capture nuanced domain semantics that rule‑based systems miss, while the flexible conformance ensures that structural changes do not break the alignment.

Key contributions are: (1) a flexible modeling framework (JSMF) that supports both rigid and relaxed conformance across DT layers; (2) the SSM‑OM method that integrates LLM reasoning with formal semantic validation; (3) empirical validation on a realistic air‑quality scenario and systematic benchmarking on OAEF datasets. Limitations include dependence on LLM access (cost, latency), the need for carefully crafted prompts, and the current lack of an automated trigger for detecting metamodel drift in real‑time streams. Future work will explore lightweight, on‑premise LLM deployment, automated drift detection, and tighter integration with DT orchestration platforms.

In summary, the paper presents a novel, scalable solution for heterogeneous model alignment in Digital Twins, combining adaptive metamodel engineering with AI‑driven semantic grounding to achieve robust, automated, and maintainable DT ecosystems.