[Paper] Heterogeneous Model Alignment in Digital Twin

Source: arXiv - 2512.15281v1

Overview

The paper proposes a new way to keep the many different models that make up a digital twin (DT) in sync as they evolve. By combining adaptive conformance techniques with large‑language‑model (LLM)‑driven alignment, the authors show how to automatically discover and maintain semantic links across abstraction layers, reducing the manual effort that typically plagues DT projects.

Key Contributions

• Adaptive Conformance Mechanism: A runtime‑aware method that lets metamodels evolve together with the concrete models they describe, preserving semantic coherence.
• LLM‑Validated Alignment Process: Uses a large language model to ground model correspondences in domain knowledge, automatically generating and checking mappings.
• Scalable Multi‑Layer Alignment Framework: Works across heterogeneous model types (ontologies, SysML, BPMN, etc.) without requiring hand‑crafted static mappings.
• Empirical Validation on Real‑World Use Cases: Demonstrated on an air‑quality monitoring DT and benchmarked against several OAEI (Ontology Alignment Evaluation Initiative) tracks.
• Open‑Source Prototype: The authors release a prototype implementation that can be plugged into existing DT pipelines.

Methodology

1. Model & Metamodel Extraction – The DT’s constituent models (e.g., sensor data schemas, simulation models, business process diagrams) are represented as graph‑structured metamodels.
2. Adaptive Conformance Layer – A set of rules monitors changes in any model and propagates required updates to related metamodels, ensuring that the “contract” between a model and its metamodel stays valid.
3. LLM‑Driven Alignment –
   • A prompt‑engineered LLM (e.g., GPT‑4) receives pairs of model fragments plus domain glossaries.
   • It proposes candidate correspondences (class‑to‑class, property‑to‑property, etc.) and scores them for semantic similarity.
   • An automated validator checks structural consistency (e.g., cardinality, hierarchy) before committing the mapping.
4. Iterative Refinement – Misalignments detected during simulation or runtime trigger a feedback loop that re‑invokes the LLM with updated context, gradually improving alignment quality.
5. Evaluation – The framework is applied to:
   • An air‑quality DT (sensor network ↔ atmospheric model ↔ city‑level policy model).
   • Standard OAEI benchmark datasets (e.g., Anatomy, Conference, Knowledge Graph tracks) to compare precision/recall against state‑of‑the‑art ontology aligners.

Results & Findings

• Automation Gains: Manual mapping effort dropped from ~30 hours per model pair to <2 hours of initial setup, with subsequent updates handled automatically.
• Semantic Consistency: No observed violations of domain constraints (e.g., unit mismatches) after alignment, even when models were updated mid‑simulation.
• Scalability: Alignment time grew linearly with the number of model elements, making the approach viable for DTs with tens of thousands of entities.

Practical Implications

• Faster DT Deployment: Engineers can integrate new subsystems (new sensors, updated simulation kernels) without rewriting extensive mapping code.
• Reduced Maintenance Cost: The adaptive conformance layer automatically propagates schema changes, cutting down on error‑prone manual refactoring.
• Improved Decision Support: Consistent semantics across layers mean that predictive analytics and optimization algorithms receive trustworthy, harmonized inputs.
• Cross‑Domain Portability: Because the LLM is grounded in domain glossaries, the same alignment pipeline can be reused for manufacturing, smart‑grid, or healthcare DTs with minimal re‑configuration.
• Toolchain Integration: The prototype can be wrapped as a micro‑service, allowing existing DT platforms (e.g., Azure Digital Twins, Siemens MindSphere) to call the alignment API during CI/CD pipelines.

Limitations & Future Work

• LLM Dependency: Alignment quality hinges on the LLM’s underlying knowledge base; rare or highly specialized domains may need custom fine‑tuning.
• Explainability: While the LLM provides confidence scores, the reasoning behind specific mappings is not always transparent to engineers.
• Performance on Very Large Graphs: Experiments beyond ~50 k nodes showed increased latency; future work will explore graph‑partitioning and incremental LLM prompting.
• Standardization: The authors note the need for a common interchange format for “alignment metadata” to foster broader ecosystem adoption.

Overall, the paper offers a compelling blend of adaptive modeling and AI‑driven semantics that could make multi‑layered digital twins far more agile and reliable for real‑world deployments.

Authors

• Faima Abbasi
• Jean‑Sébastien Sottet
• Cedric Pruski

Paper Information

• arXiv ID: 2512.15281v1
• Categories: cs.SE
• Published: December 17, 2025
• PDF: Download PDF