TL;DR
Digital twins from different companies often rely on static "mappings" to translate data between them. This systematic review of over 5,800 records reveals a critical vulnerability: if one company changes its data definitions (semantic drift) during operation, current systems fail to detect or repair the broken connection, leading to silent, potentially catastrophic errors. The authors propose six formal requirements for "run-time repair" to bridge this gap.
What Problem Does This Paper Address?
Imagine a supplier’s digital twin reports "capacity" in units per 8-hour shift, and an assembler’s twin translates this into units per hour to plan production. If the supplier suddenly switches to 12-hour shifts or redefines "capacity" as units per hour without a formal notification, the assembler’s plans will be off by a factor of eight.
This is semantic drift. In a single company, a central authority can update all models simultaneously. However, in a federated system—where twins are owned by independent organizations—there is no central authority. Each party has limited visibility into the other's internal models. The authors tackle the challenge of how these independent systems can detect that a previously correct mapping has become invalid and repair it while the system is still running.
Methods
The authors conducted a systematic review following the PRISMA 2020 guidelines. They screened 5,833 records and performed a deep dive into 119 full texts, ultimately including 32 studies that met their criteria for federated digital twin interactions.
To evaluate these studies, they used the Levels of Conceptual Interoperability Model (LCIM). As shown in the table below, LCIM provides a scale from basic technical connectivity (L1) to dynamic, run-time changes in meaning (L5).
The review process itself was notable for its use of a Large Language Model (LLM) to assist in screening, with all machine decisions recorded and released for transparency. The selection flow is detailed in the PRISMA diagram:
Datasets
As a systematic review, the "dataset" consists of the existing body of academic literature. The authors searched databases including Scopus, IEEE Xplore, EBSCO, and OpenAlex for papers spanning 2015 to 2026. The final synthesis included 32 studies across diverse domains like manufacturing, smart cities, and supply chains.
Key Insights and Ideas
- The Masking Effect of Learning: A major "aha moment" is the critique of Multi-Agent Reinforcement Learning (MARL). The authors argue that a learner might simply adjust its policy to compensate for a drop in reward caused by drift, effectively "masking" the translation error rather than fixing the underlying mapping.
- The "Two Halves" Gap: The research found that the ingredients for a solution exist but are currently siloed. Cross-company matching (finding terms that mean the same thing) and single-company drift repair (detecting changes in meaning) have been demonstrated separately, but never together in a federated context.
- Neuro-Symbolic Architecture: The authors suggest a division of labor. Machine learning is great at spotting anomalies in data streams (the "neuro" part), while symbolic logic (knowledge bases and ontologies) is better suited for governing, validating, and storing the actual mapping rules (the "symbolic" part).
Key Results
The findings were stark: Zero out of the 32 included studies reached LCIM Level 5.
While many systems could parse data (L2) or use shared vocabularies (L3), every single one relied on mappings established before operation. If the meaning changed mid-run, the systems would continue to apply the old, now-incorrect rules.
To move the field forward, the authors defined six testable requirements for run-time repair, summarized here:
Limitations and Future Work
The review was limited to open-access full texts and specific databases, potentially missing proprietary or subscription-only breakthroughs.
For future work, the authors call for:
- Shared Benchmarks: The field needs datasets where controlled, unannounced semantic drifts are injected to test how quickly and accurately systems can recover.
- Repair under Partial Disclosure: Research into how twins can negotiate new mappings without exposing their entire internal proprietary models.
- Neuro-Symbolic Implementation: Developing and testing the proposed architecture that combines drift detection with formal governance of mapping records.
Summary updated