On the scaling of digital twins by aggregation

Marcelo Pias, Lawrence A. Bull, Daniel S. Brennan, Mark Girolami, Jon CrowcroftData 2025 · DOI

TL;DR

This paper proposes a conceptual framework for scaling Digital Twins (DTs) from single assets to planetary systems through a "hierarchy of scales" and modular aggregation. The authors argue that the future of DTs lies in a decentralized, peer-to-peer approach where complexity emerges from the connections between interpretable, domain-specific units rather than from a single, monolithic model.

What Problem Does This Paper Address?

Digital Twins—virtual replicas of physical objects or systems—are moving beyond their origins in engineering asset management (like monitoring a single engine) toward representing massive, interconnected systems like smart cities or global energy grids.

The authors identify a major challenge: as DTs scale, they become "systems of systems." Traditional model-centric approaches struggle with this complexity, leading to issues with data silos, lack of interoperability, and massive privacy risks. There is a need for a unified methodology that handles the hierarchical aggregation of knowledge while balancing physics-based "process understanding" with modern data-centric AI.

Methods

The paper introduces a framework built on Three Pillars:

  1. Emulation (The Mirror): The DT must evolve in parallel with the real world, allowing for "what-if" scenario testing.
  2. Evolution (Data Pathways): Overcoming data management obstacles to ensure a continuous flow of information between the physical and virtual.
  3. Aggregation (Modular Knowledge): Building complex DTs by connecting smaller, interpretable units.

To organize this, they propose a Hierarchy of Scales:

  • System-specific: A single asset (e.g., one wind turbine).
  • Local collection: Connected IoT devices (e.g., a wind farm).
  • Ecosystems: Interacting engineered and natural systems (e.g., a regional energy grid).
  • Planetary: Distributed, parallel DTs addressing global challenges (e.g., Earth-scale climate modeling).

Technically, the authors advocate for Data-centric ML, which prioritizes high-quality, curated data and domain expertise over simply tuning complex neural architectures. They also suggest using Knowledge Transfer techniques (like federated learning and multi-view learning) to share insights across the hierarchy without compromising data ownership.

Datasets

As this is a Commentary paper rather than a primary experimental study, it does not introduce new datasets or benchmarks. Instead, it references several real-world use cases to illustrate its points:

  • Wind Energy Systems: Used to demonstrate the progression from a single turbine DT to an international energy ecosystem.
  • Smart Energy Meters: Used to discuss privacy implications across different scales of data aggregation.
  • Project Odysseus: A reference to a London-based project that used data-driven models to monitor activity during COVID-19 lockdowns.

Key Insights and Ideas

  • Complexity is in the Connections: The authors emphasize that the complexity of a large-scale DT emerges from the framework of connections between modules, not just the complexity of the modules themselves.
  • The "Sweet Spot": Successful DT design requires finding the intersection where process-based models (physics/logic) and data-centric models (ML) work together.
  • Decentralization is Mandatory: For planetary-scale DTs, a centralized server is a bottleneck and a privacy nightmare. A peer-to-peer (P2P) or "logic data highway" approach is proposed to handle jurisdictional and ownership issues.
  • Industry 5.0 and the Human-in-the-Loop: Unlike the automation-heavy Industry 4.0, the authors argue that DTs should augment human intelligence ("cobots"), especially when dealing with unquantifiable ethical, legal, or cultural constraints.

Key Results

The paper provides a qualitative roadmap for the evolution of DT technology:

  • Privacy Framework: It maps specific privacy-enhancing technologies (like differential privacy and homomorphic encryption) to the hierarchy of scales, showing that as knowledge increases through aggregation, the risk of residual personal data exposure also grows.
  • Ethical Design Issues (EDIs): The authors identify five core EDIs that must be addressed: (1) cumulative uncertainty, (2) the intricacy of optimization at scale, (3) the responsibility of representing human societies, (4) the necessity of FAIR (Findable, Accessible, Interoperable, Reusable) data sharing, and (5) the need for transparency in AI.

Limitations and Future Work

  • Technical Debt: The authors acknowledge that while a "clean-slate" decentralized architecture is ideal, the massive existing investment in current siloed systems may make this financially and practically difficult.
  • Interoperability Standards: There is a critical need for open standards and protocols to allow DTs from different countries and organizations to "talk" to each other.
  • Uncertainty Propagation: A major remaining challenge is how to accurately quantify and represent the cumulative uncertainty that builds up as many small models are aggregated into a planetary-scale system.
  • Validation at Scale: While the P2P/decentralized paradigm is theoretically sound for privacy, it has yet to be fully validated at the "Planetary" scale.

Summary updated

Abstract

Abstract We discuss the emerging technology of digital twins (DTs) and the expected demands as they scale to represent increasingly complex, interconnected systems. Several examples are presented to illustrate core use cases, highlighting a progression to represent both natural and engineered systems. The forthcoming challenges are discussed around a hierarchy of scales , which recognises systems of increasing aggregation. Broad implications are discussed, encompassing sensing, modelling, and deployment, alongside ethical and privacy concerns. Importantly, we endorse a modular and peer-to-peer view for aggregate (interconnected) DTs. This mindset emphasises that DT complexity emerges from the framework of connections (Wagg et al. [2024, The philosophical foundations of digital twinning, Preprint ]) as well as the (interpretable) units that constitute the whole.
BibTeX
@inproceedings{pias2025scaling,
  author = {Marcelo Pias and Lawrence A. Bull and Daniel S. Brennan and Mark Girolami and Jon Crowcroft},
  title = {{On the scaling of digital twins by aggregation}},
  booktitle = {Data},
  year = {2025},
  doi = {10.1017/dap.2024.86},
}
Figure