PaperlayerView exact revision on arXiv

Source-linked AI summary

Digital Twin: Values, Challenges and Enablers

Adil Rasheed, Omer San, Trond Kvamsdal

arXiv:1910.01719v1eess.SP

TL;DR

Digital twins address the challenge of extending physical-asset modeling into synchronized, data-enabled monitoring, prediction, control, and optimization across the asset life cycle. The paper reviews applications, technologies, challenges, and stakeholder roles, concluding that hybrid physics-based and data-driven modeling is especially promising for robust digital-twin platforms.

  • Problem

    Digital-twin concepts span many applications, but building systems that remain synchronized with physical assets while supporting real-time use involves substantial technical and cross-domain challenges.

  • Method

    The paper reviews five diverse application areas, enabling technologies, digital-twin platforms, hybrid modeling approaches, and stakeholder recommendations.

  • Results

    The review identifies hybrid physics-based and data-driven modeling as particularly appealing for developing robust digital-twin platforms.

  • Takeaways & Limitations

    Digital twins can support real-time prediction, monitoring, control, optimization, and improved decision making throughout an asset’s life cycle and beyond.

Abstract

from arXiv · show

A digital twin can be defined as an adaptive model of a complex physical system. Recent advances in computational pipelines, multiphysics solvers, artificial intelligence, big data cybernetics, data processing and management tools bring the promise of digital twins and their impact on society closer to reality. Digital twinning is now an important and emerging trend in many applications. Also referred to as a computational megamodel, device shadow, mirrored system, avatar or a synchronized virtual prototype, there can be no doubt that a digital twin plays a transformative role not only in how we design and operate cyber-physical intelligent systems, but also in how we advance the modularity of multi-disciplinary systems to tackle fundamental barriers not addressed by the current, evolutionary modeling practices. In this work, we review the recent status of methodologies and techniques related to the construction of digital twins. Our aim is to provide a detailed coverage of the current challenges and enabling technologies along with recommendations and reflections for various stakeholders.

I. INTRODUCTION

The paper frames digital twins as synchronized virtual representations that use real-time physical-asset data for decisions, control, and hypothetical analysis. It reviews their values, application areas, challenges, enabling technologies, and stakeholder implications.

  • Digital twins connect physical assets with virtual representations through synchronized information, data, functions, and communication capabilities.
  • The paper asks what digital twins are, whether the concept is new, what value they provide, which challenges they face, and which technologies enable adoption.
  • Real-time asset data supports informed decision making and real-time optimization or control, while perturbed data supports offline what-if analysis, risk assessment, and uncertainty quantification.
  • The review covers eight value additions, state-of-the-art technologies across five application areas, common challenges, enabling technologies, socio-economic impacts, and stakeholder recommendations.

II. VALUE OF DIGITAL TWIN

Digital twins are presented as sources of operational, analytical, collaborative, and sector-specific value. Their benefits include remote control, safety, scenario analysis, improved teamwork, faster decisions, personalization, and applications in health care.

  • Digital twins enable remote monitoring and control, greater efficiency and safety, and what-if analyses for risk assessment without jeopardizing the real asset.
  • Digital twins can improve intra- and inter-team synergy and collaboration by giving teams access to information and greater autonomy.
  • Real-time quantitative data and analytics support more informed and faster decision making, while historical requirements and preferences support personalized products and services.
  • Five diverse applications are examined to understand digital-twin state of the art, common challenges, corresponding solutions, and future needs for greater physical realism.
  • Health care: Health-care digital twins are motivated by wearable devices, organized health data, personalized medication, and integration between engineering and medicine.
  • Health care: Human-body digital twins could support longitudinal data collection and surgical training, but technical, ethical, and privacy issues remain unresolved.

B. Meteorology

Meteorology is presented as a mature digital-twin application that combines physical models, high-fidelity simulators, heterogeneous observations, and data assimilation for weather prediction. Its remaining limitations are uneven observation coverage and the need for near-real-time assimilation and nature-informed modeling.

  • Meteorological digital twins combine regional solid models, high-fidelity physics simulators, and multisource big data to produce short- and long-term weather predictions.
  • Weather forecasting is described as a mature digital-twin concept, but service quality remains poor where weather stations or observation data are insufficient.
  • Large-volume, varied observations provide atmospheric and oceanic initial conditions for forecasting, while observing system simulation experiments mimic observation-analysis procedures.
  • Near-real-time forecasts support environmental awareness and decisions in air-traffic management, severe-weather monitoring, and wind-turbine adaptation.
  • Hybridization combines data-driven tools with established physics to provide nature-informed indicators for near-time events.
  • Synthetic observations from realistic nature runs are used to test new models and assimilation tools before operational adoption.

D. Education

The paper situates digital twins within broader smart-city, infrastructure, energy, transport, and industrial applications while emphasizing interoperability, security, dependability, sustainability, reliability, and predictability. It presents these cross-domain challenges as requiring enabling technologies.

  • Digital-twin technologies are becoming more mature for smart solutions in construction, transportation, and energy sectors.
  • Smart cities: Smart-city work pairs physical objects with digital counterparts and examines scalable cloud models for urban planning and city science.
  • Energy: Smart-grid research applies digital twins to power-system control and faster-than-real-time prediction of dynamic system behavior.
  • Transportation: Transportation and aviation studies use cyber-physical integration, mobile cloud computing, and cyberspace–physical-space integration to address industry challenges.
  • Industry: Industry 4.0 work applies digital-twin and related technologies to improve energy efficiency, production costs, greenhouse-gas emissions, and built-environment diagnostics.
  • Common challenges: Across application areas, major challenges include interoperability, security, dependability, sustainability, reliability, predictability, data management, latency, real-time simulation, and generalization.

V. ENABLING TECHNOLOGIES

The paper identifies five categories of enabling technologies for digital twins, beginning with physics-based modeling and its supporting experimental, geometric, and numerical techniques.

  • Enabling technologies are organized into physics-based modeling, data-driven modeling, big data cybernetics, infrastructure and platforms, and human-machine interface.
  • Physics-based modeling: Because assumptions and incomplete understanding enter at multiple stages, physics-based models capture only part of the known physics.
  • Physics-based modeling: Physics-based modeling starts from observing phenomena, developing partial understanding, expressing it mathematically, and solving the resulting equations.
  • Physics-based modeling: Experimental modeling uses laboratory or full-scale experiments and surveys to develop correlations or models for quantities that are difficult or expensive to measure directly.
  • Physics-based modeling: Three-dimensional models are the starting point for most digital twins and provide inputs to physical simulators, making geometric quality crucial.
  • Physics-based modeling: Locally Refined B-splines can produce compact object-geometry representations and high data-compression ratios useful for digital twins.

3) High Fidelity Numerical Simulators:

High-fidelity numerical simulators provide physical realism and support predictive digital twins, but their computational demands constrain operational use. The paper also presents data-driven modeling as a complementary approach enabled by abundant sensor data and computational resources, while emphasizing data-quality challenges.

  • High Fidelity Numerical Simulators: Governing physical equations add realism to digital twins but usually require numerical computer solutions because of their complexity.
  • High Fidelity Numerical Simulators: Physics-based models are interpretable and broadly generalizable, but can be numerically unstable, computationally demanding, and sensitive to modeling and input uncertainty.
  • High Fidelity Numerical Simulators: High-fidelity simulators have mainly supported design because their computational efficiency must improve by several orders of magnitude for fuller digital-twin use.
  • High Fidelity Numerical Simulators: High-fidelity models can support predictive digital twins through reduced-order models, while advanced simulators already enable patient-specific cardiovascular treatment models.
  • Data-driven modeling: Data-driven modeling treats data as a manifestation of known and unknown physics and has become popular with abundant data, accessible libraries, inexpensive computing, and training resources.
  • Data-driven modeling: Digital-twin data requires improved quality, online compression, outlier detection, missing-data filling, and stronger management and ownership resolution.

3) Data Privacy and Ethical Issues:

The paper presents digital-twin enabling technologies spanning secure information systems, machine learning, big data cybernetics, and hybrid modeling. These approaches address monitoring, learning from data, control, and the integration of physics-based and data-driven models.

  • Blockchain provides a distributed, secured, verifiable, traceable, and transparent record suited to digital twins where information security is critical.The paper links this role especially to safety-critical applications and privacy concerns.
  • Machine learning methods support digital twins through supervised, unsupervised, and reinforcement learning for prediction, classification, anomaly detection, and data quality improvement.Examples include resolution upscaling, denoising, missing-data completion, time-series modeling, and forecasting.
  • Data-driven models improve as more experience data are supplied, but complex models can be unstable during training, uninterpretable, and biased by their training data.These limitations are especially consequential in safety-critical applications.
  • Big data cybernetics combines more complex physics-informed controllers with data-based estimation of quantities that cannot be measured directly.Its workflow continuously loops as new data streams become available.
  • Hybrid Analysis and Modeling combines physics-based interpretability and robustness with data-driven accuracy, efficiency, and automatic pattern identification.The paper places HAM at the intersection of big data, physics-based modeling, and data-driven modeling.

1) Data Assimilation:

Data assimilation combines computational models with observations to estimate the state of a physical system. In operational forecasting, it initializes models by statistically combining forecasts and observations while accounting for uncertainty.

  • Data assimilation estimates a physical system’s state by fitting computational models to observations.The paper also relates this process to filtering, estimation, smoothing, and prediction in engineering.
  • Operational atmospheric data assimilation initializes forecast models by combining short-range forecasts with observations and their estimated uncertainties.Methods differ in how they treat background error and solve the analysis equations.
  • Reduced-order models represent high-dimensional scientific problems with embedded low-dimensional structures, trading some accuracy for substantially greater computational speed.They can emulate complex full-order models or processes at the interface of data and domain-specific science.
  • Nonintrusive data-driven reduced-order models avoid requiring exact governing equations and can protect exchanged data, executables, and intellectual property.They can use experimental, sensor, and large-scale simulation data when detailed governing equations are unknown.
  • Hardware-in-the-loop simulation is presented as a route toward near-real-time predictions within digital-twin platforms.

4) Other Hybridization Techniques:

The paper reviews hybridization techniques that combine physical knowledge, data, and machine learning to improve efficiency, interpretability, or model construction. These include physics-informed learning, domain-guided feature engineering, hybrid physics-AI models, adaptive solution spaces, and sparse or symbolic regression.

  • Replacing equations with machine learning can speed repeated-process simulation, but interpolation-based models fail in unexpected situations because they do not extrapolate.The approach is therefore scoped to cases where generalization is not necessary.
  • Hybrid physics-AI models represent known physics with governing equations and unknown physics with black-box DNN or LSTM networks.The equation-based component supplies a sanity check for unexpected black-box behavior.
  • Domain-informed feature engineering can reduce algorithmic complexity while improving accuracy by supplying features learned from scientific knowledge.The paper cites material-discovery applications as examples.
  • Hybrid approaches combine deterministic and machine-learning components in general circulation models for climate modeling and weather prediction.
  • Adaptive nonparametric probabilistic reduced-order methods use data to adapt the solution subspace rather than only adjusting computational-model parameters.The approach combines physics-based modeling and data with a mathematical and statistical foundation.
  • Physics-informed learning addresses challenges including physical-law incorporation, interpretability, nonlinearities, conservation properties, and massive training-data requirements.One cited strategy regularizes a machine-learning cost function using residuals of governing equations.
  • Sparse regression recovers nonlinear partial differential equations from candidate feature terms, while symbolic regression searches for governing structures from data when accurate physical laws are unavailable.

D. Infrastructure and platforms

Digital-twin infrastructure combines cloud, edge, fog, IoT, and data-management platforms to store, process, integrate, and analyze large volumes of sensor and business data across applications.

  • Data infrastructure: Big-data infrastructure provides blending, integration, storage, centralized management, interactive analysis, visualization, accessibility, and security for high-volume data.Hadoop can execute tasks where the data are hosted, avoiding copies to local memory.
  • Cloud and IoT platforms: IoT platforms connect sensors and devices to cloud services for real-time monitoring, processing, control, and domain-specific applications.Platforms discussed include Cisco, Bosch, IBM Bluemix, AWS, Microsoft Azure, SAP, Salesforce, and Oracle.
  • Cloud and IoT platforms: Cloud platforms support IoT data storage, management, analysis, and application deployment across healthcare, agriculture, transportation, energy, and smart-home systems.Examples include biomedical genome mapping, environmental sensor networks, smart homes, and data-driven agriculture.
  • Cloud and IoT platforms: Software-as-a-service and platform-as-a-service models improve scalability, maintenance, management, and resource costs for IoT applications.Salesforce IoT is described as SaaS, while Oracle IoT is described as PaaS.
  • Challenges: Security and privacy remain major implementation challenges for IoT devices and architectural models.The discussion specifically identifies industrial security threats and privacy concerns in IoT deployments.

3) Communication Technologies:

Digital twins depend on communication technologies that deliver information among distributed components with sufficiently low latency, while expanding device connectivity creates spectrum, interference, and scalability challenges.

  • Communication requirements: Reliable digital-twin operation requires information from different components to reach its intended target on time.The paper illustrates this requirement with robotic surgery, where digital actions should manifest in reality without latency.
  • 5G and wireless technologies: 5G, millimeter waves, small cells, Massive MIMO, and beamforming are presented as technologies for supporting many connected devices and correcting interference.5G uses higher frequencies, while small cells and antenna arrays address coverage and capacity requirements.
  • Alternative networking: LoRaWAN is gaining attention for IoT applications because it keeps network structures and management simple.
  • Alternative networking: Unmanned aerial systems can assist ground base stations during crowded events, disasters, and other temporary situations requiring additional network resources.UAS-based sensing has also supported in-situ atmospheric observations for weather forecasting.
  • Computational communication: Communication becomes more challenging as processing elements increase or the number of grid points per element decreases, creating a bottleneck for exascale systems.The paper connects this challenge to the limits of current transistor scaling and increasing architectural heterogeneity.
  • Distributed computing: Cloud, edge, and fog computing distribute data storage, management, and processing between remote servers and network nodes.Fog computing is described as a blend of cloud and edge computing.

5) Digital Twin Platforms:

Digital-twin platforms combine cloud services, data integration, physics-based and machine-learning models, multiphysics tools, and real-time asset data to support simulation, optimization, and prediction.

  • Industrial platforms: KognifAI combines cloud applications for data accessibility and processing with cybersecurity, identity, and encryption capabilities.Kongsberg provides the platform for energy, oil and gas, and maritime industries through PaaS and SaaS services.
  • Model-based platforms: MapleSim creates dynamic machine models from CAD data, including forces and torques, to support digital-twin development and machine design.Its digital-twin module aims to implement virtual plant models without necessarily requiring expert knowledge.
  • Model-based platforms: The initial motor size for a screwing machine was undersized by a factor of 10, a discrepancy identified through a MapleSim digital twin.The paper states that this would have caused machine failure and excessive losses.
  • Industrial platforms: Cognite Data Fusion extracts useful information from industrial data and provides open-source APIs, SDKs, and libraries for developers and analysts.Its integration with Siemens information management benefited Aker BP by optimizing offshore maintenance and reducing costs.
  • Model-based platforms: ANSYS Twin Builder supports multidomain systems, multiple fidelities, multiphysics solvers, reduced-order models, third-party integration, embedded software, and system optimization.The platform was also used to build a pump digital twin using real-time sensor data for performance improvement and prediction.
  • Industrial platforms: Oracle IoT Cloud organizes digital-twin services into virtual-twin, predictive-twin, and twin-projection pillars.Predictive twins may use FEM/CFD or statistical and machine-learning models, while projections connect insights to backend applications.

1) Augmented and Virtual Reality:

AR, VR, natural-language, and gesture interfaces are presented as technologies for richer human–digital-twin interaction, supported by advances in sensing, deep learning, and communication.

  • AR and VR: AR and VR can provide new perspectives for digital-twin interaction across engineering, design, medicine, and education.The paper places these technologies among key tools for detailed visualization of physical assets.
  • Natural-language interaction: Voice interaction is identified as a fast human communication mode that could support seamless integration between people and machines.The paper links this opportunity to recent advances in deep learning and LSTM-based language translation and interpretation.
  • Gesture interaction: Gesture recognition can use RF and millimeter-wave radar, ambient light, cameras, sound, and wearable devices.The paper connects these sensing technologies with more robust digital-twin systems alongside faster IoT communication.
  • Socio-economic considerations: Digital-twin automation raises workforce acceptability concerns, while human–machine task allocation is presented as a way to support safety and creativity.The paper emphasizes keeping humans involved in coordinating AI developments and checking AI results.

VII. CONCLUSIONS AND RECOMMENDATIONS

The paper defines a digital twin as a data- and simulator-enabled virtual representation supporting real-time prediction, monitoring, control, and optimization across an asset’s life cycle. It organizes the concept into virtual, predictive, and projection pillars, then recommends coordinated contributions from stakeholders while highlighting realism, standardization, and hybrid modeling challenges.

  • A digital twin is a virtual representation of a physical asset enabled by data and simulators for prediction, monitoring, control, and optimization throughout its life cycle.
  • The framework comprises a Virtual Twin, a Predictive Twin, and Twin Projection, which integrates predictive insights into business operations and processes.
  • Stakeholder recommendations: Industry can advance adoption by sharing asset datasets, contributing practical knowledge to research, and projecting predictive-twin insights into business applications.
  • Stakeholder recommendations: Academia and research institutes are expected to develop enabling technologies, release them as open-source software, and train interdisciplinary teams spanning application knowledge, mathematics, and computer science.
  • Stakeholder recommendations: Government, funding agencies, and society are encouraged to support inclusive regulation, open public data, open-source infrastructure, and new skills for broad and ethical adoption.
  • Challenges and enablers: The paper identifies limited physical realism and missing cross-domain standards as important boundaries, while proposing hybrid physics-based and machine-learning modeling as a route toward more realistic digital twins.
  • Conclusion: Digital twins offer perspectives across fields, and hybrid analysis and modeling can support more informed decisions in applications involving big data cybernetics, security, digitalization, automatization, and intelligentization.
Loading 1910.01719v1…