{
  "id": "digital-twin",
  "title": "Digital twin",
  "lang": "en",
  "source_lang": "en",
  "summary": "A digital twin is a data-connected representation of a physical entity or process used to understand, predict, and support its behavior throughout its life cycle.",
  "category": "technology",
  "subcategory": "engineering",
  "type": "concept",
  "aliases": [
    "Digital counterpart",
    "Virtual twin"
  ],
  "generated_at": "2026-11-05",
  "model": "pro/gpt-6.1-sol",
  "url": "https://aiwiki.page/en/digital-twin/",
  "translations": {
    "zh-Hans": {
      "url": "https://aiwiki.page/zh/digital-twin/",
      "json": "https://aiwiki.page/zh/digital-twin.json"
    }
  },
  "sections": [
    {
      "heading": "",
      "text": "A digital twin is a digital representation of an identifiable physical entity, process, or system that is maintained through exchanges of data with its real-world counterpart. It combines information about what the counterpart is, its current or historical condition, and models of how it behaves. Depending on its purpose, a twin may represent an individual machine, a manufacturing process, a building, a transport network, or a larger interconnected system.\n\nThe distinctive feature is not photorealistic visualization but a continuing relationship between representation and represented system. Measurements, maintenance records, operating conditions, and configuration changes can update the representation; analysis of that representation can then support inspection, planning, or control. A twin may therefore provide capabilities extending beyond a conventional drawing, database, or stand-alone simulation, although it commonly incorporates all three.\n\nDigital twins bring together developments in sensing, computer-aided design, numerical modeling, the Internet of Things, and industrial information systems. Their usefulness depends on whether the representation is accurate and timely enough for a particular decision—not on whether it reproduces every detail of reality. Applications range from estimating the remaining life of a rotating component to evaluating changes in a factory or assessing how an urban drainage system could respond to heavy rainfall."
    },
    {
      "heading": "Definition, boundaries, and related concepts",
      "text": "Definitions of a digital twin differ across engineering disciplines and standards. A common core consists of a real-world referent, a digital representation, and mechanisms that keep relevant aspects of the representation aligned with the referent. The alignment may concern geometry, operating state, material condition, configuration, or some combination of these. An important qualification is that synchronization occurs at a frequency and level of fidelity appropriate to the application; it does not necessarily mean continuous, instantaneous updating.\n\nA useful distinction separates a generic model from a representation of a particular instance. A model of a turbine design describes a class of machines. A twin associated with one installed turbine can additionally include its serial number, manufacturing deviations, operating history, repairs, and inferred damage. Some usage also extends “digital twin” to designs before a physical object exists, especially when the same representation will later be connected to production and operational data. In that case, the connection to a future physical instance is prospective rather than already established.\n\nOne influential classification distinguishes a digital model, updated mainly through manual exchanges; a digital shadow, receiving automated data from the physical system; and a digital twin, with automated exchanges in both directions. This terminology is useful but not universal. Many accepted industrial twins provide monitoring and recommendations without issuing commands to equipment. Bidirectional exchange may also involve approved work orders or revised operating plans rather than direct automatic control.\n\nA twin is likewise not synonymous with a three-dimensional scene. A thermal network, a set of differential equations, or a statistical representation may be more useful than detailed geometry. Conversely, an immersive representation of a factory may show equipment accurately while having little information about its condition or performance. Visualization is an interface to a twin, not sufficient evidence that one exists.\n\nSeveral neighboring concepts overlap without being identical. Building information modeling organizes information about built assets, while product life-cycle management manages product-related information and processes. Either may supply a twin’s identity, configuration, and engineering records. A simulation investigates behavior under specified assumptions; a twin may run many simulations while also maintaining a persistent connection to a particular system. A dashboard displays observations, whereas a twin can additionally relate those observations to structure, inferred states, and possible futures. In practice, boundaries are often determined by the intended use and the responsibilities assigned to the system."
    },
    {
      "heading": "Historical development and standardization",
      "text": "Digital twins developed from several older engineering practices rather than a single invention. Engineers have long used scale models, mathematical models, test rigs, and operational records to understand machines that cannot be fully observed in service. The Apollo program, for example, used ground-based simulators and extensive telemetry to support spaceflight operations. Such arrangements are often described retrospectively as precursors, but they should not be treated as identical to contemporary, network-connected digital twins.\n\nThe modern concept is closely associated with Michael Grieves’s work on product life-cycle management at the University of Michigan beginning in 2002. That work described a physical space, a virtual space, and information links between them. Terminology evolved as the idea spread. John Vickers of NASA is commonly associated with the later adoption of the expression “digital twin.” These contributions helped frame the twin not simply as a computer model but as an enduring relationship between physical and informational representations.\n\nAerospace research gave the concept a particularly demanding technical interpretation. NASA’s 2010 modeling and simulation roadmap discussed digital twins, and a widely cited 2012 paper by Edward Glaessgen and David Stargel described a digital-twin paradigm for future NASA and United States Air Force vehicles. Its emphasis was on integrating high-fidelity simulation, onboard measurements, and historical information to reflect the condition of an individual vehicle. Fatigue, damage accumulation, and uncertain operating loads were central concerns: two nominally identical aircraft can develop different conditions because they experience different histories.\n\nDuring the 2010s, cheaper sensing, improved connectivity, cloud computing, and advances in numerical methods broadened the concept’s industrial reach. Manufacturing initiatives connected it to production equipment and enterprise information systems. Building and infrastructure applications connected it to spatial models, inspection records, and asset-management systems. The expression also spread into marketing, where it sometimes became a loose label for almost any digital representation. This expansion made precise statements about connectivity, scope, and validation increasingly important.\n\nStandardization has sought to make implementations more comparable. The ISO 23247 series, first published in 2021, establishes a digital-twin framework for manufacturing, addressing general principles, reference architecture, digital representations of manufacturing elements, and information exchange. Such standards do not prescribe a single universal model or software product. Rather, they provide structures within which manufacturers and technology providers can define components and interfaces. Industry-specific standards remain important because equipment identity, engineering meaning, safety obligations, and acceptable uncertainty differ substantially between a machine tool, an aircraft, and a municipal water network."
    },
    {
      "heading": "Architecture, identity, and information exchange",
      "text": "A digital-twin architecture usually contains several interacting layers: physical equipment and instrumentation; communication and data collection; persistent information storage; models and analytical services; and interfaces through which people or other systems use the results. These functions need not reside in one application. A practical implementation may combine embedded software, local servers, remote computing services, maintenance databases, and specialist simulation tools.\n\nThe foundation is reliable identity. Every observation must be associated with the correct asset, component, location, and configuration. Replacing a motor, changing a sensor, or updating a controller can alter the meaning of later measurements. A twin therefore needs more than an equipment name: it needs records of component relationships and their validity over time. Without configuration control, a model may interpret data using the characteristics of a component that is no longer installed.\n\nOperational information commonly arrives from programmable logic controllers, supervisory control and data acquisition systems, inspection instruments, and business systems. OPC Unified Architecture can convey industrial data together with structured descriptions, while an application programming interface may expose maintenance histories or simulation services. Communication protocols move information, but semantic agreement determines whether different systems interpret it consistently. A value labeled “pressure,” for example, remains ambiguous without its unit, measurement location, reference convention, and timestamp.\n\nTime handling is especially important. A sensor timestamp may record when a quantity was sampled, while a database timestamp records when the message arrived. Networks can delay, duplicate, or reorder messages. Slowly changing configuration records must also be reconciled with rapidly sampled signals. A twin intended for accident reconstruction may require a different temporal resolution from one used for annual capital planning. Storing provenance—the origin and transformation history of information—helps distinguish direct measurements from calculated, corrected, or manually entered values.\n\nArchitecture also reflects physical constraints. Computation close to equipment can reduce communication delays and allow operation during network outages. Centralized resources can combine information across a fleet or execute computationally expensive analyses. The choice is therefore not simply between local and remote computing; many twins distribute work across both. Critical protective functions generally remain within appropriately engineered local systems rather than depending on an external analytical service.\n\nInteroperability extends beyond file exchange. Two models must agree about coordinate systems, asset boundaries, units, variable definitions, and the conditions under which results are valid. A persistent engineering information chain, often called a digital thread, helps connect requirements, design revisions, manufacturing records, and operational evidence. It supports the twin, but does not itself guarantee that behavioral models remain accurate."
    },
    {
      "heading": "Behavioral models and computational methods",
      "text": "A twin’s models translate observations and assumptions into descriptions of behavior. Physics-based approaches use governing relationships such as conservation of mass, momentum, and energy. A heat exchanger twin, for example, may combine mass balances and an energy balance with heat-transfer relationships to estimate outlet temperatures and detect declining performance. A structural twin may use finite-element analysis to estimate stresses from measured loads. The appropriate model depends on the decision: predicting overall energy consumption requires different detail from locating a small fatigue crack.\n\nA general discrete-time formulation is\n\n\\[x_{k+1}=f(x_k,u_k,\\theta)+w_k, \\qquad y_k=h(x_k,u_k,\\theta)+v_k,\\]\n\nwhere \\(x_k\\) denotes an internal state, \\(u_k\\) an input or operating condition, \\(\\theta\\) model parameters, and \\(y_k\\) measured outputs. The terms \\(w_k\\) and \\(v_k\\) represent process and observation uncertainties. This formulation separates what the model assumes about evolution from what instruments actually observe. It can describe systems ranging from a battery’s internal condition to the distribution of water through a pipe network.\n\nData-driven models instead learn relationships from observations. Machine learning can estimate normal operating behavior, identify unusual patterns, or approximate expensive calculations. Such models may be useful where mechanisms are incompletely understood or too costly to model directly. Their limitations depend strongly on the training data: a model trained on steady operation may be unreliable during startup, equipment damage, or an unfamiliar environmental condition. Accurate interpolation within past experience is not equivalent to reliable prediction outside it.\n\nHybrid approaches combine mechanistic and empirical elements. A physical model may describe heat flow while a learned correction represents uncertain heat losses. Alternatively, known conservation laws can constrain a statistical estimator. These arrangements can improve plausibility and data efficiency, but they do not automatically eliminate error. A correction model may compensate for incorrect parameters during calibration and then fail when operating conditions change.\n\nComputational cost determines which methods can run at which timescale. A detailed three-dimensional calculation might require hours, whereas an operational decision may be needed within seconds. Model-order reduction, simplified equivalent circuits, lookup tables, and surrogate models can reduce this burden. Different models may coexist: a fast approximation for monitoring and a more detailed model for periodic investigation. Neither is universally superior; suitability depends on required accuracy, update frequency, and available resources.\n\nStochastic simulation adds another dimension by representing multiple possible futures rather than one trajectory. A fleet-maintenance twin may sample uncertain usage and failure processes, while a production twin may simulate variable arrival times and breakdowns. Such outputs are conditional on assumptions and should not be interpreted as exact forecasts."
    },
    {
      "heading": "Synchronization, state estimation, and uncertainty",
      "text": "Maintaining a twin involves more than copying measurements into a database. Instruments observe only selected quantities, while many states of interest—internal temperature, accumulated damage, or remaining material thickness—are hidden. Synchronization therefore often requires state estimation: combining a model’s prediction with observations to infer the system’s condition. Sensor fusion can integrate measurements with different noise characteristics, sampling rates, and spatial coverage.\n\nA simple estimator corrects a predicted state using the discrepancy between predicted and measured outputs. In schematic form,\n\n\\[\\hat{x}_k^{+} = \\hat{x}_k^{-} + K_k\\left[y_k-h(\\hat{x}_k^{-})\\right],\\]\n\nwhere \\(\\hat{x}_k^{-}\\) is the prior estimate, \\(\\hat{x}_k^{+}\\) the corrected estimate, and \\(K_k\\) a correction gain. The Kalman filter provides one established framework for calculating such corrections under specified assumptions; nonlinear and ensemble variants address more complicated systems. In large environmental and geophysical models, related procedures are commonly described as data assimilation.\n\nWhether a state can be inferred at all is a question of observability. More sensors do not necessarily resolve ambiguity if they measure essentially the same information. For example, reduced cooling performance could result from fouling, altered flow, or an incorrectly calibrated temperature sensor. Additional measurements or controlled changes in operating conditions may be needed to distinguish these causes. Parameter estimation presents a related difficulty: several different parameter combinations can sometimes explain the same observations.\n\nSynchronization also has temporal limits. A twin can be accurate about the recent past yet unsuitable for an immediate decision if its measurements arrive too late. Update intervals should reflect system dynamics and intended use. Vibration monitoring may require high-frequency acquisition, whereas a bridge inspection model may incorporate detailed surveys only periodically. “Real time” is therefore a requirement relative to a process, not a universal sampling rate.\n\nMeasurement uncertainty, parameter uncertainty, uncertain future inputs, and model inadequacy all affect predictions. These sources should not be collapsed into a single unexplained accuracy claim. A confidence interval for an estimated parameter, for instance, does not necessarily capture the error caused by omitting a physical mechanism. Uncertainty quantification examines how such uncertainties influence outputs and decisions.\n\nPrediction errors can also change over time. Sensor drift, repairs, wear, software changes, and new operating regimes may invalidate earlier calibration. Residuals—the differences between observations and model predictions—can reveal such changes, but a small residual is not sufficient proof of correctness if the model is being continually adjusted to fit the same data. Independent measurements, withheld operating periods, and explicit limits on model use help distinguish genuine predictive ability from retrospective agreement."
    },
    {
      "heading": "Engineering life cycle, maintenance, and control",
      "text": "A twin can support different activities as an asset moves from design through manufacture, operation, and retirement. During design, models can compare configurations and evaluate expected loads. During production, measurements of geometry, material properties, and assembly conditions can make the representation specific to an individual item. Operational records then add evidence about how that item has actually been used. Continuity across these stages requires careful management of revisions: the designed, manufactured, installed, and currently operating configurations are not always identical.\n\nVirtual commissioning is one practical application. A simulated production line can be connected to control software before the complete physical line is available. Engineers can test sequencing, interlocks, material flow, and responses to selected faults without moving real machinery. This can reveal interface errors, but the result is bounded by the simulation’s coverage. A collision or fault omitted from the virtual environment cannot be ruled out merely because simulated tests succeeded.\n\nIn operation, twins can support condition monitoring and predictive maintenance. A rotating-machine model might combine speed, load, vibration, and temperature with knowledge of bearings and lubrication. Rather than treating a high vibration level as universally abnormal, it can compare the observation with what is expected under the current operating condition. Reliability engineering then connects condition estimates to failure mechanisms, inspection intervals, and maintenance consequences.\n\nRemaining useful life is a particularly demanding output. It depends not only on present degradation but also on future loading, environmental exposure, maintenance, and the chosen failure threshold. A numerical estimate should therefore be understood as conditional. A component expected to last a certain period under one duty cycle may behave differently after reassignment to heavier service. Some applications are better served by a probability of exceeding a damage threshold within a defined interval than by a single countdown to failure.\n\nTwins may also support model predictive control, in which a model evaluates future trajectories and an optimizer selects actions subject to constraints. However, a twin and a controller are not the same thing. A twin can serve only as an advisory system, while an existing controller remains responsible for execution. If recommendations are connected to automatic actuation, timing, stability, failure response, and functional safety become system-level requirements.\n\nAt retirement or refurbishment, accumulated records can inform component reuse, remanufacturing, and disposal. That value depends on data continuity. A sophisticated operational model may be of little use to a later owner if records cannot be transferred, interpreted, or trusted. Life-cycle support therefore involves contractual and organizational arrangements as well as modeling technology."
    },
    {
      "heading": "Manufacturing, energy, and transport applications",
      "text": "Manufacturing applications span individual machines, work cells, production lines, and entire factories. A machine-tool twin may estimate thermal deformation and its effect on dimensional accuracy. A production-system twin may instead represent queues, processing times, equipment availability, and routing choices. These are different modeling problems: a highly detailed mechanical model does not automatically predict factory throughput, which depends on interactions among machines, workers, buffers, and schedules.\n\nDiscrete-event simulation is useful when operational behavior is governed by events such as an arriving part, a completed operation, or a machine failure. Connected to current production records, it can compare alternative schedules or estimate the effect of losing a critical workstation. The principal difficulty is often representing actual operating rules, rework, material shortages, and human interventions rather than reproducing nominal cycle times. Integration with logistics systems extends the analysis to stock availability and inbound or outbound transport.\n\nEnergy assets provide examples of multiscale modeling. A wind turbine twin can relate wind conditions, rotor behavior, structural loading, and power output. A gas turbine twin may track performance changes associated with compressor fouling or component deterioration. In both cases, changes in ambient conditions must be separated from changes in equipment condition; otherwise, normal environmental variation can be mistaken for degradation.\n\nAn electrical grid twin may combine network topology, equipment parameters, power measurements, and switching states. State estimation and power-system simulation long predate the digital-twin label, and adopting that label does not by itself add a new analytical capability. Its practical significance lies in connecting previously separate information sources, maintaining configuration consistency, and supporting repeated operational or planning analyses. A model with an outdated switch position can be less useful than a simpler model with correct topology.\n\nFor a lithium-ion battery, internal state estimates can connect cell voltage, current, and temperature to state of charge and state of health. A twin may extend beyond the real-time functions of a battery management system by combining manufacturing information, service history, and fleet-level analysis. Aging models must account for the distinction between time-dependent deterioration and deterioration associated with cycling, as well as their dependence on operating conditions.\n\nTransport applications range from vehicles to networks. A railway twin may connect infrastructure geometry, inspection results, train movements, and maintenance records. A vehicle twin may represent fatigue-sensitive components or thermal behavior. At a seaport, the modeled system may instead center on berth allocation, cargo handling, storage, and connecting transport. The common principle is an updated representation tied to decisions; the relevant physics, timescales, and operational constraints differ substantially across these examples."
    },
    {
      "heading": "Buildings, infrastructure, and environmental systems",
      "text": "Built assets pose a different life-cycle problem from many manufactured products: they often remain in use for decades while undergoing alterations, changes of ownership, and incomplete documentation. A building twin may begin with a design model but must eventually reflect the asset as constructed and subsequently modified. Connecting spaces, equipment, and documents to persistent identifiers is essential if operating data are to remain interpretable through renovation and equipment replacement.\n\nBuilding-energy applications link geometry and construction properties with occupancy, weather, and equipment operation. Measurements from heating, ventilation, and air-conditioning systems can be compared with predicted loads or equipment performance. Discrepancies may indicate incorrect schedules, sensor faults, abnormal control behavior, or errors in the building model. Occupant behavior introduces substantial uncertainty, so a detailed geometric representation alone does not guarantee accurate energy prediction.\n\nCivil-infrastructure twins can combine design calculations, inspection records, and measurements of structural response. For a bridge, strain, acceleration, temperature, and traffic information may help interpret changes in behavior. Seasonal thermal effects must be distinguished from structural deterioration. Nondestructive testing and periodic close inspection can provide evidence unavailable from permanently installed sensors. A twin is therefore often a framework for combining heterogeneous evidence rather than a replacement for physical examination.\n\nWater systems connect hydraulic models to information about pipes, pumps, valves, storage, and demand. Observed flows and pressures can help assess leakage or investigate whether a proposed operating change would maintain service. Urban drainage applications may use models such as the Storm Water Management Model to examine runoff and sewer behavior. Rainfall forecasts introduce additional uncertainty, and the spatial resolution of precipitation data can matter as much as the detail of the drainage model.\n\nAt city scale, the term “digital twin” can describe a federation of models rather than one comprehensive simulator. Transport, energy, flood risk, land use, and building systems have different owners and update cycles. Urban planning questions also involve policy choices and human responses that cannot be inferred from geometry alone. A model of travel times, for example, does not by itself establish how households will respond to a new housing policy.\n\nEnvironmental twins extend similar methods to watersheds, coasts, and other natural systems. Remote sensing can provide spatial observations that ground instruments cannot economically supply, while models connect those observations to physical processes. Their boundaries are necessarily open: weather, upstream conditions, and human activity can change independently of the modeled system. This makes explicit scenarios and uncertainty ranges especially important when such representations support assessments of infrastructure resilience."
    },
    {
      "heading": "Validation, security, and limits of reliability",
      "text": "A twin’s credibility must be assessed against a specified purpose. Verification asks whether the computational implementation correctly represents its mathematical and software specifications. Validation asks whether the model adequately represents relevant real-world behavior for the intended application. Calibration adjusts parameters using evidence; it is not equivalent to independent validation. A model calibrated against one operating period may reproduce that period closely yet perform poorly under different loads or after equipment changes.\n\nAssessment can draw on controlled tests, historical operating records, independent instruments, and comparisons with alternative models. Useful performance measures depend on the decision. Average temperature error may be appropriate for one energy assessment, whereas missed detection of a rare damaging condition matters more for a maintenance system. Error should also be evaluated across operating regimes rather than only as a single aggregate score. Good performance during frequent, steady conditions can conceal serious deficiencies during uncommon transients.\n\nA digital twin creates a substantial cybersecurity surface because it links operational data, engineering information, and sometimes command pathways. False measurements can corrupt state estimates; altered configuration records can make a correct calculation apply to the wrong system. An attacker may also learn sensitive operational details from a read-only twin. Authentication, access control, integrity checking, network separation, and auditable changes address different parts of this problem. The ability to issue commands warrants particular scrutiny because a compromise may then affect physical operation.\n\nHuman-centered twins and detailed building or workplace models can also raise data privacy issues. Occupancy traces, movement histories, and equipment-use records may reveal information about identifiable individuals even when they were collected for operational purposes. Data ownership and intellectual property complicate exchanges among equipment suppliers, operators, software providers, and asset owners. Access to a machine’s measurements does not necessarily confer rights to its manufacturer’s proprietary model.\n\nEconomic and organizational limits can be decisive. Instrument installation, communications, model development, software integration, validation, and long-term maintenance all incur costs. Adding detail may increase expense without changing a decision. A simpler, well-maintained representation can therefore be more valuable than a comprehensive but poorly supported one. Evaluation should compare the twin with an explicit alternative, such as conventional inspection, an existing monitoring system, or a smaller modeling effort.\n\nOpen technical problems include transferring models between similar assets, detecting when a model has left its validated domain, and maintaining trustworthy interactions among independently developed twins. Coupled models can exchange inconsistent quantities or advance on incompatible timescales, producing numerical or conceptual errors. As systems evolve, each update can affect earlier validation evidence. Long-lived deployments consequently require documented assumptions, version histories, identifiable data dependencies, and procedures for revising or withdrawing outputs when those assumptions no longer hold."
    }
  ],
  "links": [
    "sensor",
    "building-information-modeling",
    "apollo-program",
    "programmable-logic-controller",
    "scada",
    "application-programming-interface",
    "mass-balance",
    "energy-balance",
    "finite-element-method",
    "machine-learning",
    "training-data",
    "sensor-fusion",
    "kalman-filter",
    "observability",
    "measurement-uncertainty",
    "commissioning",
    "model-predictive-control",
    "logistics",
    "wind-power",
    "gas-turbine",
    "electrical-grid",
    "lithium-ion-battery",
    "battery-management-system",
    "railway",
    "seaport",
    "air-conditioning",
    "storm-water-management-model",
    "urban-planning",
    "remote-sensing",
    "infrastructure-resilience",
    "calibration",
    "cybersecurity",
    "data-privacy",
    "intellectual-property"
  ],
  "sources": [],
  "license": "https://creativecommons.org/licenses/by/4.0/"
}