TL;DR: A digital twin is a digital representation of a real thing or process that is kept in step with it using real data. NIST notes there is still no single agreed definition, and the term covers everything from a live asset dashboard to a physics simulation that steers a machine. It helps to think in levels of fidelity: a structured record of the asset, live monitoring, diagnostics, prediction through simulation, and finally closed-loop control. Each level needs better data plumbing from IoT and costs more to build and maintain. Twins pay off where assets are valuable, behaviour is measurable and decisions are frequent; they rarely pay off as a 3D showcase with no decision attached.
What a digital twin is, and is not
The most quoted definition comes from the Digital Twin Consortium (2020): a digital twin is "a virtual representation of real-world entities and processes, synchronized at a specified frequency and fidelity." NIST's report Security and Trust Considerations for Digital Twin Technology (NIST IR 8356, February 2025) cites it, and adds an honest caveat: despite many definitions from researchers, standards bodies and industry, there is no agreed-upon definition. ISO/IEC 30173:2023 now sets out digital twin concepts and terminology, which helps teams use the words consistently.
Three ideas are common to the serious definitions:
- A specific counterpart. The twin represents a particular pump, building, production line or process, not just a generic type. NIST distinguishes a digital twin definition (a machine-readable description of a type of entity) from the twin instances created from it.
- Synchronisation. The twin is updated from the real thing at some frequency, from near real time to daily. A model nobody updates is a model, not a twin.
- A purpose. Fidelity and frequency are chosen for a use case. NIST observes that what twins add over traditional modelling and simulation is real-time monitoring of entities while dynamically updating their digital counterparts.
What a twin is not: a 3D model on its own, a CAD file, or a dashboard of sensor values with no model of the asset behind it. Those can be parts of a twin.
Levels of fidelity
There is no single official maturity scale, so the levels below are a practical framing rather than a standard. They are useful because each level answers a different question and needs different data.
| Level | Question it answers | What it needs |
| 1. Descriptive | What is this asset, what is it made of, what has been done to it? | Asset registry, documents, maintenance history, structured data model |
| 2. Monitoring | What is happening right now? | Live IoT data mapped to the asset model |
| 3. Diagnostic | Why did it happen? | History, events, context (operating mode, batch, operator), analytics |
| 4. Predictive | What will happen if...? | Simulation or machine learning models calibrated against real data |
| 5. Prescriptive or closed-loop | What should we do, and can the system do it? | Validated models, optimisation, safe control interfaces, strong security |
Most of the value for most organisations sits in levels 1 to 3. They are also prerequisites for 4 and 5: a predictive model fed by an inconsistent asset registry will predict the wrong thing confidently.
Data plumbing: from IoT to twin
A twin is only as good as the data feeding it. The plumbing usually looks like this:
- Sensors and control systems produce measurements and events. The device, gateway and messaging layers are described in our guide to industrial IoT architecture.
- Ingestion validates and timestamps data and attaches it to the right asset identity.
- An asset model describes what exists and how it relates: site, line, machine, component, sensor. This is the heart of the twin.
- Storage combines time-series data with the asset model and business records (work orders, batches, quality results).
- Analytics and simulation run against that combined view.
- Applications present it to people, or feed decisions back to operations.
Standards for describing assets
Modelling every asset from scratch is slow and leads to models that cannot be shared. Existing standards help:
- Asset Administration Shell (AAS). Specified by the Industrial Digital Twin Association and standardised as IEC 63278-1:2023, the AAS defines a structure of submodels for exchanging information about industrial assets between applications. It is increasingly used in manufacturing supply chains.
- ISO 23247. A digital twin framework for manufacturing; Part 1 (2021) covers overview and general principles, with applications such as real-time control, off-line analytics, predictive maintenance and production control.
- DTDL. Microsoft's Digital Twins Definition Language is based on JSON-LD and is programming-language independent; models describe properties, relationships and components of entity types.
- FMI. The Functional Mock-up Interface, a Modelica Association project, is a free standard that defines a container and an interface to exchange dynamic simulation models between tools.
Open-source platforms such as Eclipse Ditto, a framework for building digital twins of internet-connected devices, provide the state-synchronisation layer if you prefer not to build it yourself.
Data quality is the real work
- Consistent asset identifiers across IoT, maintenance, ERP and engineering systems.
- Units, timestamps (UTC) and sampling rates documented and enforced at ingestion.
- Context captured alongside measurements: operating mode, product, setpoints.
- A way to handle gaps, sensor drift and replaced components without corrupting history.
Simulation: when the twin starts to predict
Level 4 adds the ability to ask "what if". Two broad approaches exist, and they combine well:
- Physics-based or engineering models (thermal, mechanical, flow, discrete-event process models) that encode how the system works. They extrapolate better to situations not seen before, but need expertise to build and calibrate.
- Data-driven models (statistics and machine learning) that learn patterns from history. They are quicker to build where data is plentiful, but are only reliable within the conditions they were trained on.
Calibrate any model against measured reality and keep checking it: a twin whose predictions drift from what the sensors report should raise an alert about itself. Discrete-event simulation of a warehouse or line is a particularly practical case, for example sizing a robot fleet before buying it (see our guide to warehouse robotics and automation software).
Where digital twins pay off
- High-value assets with costly downtime, where condition monitoring and earlier fault diagnosis justify the investment.
- Processes with many interacting parts, such as production lines, logistics hubs and building systems, where people struggle to see the whole picture and simulation can test changes safely.
- Design and commissioning, where testing control logic or layouts virtually before building can save rework.
- Portfolios of similar assets (fleets of machines, chains of buildings), where one model definition can be reused across many instances and compared.
Where they usually do not
- No decision attached. If nobody will act differently because of the twin, it is a visualisation project.
- Cheap, replaceable assets where running to failure costs less than monitoring.
- Poor or missing data, where most of the budget goes into fixing data foundations. That work may still be worth doing, but call it that.
- Fidelity beyond the need. A photorealistic 3D scene adds cost without improving a maintenance decision.
- No owner. Twins need ongoing maintenance as assets, sensors and processes change. Without an owner and budget they decay.
Security and trust
A twin aggregates detailed operational data and, at higher levels, can influence real equipment. NIST IR 8356 discusses both familiar and new cybersecurity concerns for twins, and trust questions about whether a twin's outputs can be relied on. Practical steps: treat the twin platform as part of your operational technology security boundary, keep any write path to physical systems separate, authenticated and audited, and record the provenance and version of models whose outputs inform decisions.
A sensible way to start
- Pick one asset class and one decision, such as when to service a fleet of compressors.
- Build the asset model and connect the data you already have before adding sensors.
- Deliver monitoring and diagnostics that the people making the decision actually use.
- Add prediction only where the history supports it, and measure whether decisions improve.
- Reuse the model definitions for the next asset class.
How ApexlyTech can help
We build the data and software layers of digital twins: IoT ingestion, asset models, time-series storage, analytics and the web and mobile applications that put them in front of operators and managers. See our IoT solutions, cloud solutions and custom software services, or contact us to scope a first use case.
Frequently asked questions
What is a digital twin in simple terms?
A digital representation of a specific real-world asset, system or process, kept up to date with data from it, and used to monitor, understand, predict or improve how it behaves.
What is the difference between a digital twin and a simulation?
A simulation models how a system could behave. A digital twin is tied to a specific real counterpart and is kept synchronised with it using live or regular data. Simulation is often one capability inside a twin.
Do I need 3D visualisation for a digital twin?
No. Many valuable twins are data models with dashboards and analytics. 3D helps where spatial understanding matters, such as layouts or building systems, but it is not what makes something a twin.
Which standards are relevant to digital twins?
ISO/IEC 30173 for concepts and terminology, ISO 23247 for manufacturing, the Asset Administration Shell (IEC 63278-1) for describing industrial assets, DTDL for modelling in some platforms, and FMI for exchanging simulation models.
How much data do I need to start?
Enough to support the first decision you want to improve. Start with the data you already have, a clean asset model and consistent identifiers, then add sensors where gaps block that decision.
Sources