Guidance

Digital twin (easy-read version)

Published 29 October 2025

A digital twin is a computer-based version of something real, for instance, an object, place, or process. It can send and receive information to and from the real thing, helping people make decisions based on current and relevant data.

What we mean by data and information

Data means any kind of raw facts or figures, and should be understood broadly here. Information is what you get when data is put into context or given meaning, making it useful for understanding a situation or making decisions.

This means the digital twin closely copies its real-world counterpart in all relevant aspects, in that timeframe and without statistical bias - bias is a statistical term here. Without bias means the method doesn’t consistently over-estimate or under-estimate the true answer.

Digital twins are complex, and so in this context, the definition can be relaxed to a distribution containing the true value within a stated percentile. In other words, there is no substantial over or underestimate even if the uncertainty is high.

A digital twin can assess the following with a certain level of assurance:

• functionality
• degradation
• impact

Defining a digital twin

A digital twin must be tied to a real-world process, environment or object. It must also apply a known tolerance and perform without statistical bias, as well as predict what the associated real-world counterpart would do given a trigger.

When combined with the basic requirements (listed on this page) and a shared way of exchanging data, this:

• makes sure digital twins can easily work together with other digital twins
• provides a known level of assurance in any situation within a specified validation and assumption envelopes (including after integration with other digital twins)
• ensures digital twins can be used for test and evaluation and safety case development

The different states of a digital twin

A digital twin can be in any 1 of the following 3 states at a time:

• connected - if it is currently being fed with data from the real-world counterpart
• semi-connected - if it is fed with simulated data but contains at least 1 real world data feed (semi-connected twins can be used to assess the impact of potential future changes on the digital twin and its real-world counterpart)
• disconnected - if it mimics the physical counterpart up to the point at which is disconnected, but is currently being run using only simulated data and is ready to be ‘twinned’ or reconnected with its physical counterpart

A digital twin may move between these states and remain a digital twin, provided its current state is recorded (see figure 1).

A disconnected digital twin can only return to being a connected digital twin if the reconnection occurs within a timeframe in which it will still accurately mimic its real world counterpart.

A digital twin must also be independent of the physical entity, except for the input of data streams into the digital twin - the physical entity must maintain safe functionality even without a continuous connection to the twin.

The communication flow to the twin will usually be of sensed data, while the twin sends back decisions that can either change how the real-world system behaves, or give insights into how it might perform in the future.

A digital twin can therefore inform a control system for the real world counterpart, but not be an integral part (the control system may not function optimally or fully without a connection to the digital twin, but must function safely) of that control system.

Consider a digital twin of a building

A digital control system will monitor the temperature inside the building and adjust windows and air conditioning to maintain a fixed temperature. Regardless of the building’s digital twin, the control system must still function.

The control system is ‘reproduced’ within the digital twin along with characteristics of the building outside of the control system.

In these circumstances, the digital twin could provide data to inform the building control system of the state it should be running in.

Figure 1: illustration of the required and potential data flows between the real world and the states of a digital twin. Here the semi-connected and disconnected digital twins would be branched (copied) from the connected digital twin and run to answer 'what-if' analysis. Following these analyses digital twins in this state could be retired or merged with the ongoing connected twin. Two digital twins can be composed together to allow data flow between them, without necessarily changing the state of the individual twins.

Requirements for a digital twin

A digital twin must do all of the following:

• mimic its real-world counterpart to a known and unbiased level of tolerance
• have set boundaries where it’s been checked and proven to work properly (validation envelope)
• have an associated assumption set
• run in a timeframe appropriate for the required decisions and assumptions
• be based on physical parameters and specifications
• be associated with a single known physical entity
• allow a 2-way data flow in to and out of the real world

We will now define these requirements further.

Mimicing its real world counterpart to a known and unbiased level of tolerance

A digital twin must accurately mimic all the important features of its real-world counterpart it’s meant to represent for its specific purpose.

If you have several digital twins of the same object for different uses, they should work together consistently. Where possible, these separate twins could be combined into one more comprehensive version, or linked together as a connected set.

A statistical, data or simulation-based model can only be a digital twin if it has been trained specifically on a physics-based (or theory-based) representation of the real thing with a known and unbiased level of uncertainty. Or, if it has been trained using data from the specific real-world counterpart. If a model is physics-based, it can only be a digital twin if it is associated with a known tolerance and fidelity (how closely it matches reality).

Having a specified validation envelope

A digital twin will only mimic its physical counterpart within known validation windows (the defined limits where the model has been tested and proven to work accurately).

For instance, if the twin is a physics-based model trained on data within a known temperature range, there can be confidence within that temperature range. Outside of that range, there will be a point at which the physics and data changes and the digital twin will no longer be valid.

To meet the criteria of a digital twin, you must clearly define what conditions it can accurately predict and consider these limits when asking ‘what-if’ questions.

If the real-world thing must unexpectedly operate outside of the validation range, you must carry out an immediate re-assessment of the digital twin.

Having an associated assumption set

Even when used correctly, the digital twin will only copy certain characteristics of the real thing. It may reflect size, shape, weight, motion and temperature for instance, but not model the effect of an external force - it might not be able to predict what happens if you push or pull the object.

Consider a digital twin of a car brake system

A digital twin can still qualify as such even if it can’t predict what happens in a car crash, as long as this limitation is clearly documented and users know about it.

Running in a timeframe that is appropriate for the required decisions and assumptions

A digital twin must process information at the same speed as what’s happening in the real-world.

To achieve this, information must flow both ways between the digital twin and the real thing fast enough to keep them in sync. The timing needs to match what you’re using it for.

For example, if a manufacturing plant runs on a 24-hour schedule with 8 non production hours, the twin must be capable of updating to current status within that 24-hour window.

Equally, an aircraft engine digital twin must be capable of running to the current status of the real-world thing after landing and before the status is next updated. If the digital twin is to be run in order to evaluate ‘what-if’ scenarios in a semi-connected or disconnected state, the twin may need to run faster than real time. This capability is not, however, a requirement to meet the definition of a digital twin.

Be based on physical parameters and specifications

A digital twin must work like the real-world thing it copies, with all the same limitations and specifications. If a digital twin is developed for a new design of a real-world thing (physical entity), you can’t give it components or properties that don’t exist in real life.

For example, a digital twin would not contain models of metals with properties that can’t be made. To analyse the impact of idealised properties, a digital twin would need to be used with a separate model or simulation - you’d need to use the digital twin alongside a separate computer model.

Be associated with a single known physical entity and allow a 2-way data flow into and out of the real world

To create digital twins of real things, you need a template for each type of physical entity you want to copy. This template must be able to send information to and receive information from the real world (2-way data flow).

If a connected digital twin gets cut off from its real car and starts using fake data instead, it becomes ‘disconnected’ again.

Consider the digital twin of a specific car type

Every manufactured car will have its own digital twin copy linked to it. When the digital twin is linked to its real car, it will be ‘connected’.

Before the car is built, the digital twin that meets the criteria set out on this page, and allows for but does not yet have a 2-way data flow, is ‘disconnected’.

A ‘connected’ digital twin that becomes severed from it’s real-world counterpart and fed from a simulation environment to assess performance, is also ‘disconnected’.

Sometimes a digital twin uses a mix of real-world and simulated inputs which is ‘semi-connected’.

For example, a digital twin that depends on both ambient temperature and external vibration - if the current ambient temperature is drawn from the real world, but the vibration level was simulated using a model to understand the impact of stress on the real world object, this is ‘semi-connected’. This is because it’s only partially linked to the real world.

Where digital twins can be applied

Digital twins of assets and capabilities

This area covers digital copies of parts, component, platform, system and manufacturing (factory) digital twins.

Different industry partners have digital copies of individual parts that need to be combined at the platform level, and even though this is challenging, there has been some success with this in the past.

The issue is the need for digital copies that work at different levels of detail, and getting reliable results quickly.

Digital twins of environments

This area covers infrastructure, transport, manufacturing, environment and mega-structure digital twins.

A mega structure is a complicated mix of of both component, platform, infrastructure and environment twins (found in an urban environment, like a city, for instance).

This needs complex digital copies that work at different levels of detail, with live data feeds to accurately show what’s happening in real time. The system also needs to send information back to the actual infrastructure.

This is currently unproven technology with substantial open research and technological challenges.

Digital twins of humans

A true complete digital twin of a human is currently impossible.

There is a focus on creating digital twins of certain parts of a human though, such as a human heart, which would be advantageous in lots of ways to medicine, as well as equipment design. However, even at this level, we can’t yet connect this type of digital copy to computer models that could work out the forces a body puts on its surroundings or clothes.

Adding a digital layer showing a person’s health on top of their physical digital copy is currently very difficult, but it might be possible in the future. Adding another layer showing the person’s psychological state however, would make the whole human digital copy impossible to create with current technology and would raise serious ethical concerns.

Threats and security of digital twins

You should consider the different complex ethical and security implications associated with digital twins, even though the security considerations should be proportionate to the specific use case.

We suggest looking at the following different security and ethical considerations:

• we don’t yet know what type of unwanted affect an adversary gaining access to a digital twin might have - to ensure capability and safety are maintained, there will need to be some assurance that the digital twins are both representative and error-free when in use

• to maximise the benefits of digital twins, the large-scale passing of data for maintenance and analysis is crucial (we don’t currently know what the best methods are to secure this data unfortunately, and are not captured in policy and guidance)

• we don’t really understand how digital copies of humans would affect healthcare, finance and people’s control over their own data, so it’s hard to know what ethical problems this might cause