Defining Digital Twin: History, Attributes and Examples

"Doublethink means the power of holding two contradictory beliefs in one's mind simultaneously, and accepting both of them." George Orwell, 1984

 

Among the many ambiguous terms used in industry, the term "Digital Twin" seems to have a highly malleable meaning these days. The term is being applied to a wide variety of applications which significantly differ. Orwell’s quotation seems applicable when observing the modern usage of the term. Why does this matter? Industry seems to be able to survive a tower of Babel. However, many documented cases exist in which this confusion has resulted in mismatched expectations and failure to produce value from the investment. Among those are:

• Oil & gas and utilities firms found only 11% of their digital twin projects produced the expected benefits
• In 2024, a major Southeast Asian port authority invested $12 million in a digital twin platform and abandoned it 18 months later
• An automated highway maintenance project found that the lack of a common understanding of the Digital Twin concept was a roadblock

Digital twin attributes

The original definition came from Dr. Michael Grieves in the “product lifecycle management” concept beginning with presentations in 2002 at a Society of Manufacturing Engineers meeting and in University of Michigan executive courses, and later in the 2005 publication “Product lifecycle management: the new paradigm for enterprises” (International Journal of Product Development, Vol. 2 No. 1/2, pp. 71-84). NASA engineer John Vickers named this concept “digital twin” in a 2010 NASA technology roadmap. The attributes of “digital twin” required the following attributes:

1. It is a model that behaves like the physical entity it represents
2. Data from the physical entity is used to maintain the accuracy of the model
3. Information from the model is used to impact the physical entity

Any application that includes all those attributes can be considered a digital twin.

It may be obvious that such a definition embraces a very wide collection of applications. That may be unsatisfactory, but it is correct. Digital Twin is an umbrella term for a wide variety of specific twins with more narrow application. IBM represents Digital Twin as shown below, with Asset Twin, Simulation Twin, Design Twin, Operational Twin and others as the more specific instances under the umbrella.

If we accept this approach, then our definition states:

• There is no formal authority on the definition of Digital Twin
• There are a number of common descriptions that can widely differ
• It is a broad umbrella term
• It is hard to say what it isn’t because there aren’t many attributes that would exclude it

Part of the confusion arises when a specific instance is considered a digital twin, but others aren’t. This may arise from a parochial attitude. If the digital twin you use differs from another, you may conclude that only your instance is a digital twin.

Analyzing attributes

Consider Attribute 1. It does not state what kind of model it must be. It could be steady state or dynamic. It could be linear or nonlinear. It could be simple or complex. It could be first principle or heuristic. There is no restriction on how large the physical entity is. It could be an enterprise, a facility, a unit operation, or even a single loop. It also doesn’t explain how realistic the behavior must be. Those details can be considered specific instances, but the attribute is rather broad.

Regarding Attribute 2, notice there is no requirement that the data is provided through a realtime connection to the physical entity. If historical data is obtained offline and used for model training through supervised learning, it can still be a digital twin. With modern control systems and adaptive methods, it is possible to use unsupervised learning for this purpose, but with noise and outliers, this is risky. Most industrial uses require human pre-processing of data for a digital twin. This attribute does require that the physical entity is actual, not conceptual. A model of a theoretical reactor that behaves like most reactors is not a digital twin unless you can get data from an actual reactor to train the model.

Lastly, Attribute 3 does not require closed loop connectivity to the physical entity. Frankly, this is not desirable with the current state of technology, and may never be. Notice it does not say that data is used, but it is information obtained from the model that helps the operation. This information can be presented to operators to help them understand the behavior and better operate.

Example of dissent

For a contrast, below is a definition of digital twin offered by a well recognized source in industry that we will keep anonymous:

1. A process model must adapt from realtime data, and the learning is unsupervised (no human review, approval or intervention).
2. There is automatic flow of data from the digital twin back to the process that can influence or control the physical process.
3. It must encompass the entire process.

There are no applications in the manufacturing industry that meet those requirements:

1. Realtime data can be noisy, inaccurate or invalid. There are times when the process is down, and the data is useless. If a human does not pre-process the data to remove noise and outliers, the model learns garbage. We know of no applications in industry that entirely replace subject matter or data analytics engineers with unsupervised software pre-processing.
2. It is very risky to automatically connect an adapting model to a physical process. Every deployment we know of requires a human in the loop to evaluate and approve predictive process models before they can impact the process. Connecting this application to valves and motors with no human in the loop is dangerous.
3. Most applications of realtime prediction models are applied to unit processes, not an entire process. For example, in a paper mill there can be such applications for each paper machine, a coater kitchen, a lime kiln, each evaporator set, each continuous digester, and each boiler. In a chemical facility the process model may apply to one reactor. If a PID loop is not big enough to be considered a digital twin, how many inputs and outputs are required to make it big enough? There is no clear definition of the size of a digital twin.

Therefore, digital twins as defined this way do not exist. While the attributes do describe a digital twin, they improperly exclude many digital twins that exist today by including unnecessary requirements.

ISO 23247

Many regard the ISO 23247 series of standards to be the definitive source for digital twins. The definition used in 23247 is:

3.2.2 <manufacturing> fit for purpose digital representation of an observable manufacturing element with a means to enable convergence between the element and its digital representation at an appropriate rate of synchonization

An examination of this definition shows alignment with the attributes offered in the original definition.

• Fit for purpose: A model of a physical entity is clearly fit for the purpose intended
• Digital representation: This seems too broad a description. This could include a digital drawing that demonstrates no behavior. Model is a better word to describe the purpose
• Observable manufacturing element: We choose physical entity to ensure the subject we are discussing actually exists
• Means to enable convergence: This can be accomplished through realtime connectivity or manual pre-processing of data and supervised model training. Convergence means we can maintain the accuracy on matching the model as the physical entity changes. This does not limit the means to automation. A manual process with a human in the pre-processing and model training effort can be the means of convergence.
• Appropriate rate of synchonization: For some applications, a quarterly or semi-annual retraining of a model may be appropriate rate of synchronization. In industry we understand that models should not be retrained for transient conditions. In some cases, such as non-linear gains, using adaptive identification with sufficient filtering may be appropriate at rates of minutes or seconds per cycle.

We believe our list of attributes provides more clarity.

What is it?

Some examples of digital twins in industry today are:

• Multivariable Predictive Control (MPC): This technology has long been utilized for controlling unit operations such as distillation columns, reactors, and paper machines. While it does not fully model the unit operations, it does model the Control Variables pertinent to product specifications as a function of Manipulated and Disturbances variables. The dynamic model demonstrates the behavior of the system. Routine model retraining has been long recognized as a requirement to ensure performance is maintained. Current implementations include pseudo random binary sequence (PRBS) to automatically generate new model recommendations. The model outputs drive the process to setpoint ranges and optimal conditions at a rate of minutes to seconds.
• Proportional Integral Derivative (PID) loop: Many would object to a PID loop being considered a digital twin. With only one input and one output, it seems too simple. Many also consider it to be model free. For many practitioners, we understand that a PID algorithm is a process model of a second order linear gain process with no deadtime. Tuning is derived from single loop process model identification and tuning rules based on process data. There are applications that continuously monitor model identification, in a similar way to the PRBS technique described above for MPC.  Lacking a constraint of how big the physical entity must be, the combination of sensor and actuator constitutes a physical entity that can be simulated and used for improved operation.
• Offline process simulation: It is common for manufacturers to have a training system that has no realtime connection to a physical process, but is used to train operators on how the process will respond. A process model exhibits behavior that represents the process. If an operator accidentally stops a motor, the simulation will show the undesired response without doing any actual damage. If this model is maintained to ensure it matches the actual process, it is synchronized. The information the operator obtains is in their head when they have their hands on the real thing.
• Product design and development: Digital Twins can be used to simulate the product behavior before it is built and shipped to the field for use in production. The focus in this stage is to optimize the design of the product and accelerate the time to market of the physical assets. Use of Digital Twins, allows the product designer to accelerate the design cycles and reduce the physical prototyping costs, before manufacturing the product at scale. A first principle model can be used when designing a mechanical entity to understand its behavior as it is being designed. This provides faster development and help ensure safe design. The synchronization occurs when the entity is produced and its behavior matched to the model. Adjustments to the design model can be new first principle refinement or tuned parameters in the first principle model.
• Manufacturing process: At this stage Digital Twins have two-pronged use. Digital Twin of the manufacturing process can help to optimize the throughput and the quality of the physical asset being manufactured. Digital Twins help to trace back the defects encountered with specific batches of production, once the assets are in use in the field. Another use of Digital Twins is for the machines being used in the manufacturing. These machines and assembly lines can also be optimized using their own Digital Twins.
• Field operations of the industrial asset: Once the physical asset is shipped and being used in the field operations, Digital Twins can be used to drive asset optimization and including the operating energy efficiency of the asset. Let us take the example of the fuel efficiency of the aircraft which is determined by how efficiently the fuel combustion is taking place in the get engines. By recording the temperature profile inside the jet engine’s fuel nozzle, the Digital Twin can help to diagnose any inefficiency in the combustion process. This helps to prolong the life of the expensive asset as often reduce the operating costs as well.
• Predictive maintenance: Digital Twins can be used to predict the right time to perform the maintenance on critical infrastructure assets, thereby reducing the unscheduled downtime. In addition, Digital Twin is often able to predict what life limited or consumables parts are likely to require replacement, during the maintenance activity. This can improve the overall efficiency of the maintenance process and reduce the downtime of the asset, from the operations.
• Asset retirement: Digital Twins can be used for the over asset lifecycle management and decide when to de-commission an asset and or replace it, in the field. The asset life models can help to make such decisions.

There can be many other manifestations that include the attributes required.

What it is not?

It is very helpful to list examples that do not qualify under the definition. Below are some examples:

• Computer Aided Design drawing: This is a digital representation of a physical object, but does not contain a model showing how it will behave in use.
• Historical process data: While the data is derived from a physical process, there is no behavior until it is used for analysis or modeling
• Open loop control configuration: Configuration of logic or algorithms alone are not a digital twin. Until there is simulated feedback to show the behavior, it is not a digital twin.

Any attribute that is lacking disqualifies the application as a digital twin.

Conclusion

This article is not the first attempt to provide clarity on the definition of Digital Twin, and it may not be the last. Some other examples include:

• In March 2021, Louise Wright and Stuart Davidson published “How to tell the difference between a model and a digital twin”
• In October 2025, Jonas Berg posted “Digital Twin Has a Standard Definition” on LinkedIn. It is an excellent description of the definition in ISO23247-1:2021 clause 3.2.3. I differ with his evaluation of what is and is not a digital twin in that I have a broader interpretation of a process model. For an example, Jonas stated that MPC only models gains but does not account for loads and disturbances and does not predict. MPC in fact does include disturbances in the control matrix, which is a matrix of dynamic responses, and does dynamically predict the Control Variables (CV) from these models. While the CV will be bound to a physical sensor or a virtual online analyzer (a dynamic prediction of the CV), MPC uses its internal CV prediction for planning control moves.

It is hoped this article will provide the reader with a better understanding amidst the confusion that exists in industry.