Digital twin simulation, explained
A digital twin is only as useful as the model inside it. This guide explains how CFD, reduced order models and the executable digital twin make a twin predictive, and which Simcenter tools build each part.
What is a digital twin?
A digital twin is a virtual model of a real product, machine or process that is fed with data from it and uses simulation to predict how it behaves, so that people can make better decisions about the real thing. In a mature twin the link runs both ways: decisions go back to the asset as closed-loop feedback.
Siemens describes a digital twin as “a virtual model mirroring a real-world object or system using sensor data and simulations”, and adds that a fully mature twin can be connected to the physical entity “to provide closed-loop feedback”.
The word covers two things. A twin of a design sits close to an ordinary simulation model: it describes every pump built to that design. The twin of pump number 214 is tied to one physical pump and also carries what has happened to it: its measured operating conditions, its service history and its current state. The people who defined the concept call these the prototype and the instance (see Types of digital twin). Most of this guide is about the second kind, where simulation meets live data.
Keeping an instance twin in step means two things: feeding it measurements, and recording what changes on the real asset, such as parts replaced or overhauled, so that the model still describes the asset as it is today.
How a digital twin works
A twin of an asset in use runs this loop:
- 1Physical asset
A product, machine or plant in use, or a prototype on a test rig, with a limited number of sensors.
- 2Data
Measurements and operating conditions: loads, temperatures, flows, speeds.
- 3Model
Simulation, a reduced order model or AI turns the data into a full picture, including values where no sensor can be fitted, and predicts what happens next.
- 4Decision
Adjust a setting, plan maintenance or change the next design. In a mature twin this goes back to the asset directly, autonomously or with a person in the loop.
The model is what makes the loop predictive. Siemens notes that digital twins “often incorporate simulation and modeling techniques” to test how an asset behaves under different conditions. Without a model, a twin is a dashboard of past measurements; with one, it can answer “what if?”
Digital twin vs simulation
A simulation model describes a design or a scenario with assumed inputs. A digital twin of a built asset is tied to that one asset, uses measured data from it and is used through its life. The twin uses simulation to make its predictions; it is not an alternative to it.
| Aspect | Simulation model | Digital twin of a built asset |
|---|---|---|
| Represents | A design or a scenario | One specific physical asset |
| Input | Assumed loads and conditions | Measured data from that asset |
| Used | Mainly while designing | Through the asset’s life, including operation |
| Link to the real thing | None needed | Data in; in a mature twin, decisions back out |
| Speed | As long as the physics needs | Fast enough to keep up, often real time |
A twin of a design sits in between: it starts from requirements, CAD and simulation, and becomes an instance twin once it is linked to one physical product. Our guide to CAE simulation explains how such models are built and checked against tests.
How CFD and simulation feed a digital twin
The physics in a twin has to be accurate and fast enough to keep up with the asset. Those demands pull in different directions, so twins are built in stages:
- 1Detailed model
CFD, finite element or system simulation, validated against tests. Simcenter STAR-CCM+, Simcenter 3D, Simcenter Amesim, Simcenter Flomaster.
- 2Reduced model
A reduced order model or AI surrogate, derived from or trained on the detailed model, small enough for real time.
- 3Executable twin
The model packaged to run on its own, as an FMU or, for an AI surrogate, as an ONNX model, on an edge device or in the cloud and connected to live data.
System models are often fast enough to skip the middle stage: Siemens lists full-fidelity system simulations among the forms an executable twin can take, and its Flomaster example below runs live beside a plant. Reduction matters most for 3D CFD and finite element models.
What is a CFD digital twin?
A CFD digital twin takes its fluid flow and thermal physics from computational fluid dynamics. A full CFD run is too slow to keep up with the asset, so its results train a reduced order model or an AI surrogate, which runs in the twin and estimates the flow field from a few measured inputs.
Offline, once
- Design studySimcenter STAR-CCM+ Design Manager runs the CFD model across the operating conditions the asset will see. Our DOE calculator estimates how many runs that takes.
- Train the ROMA scalar model for a few outputs, such as pressure drop, or a field model that rebuilds the whole 3D field.
- Check itAgainst runs left out of the training.
- ExportAs an FMU (Functional Mock-up Unit, the model format of the FMI standard) or an ONNX (Open Neural Network Exchange) model.
Online, while the asset runs
- Live inputsMeasured boundary conditions, such as inlet flow and temperature.
- ROM in the twinEvaluates in real time.
- OutputThe field, or a virtual sensor value at any point in it.
In Siemens’ example, an interpolation POD model of an exhaust manifold takes six inputs (mass flow, total temperature and turbulence intensity at the runner inlets) and returns six fields, including temperature, velocity, turbulent kinetic energy and wall y+. It reproduces its training runs exactly (100%), as an interpolating model does by construction. What counts is the 97.8% it reaches on runs it was not trained on; turbulent kinetic energy is the hardest output, at 93.5%.
Two limits are worth knowing. A ROM is only reliable inside the range of conditions it was trained on, and it inherits any modelling error in the CFD behind it, so the detailed model has to be validated first. Transient behaviour needs a dynamic ROM; Volupe’s Fabian Hasselby shows one in his article on time-dependent neural networks with Amesim’s ROM Builder.
Reduced order models (ROMs)
A reduced order model reproduces the response of a detailed model at a fraction of the cost. There are two broad families: models derived from the equations of the detailed model, such as the reduced finite element model behind Simcenter 3D Smart Virtual Sensing, and models fitted to its results, from response surfaces to neural networks. In Siemens’ words, a ROM “has a small memory footprint, is tool-agnostic and can be operated in real-time”.
Simcenter Reduced Order Modeling builds, validates and exports such models from simulation and test data, and the exported models “do not require a license or solver to run”. Simcenter Amesim has its own ROM Builder that turns training data into an Amesim block.
AI surrogates: Simcenter PhysicsAI
Simcenter PhysicsAI, which came to Simcenter from Altair, trains AI surrogate models on simulation results and gives near-instant predictions for new geometries, with the high-fidelity solver kept as the reference. Since Simcenter STAR-CCM+ 2602 it is also available as an add-on in Design Manager.
Siemens positions it to “accelerate AI-powered design exploration”: it predicts for new shapes. For a twin in operation, a ROM trained over the asset’s operating conditions is the closer fit.
The executable digital twin (xDT)
An executable digital twin (xDT) is, in Siemens’ definition, “a real-time capable behavior model which can be connected to the live IoT data stream from a physical system”. The packaged xDT contains the physics-based behaviour model and the solver needed to run it.
“In simple terms, the xDT is the digital twin on a chip.”
Siemens Digital Industries Software
Siemens describes it in five steps: model the behaviour; package it with its solver; leverage it at scale on a connected edge device or in the cloud; deploy it so that it connects to the live asset; and operate it, so that engineering insight is available beyond design. The packaged models use open standards, FMI and ONNX.
The first use Siemens names is virtual sensing: with data from a relatively small number of sensors, the xDT uses reduced order models to predict the physical state elsewhere on the object, in Siemens’ words “even in places where it would be impossible to place sensors”. Siemens also names cost-based design optimisation, virtual PLC testing before a system is built, and operations through industrial IoT.
Two Siemens examples
A plant with a virtual sensor. Siemens connected a Simcenter Flomaster model to the SIMATIC control of a real plant. The model runs in real time beside the plant, augments the measured data and provides leak detection.
Loads from a few sensors. Simcenter 3D Smart Virtual Sensing combines a limited set of measurements with a reduced order finite element model to create virtual sensors. In Siemens’ robot-arm demonstrator, a few sensors give the load across the whole part in real time.
Types of digital twin
There is no single standard list. Three framings are useful.
Prototype and instance (Grieves and Vickers)
- Pump #211instance twin
- Pump #212instance twin
- Pump #213instance twin
- Pump #214instance twin, its own history
- Pump #215instance twin
The concept’s originators separate the Digital Twin Prototype, which describes a design (requirements, the annotated 3D model, bills of materials and processes), from the Digital Twin Instance, which stays linked to one specific physical product for its whole life, including its sensor data and service record. Twins follow the product through creation, production, operation and disposal.
Product and production twins (Siemens)
Siemens talks about the product digital twin, the models of the item itself, and the production digital twin, the models of the facilities that make it. The two interact during manufacturing. A digital thread captures data through the lifecycle, so that changes to the twin are documented and traced back to system models and requirements.
Component, asset, system and process twins (IBM)
IBM groups twins by scale: component (or part) twins of single components, asset twins of functional units made of several components, system or unit twins of how assets fit together, and process twins of how systems work together across a production facility.
Simcenter tools for building a digital twin
These are Simcenter’s digital twin capabilities, stage by stage. Volupe licenses, trains and supports the tools that link to a Volupe product page; Simcenter Reduced Order Modeling and the Executable Digital Twin Gateway are part of Siemens’ wider Simcenter portfolio.
| Tool | Role in a digital twin |
|---|---|
| Detailed model | |
| Simcenter STAR-CCM+ | CFD and multiphysics; Design Manager runs the sweeps a ROM is trained on. |
| Simcenter 3D | Structural FE models; Smart Virtual Sensing combines them with measurements. |
| Simcenter Amesim | System simulation; ROM Builder turns training data into an Amesim block. |
| Simcenter Flomaster | 1D thermo-fluid networks; can run in real time beside a plant (Siemens example). |
| Reduced model | |
| Simcenter HEEDS | Design exploration, surrogate models and the parameter studies behind them. |
| Simcenter Reduced Order Modeling | Builds, validates and exports ROMs to FMI, ONNX or Amesim. |
| Simcenter PhysicsAI | AI surrogates for design exploration on new geometries, trained on CFD or FE results. |
| Simcenter Flotherm | Electronics cooling; the separately licensed BCI-ROM module extracts boundary-condition-independent thermal ROMs, exported as FMUs for Amesim or Flomaster. |
| Executable twin | |
| Simcenter Executable Digital Twin Gateway | Deploys and manages xDT instances (Siemens). |
How to start a digital twin project
Most twins start smaller than the word suggests. A practical order:
- 1Pick one decision
The question the twin should answer: when to service, how hard to run, what to change next.
- 2Check what you measure
The sensors and test data you have, and where a virtual sensor would have to fill in.
- 3Validate a detailed model
Build the CFD, FE or system model and compare it with test data first.
- 4Reduce and compare
Train a ROM or AI surrogate and check it against detailed runs it has not seen.
- 5Deploy and keep checking
Run on live data, record changes to the asset and keep comparing with measurements. When predictions drift, re-fit or retrain the model.
Volupe helps with steps 3 and 4: Simcenter licences, training and technical support. Volupe engineers’ articles on ROMs and surrogates are collected below.
History of the digital twin
According to Michael Grieves and John Vickers, who wrote the concept’s history in 2017:
- 2002
Grieves presents “Conceptual Ideal for PLM” at the University of Michigan: real space, virtual space and the data and information links between them.
- 2003
The model is taught in the first PLM courses as the Mirrored Spaces Model, later the Information Mirroring Model.
- 2010
NASA uses the concept in its technology roadmaps.
- 2011
Grieves’ book Virtually Perfect adopts John Vickers’ term for the model: the “digital twin”.
Sources and changelog
Sources
Siemens, Digital twin.
Siemens Digital Industries Software, “Defining the digital twin”, white paper, 2024.
Siemens, Executable digital twin (xDT) and Simcenter Reduced Order Modeling.
Siemens, “Simcenter Executable Digital Twin – The leap forward in digital twin technology”, 28 April 2025.
Siemens, Simcenter PhysicsAI, 27 May 2026.
M. Grieves and J. Vickers, “Digital Twin: Mitigating Unpredictable, Undesirable Emergent Behavior in Complex Systems”, in Transdisciplinary Perspectives on Complex Systems, Springer, 2017, pp. 85–113.
IBM, What is a digital twin?
Images: Siemens Simcenter image library.
Changelog
October 2026: rewritten with Siemens and published sources; CFD digital twins, reduced order models, PhysicsAI and the executable digital twin added; unsourced market figures and claims removed.
ROMs and surrogates in practice: articles by Volupe engineers
All articlesROM Builder
time-dependent neural networks
Simcenter Amesim
Time-dependent neural networks with Simcenter Amesim’s ROM Builder
Fabian Hasselby, Volupe · 2021
Read the article →
2D field ROM
static fields in HEEDS
Simcenter HEEDS
Reduced order modelling for 2D static fields
Fabian Hasselby, Volupe · 2024
Read the article →
PhysicsAI add-on
Design Manager, release 2602
Simcenter STAR-CCM+
The PhysicsAI add-on for Simcenter STAR-CCM+
Fabian Hasselby, Volupe · 2026
Read the article →
PhysicsAI workflow
from a STAR-CCM+ study
Simcenter PhysicsAI
A PhysicsAI workflow based on a Simcenter STAR-CCM+ study
Christoffer Johansson, Volupe · 2026
Read the article →
Surrogates tab
new in HEEDS 2504
Simcenter HEEDS
Smarter tools for better surrogate results
Florian Vesting, Volupe · 2025
Read the article →
Surrogate models
reliability and optimisation
Simcenter STAR-CCM+
Surrogate models in Simcenter STAR-CCM+
Robin Victor, Volupe · 2021
Read the article →
Questions about digital twins
Questions about the model behind a twin?
Volupe’s simulation engineers
CFD, system simulation and ROMs
Fabian Hasselby wrote three of the articles above. We reply within one day.
Ask a technical question →
What is a digital twin in simple terms?
A virtual model of a real product, machine or process that is fed with data from it and uses simulation to predict how it behaves, so you can make better decisions about the real thing.
What is the difference between a digital twin and a simulation?
A simulation model describes a design or a scenario with assumed inputs. A digital twin of a built asset is tied to that one asset, uses measured data from it and is used through its life. The twin uses simulation to make its predictions.
What is a CFD digital twin?
A digital twin that takes its fluid flow and thermal physics from CFD, for example from Simcenter STAR-CCM+. Because a full CFD run is too slow to keep up with the asset, its results train a reduced order model or an AI surrogate, which runs in the twin.
What is an executable digital twin (xDT)?
Siemens’ term for a real-time capable behaviour model that can be connected to the live data stream from a physical system. It packages the model with the solver needed to run it, so it can run on an edge device or in the cloud next to the asset.
What is a reduced order model (ROM)?
A compact model that reproduces the response of a detailed simulation or test at a fraction of the cost, either derived from the model’s equations or fitted to its results. It has a small memory footprint and can run in real time.
Which Simcenter tools are used to build a digital twin?
A detailed model in Simcenter STAR-CCM+, Simcenter 3D, Simcenter Amesim or Simcenter Flomaster; a reduced model from Simcenter Reduced Order Modeling, Amesim’s ROM Builder, Simcenter HEEDS or Simcenter PhysicsAI; and packaging through FMI or ONNX. See Simcenter tools for building a digital twin.
Do I need sensors before I can start?
Not for a twin of a design: it can start from requirements, CAD and simulation. A twin of one built asset needs data from that asset, but often from only a few sensors, because virtual sensors fill in the rest.
Who came up with the digital twin?
According to Michael Grieves and John Vickers, the concept goes back to a presentation Grieves gave at the University of Michigan in 2002; the name “digital twin” came from Vickers, of NASA.
Talk to an engineer in your region
Every Volupe business developer holds an engineering degree. Find your regional contact below, or send us a message.
Ask us about Digital twin
Use this form for licence quotes, demos, training and technical questions.
What happens next
- A reply within one day.
- A scoping call with one of our engineers.
- A quote, demo or other next step that fits your enquiry.
After purchase: support from Volupe simulation engineers.