ProcessDigital Twins
How to build a digital twin that improves one operating decision first
To build a digital twin that earns its keep, start from one recurring operating decision, then choose the least complex model that can answer it, connect only the data that model needs, calibrate it against observed behavior and show its uncertainty. Keep the twin advisory until its predictions have been checked against reality. Expand scope only when the first decision is demonstrably better informed.
On this page
Start with a decision, not a replica
Many twin projects begin by modeling everything that can be modeled and end with an attractive visualization nobody consults when it matters. The alternative is to scope backwards from a decision that recurs and has consequences: whether to accept a load change, when to take a line down for changeover, how to sequence pumps through a heatwave. NIST describes a digital twin as an electronic representation that lets users view the states of a real entity and the transitions between them1; the minimum viable twin is the smallest such representation that changes how one decision is made.
A good first decision has four properties. It is made often enough to test predictions against outcomes. Someone owns it and is willing to use new evidence. The physical behavior behind it is understood well enough to model. And the data needed to describe the current state already exists or can be collected within the pilot. If no candidate decision has all four, the right first project is usually instrumentation, not a twin.
Matching model fidelity to the question the twin must answer
| Model type | Physics-based | Statistical or ML | Process simulation | Hybrid |
|---|---|---|---|---|
| Answers well | What happens under conditions never seen before | What usually happens under familiar conditions | How queues, buffers and schedules interact | Physical behavior with learned corrections |
| Needs | Engineering parameters and equations | Plenty of history covering the operating range | Process maps, cycle times and rules | Both, plus discipline about which part does what |
| Weak when | Parameters drift or are unknown | Asked to extrapolate beyond history | Physical limits matter more than flow | Nobody can tell which component caused an error |
| Typical first use | Thermal, hydraulic or structural limits | Short-term demand or energy estimates | Throughput and changeover planning | Assets with good physics but aging behavior |
Hybrid is highlighted because it is often where a first twin ends up, not because it should be the default starting point.
From one decision to a working first twin
- Name the decision
Who decides, how often, with what information today.
- Choose fidelity
The simplest model type that can answer the question.
- Inventory the data
Assets, identifiers, signals, sampling rates and gaps.
- Build the state model
Current condition of each asset, with freshness flags.
- Calibrate and bound
Fit to observed behavior and record the validated range.
- Decision views
Baseline versus scenario, with uncertainty shown.
- Expansion review
Compare predictions with outcomes before adding scope.
Eight scoping steps with their outputs and owners
Write the decision brief
Describe the decision, its frequency, the options considered, what goes wrong today and how you would know the twin helped. Keep it to one page.
Select the model type
Use the fidelity table to choose physics-based, statistical, process or hybrid modeling. Prefer the simplest type that answers the question, and note what it cannot answer.
Inventory assets and signals
List each asset in scope, its identifier in every system that mentions it, the signals available, their units, sampling rates, typical gaps and who owns them.
Build the state model
Assemble the current state of each asset from those signals, carrying the time of the last good reading. Decide how the twin behaves when a value is stale, missing or out of range, and make that visible rather than silently filling it.
Calibrate and record the validated range
Fit model parameters to historical behavior, test on a separate period, and write down the operating conditions over which the fit was checked. Outside that range the twin should warn the user.
Design decision views with uncertainty
Show baseline and scenario side by side, with ranges rather than single lines where uncertainty is material, and list the assumptions behind each scenario.
Keep the twin advisory
Route scenario results to a person who decides and acts through existing controls. Any later move toward automated setpoints is a separate change with its own safety review.
Run the expansion review
After a period of real decisions, compare predictions with what happened. Expand to new assets or decisions only if the first one is better informed and the model held up.
Data questions to answer before modeling begins
Questions and answers
Does a digital twin need a 3D model?
Only if spatial relationships matter to the decision, such as access, clearances or line-of-sight. Many useful twins are charts, schematics and tables tied to live state and a calibrated model. A 3D view can help people orient themselves, but it adds cost and maintenance, and it does nothing to make the underlying predictions more accurate.
How long does it take to scope a first digital twin?
Scoping itself, meaning the decision brief, modeling approach and data inventory, is usually a matter of weeks rather than months when the right people are available. Building and calibrating takes longer and depends mostly on data quality. If the inventory reveals large gaps, plan an instrumentation phase before committing to a build date.
Who should own a digital twin once it is running?
Ownership is usually split. Operations owns the decision and the use of the twin, an engineering or modeling custodian owns calibration and the validated range, and data owners keep each feed healthy. Write these roles down at the start, because a twin whose calibration nobody maintains quietly becomes wrong while still looking authoritative.
When should a first twin be stopped rather than expanded?
Stop or rescope when predictions do not hold up against outcomes after recalibration, when the decision owner is not using the results, or when the data needed to keep the state model current cannot be sustained. Stopping early is a valid result: it shows where instrumentation or process knowledge is missing before a larger investment.
Sources
- NIST IR 8356: Security and Trust Considerations for Digital Twin Technology — NIST · checked 10 October 2026