SEC-04 / Data Foundations

Digital twins for mid-market manufacturers: data, governance and security

A digital twin is only useful when it answers a specific operational question at the fidelity that question needs. Govern it like any engineering model: versioned, validated against the plant, clear on who owns the IP, and connected to OT through a reviewed, read-mostly path.

6 min read Updated Plant managersOT engineersCIOsProcess engineers

Key takeaways

  • Start from the decision the twin supports, then choose the type (asset, process, line or factory) and fidelity. Not the other way round.
  • A twin drifts away from reality as equipment wears, recipes change and PLC logic is modified. Plan for recalibration.
  • Version the model, its parameters and its input data together, and put changes through MOC when the twin informs operations.
  • OEM-supplied models often come with license limits. Clarify ownership and use rights before you build on them.
  • A twin that writes back to the plant is a control system. Treat the connection as a high-risk conduit.

"Digital twin" covers a lot of ground. To one vendor it is a 3D walkthrough of the plant with live tags overlaid. To another it is a physics model of a reactor. To an OEM it might be the simulation model they used to design your machine. All of these can be useful. Many end up as expensive screens that nobody uses after the launch event.

For a mid-market manufacturer, the right question is not "should we have a digital twin?" It is "which decision would a twin improve, and what data and governance would it take to keep it trustworthy?" This page answers that.

What a digital twin is

A digital twin is a virtual representation of a physical thing or process, kept in step with reality through data, and used to monitor, predict or test changes. ISO 23247 sets out a digital twin framework for manufacturing, covering the observable manufacturing elements, the twin itself, and how they exchange data. You do not need to adopt the standard to benefit from its vocabulary.

The key word is "kept in step." A model that is never compared to the plant is a simulation, not a twin. That is fine, as long as nobody treats its outputs as current reality.

Types of twin and what they need

Type Typical question Data needed Fidelity needed
Asset twin How healthy is this pump, press or robot? What happens if we change its duty? Condition signals, operating context, maintenance history, OEM specs Medium to high for the asset, low for surroundings
Process twin How do temperature, pressure and feed rate affect yield or quality? Process variables, lab and quality results, batch and recipe context High for the key variables, with good time alignment
Line twin Where is the bottleneck? What if we add a buffer or change the schedule? Cycle times, states, downtime reasons, changeovers, product mix Medium; discrete-event logic matters more than physics
Factory twin How do flows, labor and layout interact across the site? MES, ERP, layout, staffing, logistics Low to medium; broad but coarse

Two practical points:

  • Fidelity is expensive. Each step up in physical detail adds modeling effort, data requirements and calibration work. Match it to the decision.
  • Line twins are often the best first step. A discrete-event model built from MES and historian state data answers real capital and scheduling questions without high-frequency data.

Plant-floor reality: Virtual commissioning, testing PLC logic against a simulated line before the SAT, is one of the clearest wins for a twin in a brownfield plant. It shortens commissioning and catches logic errors before they reach the floor.

Data foundations for a twin

A twin is only as good as its inputs. Check these before you build:

  • Asset hierarchy. Twin components should map to the same ISA-95 hierarchy used in your historian or Unified Namespace.
  • Time alignment. Process twins combine signals from PLCs, historians and lab systems. Clock offsets of a few seconds can break cause-and-effect.
  • State and context. Running, idle, changeover and fault states, plus product and batch, are essential for line and process twins.
  • Quality of measurements. Sensor drift and historian compression feed straight into calibration errors. See industrial data quality.
  • Engineering data. P&IDs, equipment specs and PLC programs are often out of date. The twin will inherit those errors.

Governance

Versioning

A twin has at least four things that change: the model structure, its parameters (calibration), the input data, and the software it runs on. Version them together, so that for any output you can say which model, which calibration and which data produced it. Without that, you cannot investigate a bad recommendation.

Simulation vs reality drift

Every twin drifts. Bearings wear, heat exchangers foul, recipes change, someone edits PLC logic, and a new product runs on the line. Plan for it:

  1. Define tracked outputs. Pick a few outputs (throughput, a key temperature, energy use) to compare against measured values.
  2. Set a tolerance. Agree how far the twin may diverge before it is flagged.
  3. Schedule checks. Monthly for process twins in active use, after every relevant change for all twins.
  4. Link to MOC. Any MOC on equipment, recipe or control logic in scope should include "update or re-validate the twin."
  5. Label outputs. When a twin is out of tolerance, its outputs should say so on the screen.

Who owns the model: OEM IP

Many twins start from models supplied by an OEM or integrator. Before building on one, get clear answers to:

  • Do you own the model, or license it? For which sites and uses?
  • Can you modify and recalibrate it, or only the OEM?
  • What happens to the model and your calibration data if the contract ends?
  • Is your operating data used to improve the OEM's products for other customers?
  • Can the OEM access your live data through the twin, and how?

Add these to your AI vendor due diligence and the OT/AI vendor security questionnaire.

Recording the use case

If a twin informs operational decisions, especially through any AI or optimization component, register it in your AI use case register with a business owner, intended use and known limits.

Security of twin-to-plant connections

Most twins need live or near-live plant data. Some vendors also propose sending optimized setpoints back. These are very different risks.

  Read-only twin              Write-back twin
  (manageable)                (control system)

  Plant -> DMZ -> Twin        Plant <-> DMZ <-> Twin
   outbound only               new inbound path,
   publish-only creds          setpoint authority,
   no commands                 needs full risk review

For read-only twins, follow the same pattern as moving OT data to a cloud historian: outbound-only from a DMZ, scoped credentials, documented as a conduit.

For write-back twins:

  • Run a full AI risk assessment, including safety.
  • Keep limits and interlocks in the PLC and SIS. The twin proposes; the control system enforces bounds.
  • Prefer operator-confirmed setpoint changes over automatic ones, at least until performance is proven.
  • Document the conduit in your zones and conduits model and review it like any other remote write path.
  • Log every setpoint the twin proposes and whether it was applied.

Watch out: A cloud-hosted twin with write-back is functionally a remote control connection into your plant. Treat its security requirements accordingly.

Avoiding vanity twins

Signs you are building a vanity twin:

  • The main use case is the demo for visitors.
  • Nobody can name the decision it changes or who makes it.
  • It is a 3D model with live numbers, but no simulation or prediction.
  • It has not been compared to plant data since go-live.
  • It costs more per year to maintain than the decisions it informs are worth.

Fix: write one sentence for each twin, "This twin helps [role] decide [decision] by [method], and we check it against [measurement] every [period]." If you cannot write it, stop or rescope.

Common mistakes

  • Choosing the highest fidelity by default.
  • Building on stale P&IDs and PLC programs.
  • No versioning, so bad outputs cannot be traced.
  • Signing OEM terms that leave you unable to recalibrate or keep the model.
  • Allowing write-back during a pilot "temporarily" without review.

Where to start

  1. List the decisions you would like to test virtually: capital changes, schedules, recipes, PLC logic.
  2. Pick one with clear value and available data. A line twin or virtual commissioning project is often a good first choice.
  3. Check data readiness for that scope using the industrial data readiness scorecard.
  4. Agree model ownership and IP terms in writing before building.
  5. Connect read-only. Define tracked outputs and a tolerance.
  6. Add the twin to MOC checklists for in-scope equipment.

Effort: a focused line twin or virtual commissioning model is typically a project of weeks to a few months with an integrator. The recurring cost is calibration and change tracking, which is small if planned and large if ignored.

What good looks like

Each twin has a one-sentence purpose, a named owner and a tracked accuracy measure. Model, calibration and data are versioned together. MOC triggers twin updates. OEM terms are clear. Connections to the plant are read-only unless a reviewed, logged and bounded write path has been approved.

Sources and further reading

  1. ISO 23247-1:2021 Digital twin framework for manufacturing, Part 1: Overview and general principles (ISO)
  2. ISA-95 Standard (ISA)
  3. NIST SP 800-82 Rev. 3, Guide to Operational Technology (OT) Security
  4. NIST AI Risk Management Framework

This guide is general information for manufacturers, not legal advice. Confirm regulatory obligations with qualified counsel.

FAQ

Frequently asked questions

What is a digital twin in manufacturing?

A digital twin is a virtual model of a physical asset, process or system that is kept in step with its real counterpart through data. It is used to monitor, simulate or optimize. ISO 23247 provides a digital twin framework for manufacturing. In practice the term covers everything from a 3D model with live data overlays to a physics-based process simulator.

Do mid-market manufacturers need a digital twin?

Only if there is a decision it improves that cheaper tools cannot. Common good fits are virtual commissioning of a new line, operator training, throughput studies before capital spend, and process optimization on a well-instrumented bottleneck. If the question can be answered with historian trends and a spreadsheet, a twin is probably not worth the upkeep.

How do you keep a digital twin accurate?

Define the outputs you rely on, compare them to measured plant data on a schedule, and recalibrate when the gap exceeds an agreed tolerance. Trigger a review after equipment changes, recipe changes or PLC logic changes. Version every calibration so you can tell which model produced which recommendation.

Is it safe to connect a digital twin to the live plant?

Read-only connections through a DMZ, using the same outbound-only patterns as historian data, are manageable. Write-back from a twin to the plant, such as optimized setpoints, makes the twin part of the control system. That needs a risk assessment, MOC, limits enforced in the PLC, and usually an operator in the loop.

Related guides

Know where your plants stand

Score your AI governance, OT security and data readiness in about 10 minutes each. Results stay in your browser.