An MES runs production and keeps the record of what happened. A digital twin is a computer model of a product, machine or line, kept in step with live data, and used to predict results or test changes before you make them. The MES is one of the best data sources a production twin can have, so it makes sense to build the record first.
Unlike most systems an MES gets compared to, a digital twin is not a product category with a fixed function. The term covers everything from a 3D view of a machine to a physics model of a turbine.
So start by asking which twin someone means. The answer changes the data it needs, and how much it overlaps with an MES.
Which twin are we talking about?
- Product twin. A model of a product and its behaviour, usually tied to design tools, closer to PLM than to MES.
- Asset twin. A model of one machine, used for condition or failure prediction. It feeds on high-rate sensor data, closer to a historian.
- Process twin. A model of a physical or chemical process, used to tune settings. It is common in process industries.
- Production system twin. A model of a line or plant: flow, buffers, schedules and crews. This one overlaps the most with an MES, so the rest of this page focuses on it.
The ISO 23247 series sets out a framework for digital twins in manufacturing, if you need a formal reference. A simulation run once on a static data set is usually not called a twin, since nothing keeps it in step with the plant.
The core difference
The MES deals in facts and instructions. It says this order ran, this lot was used, this stop took seven minutes, and now run this job.
The twin deals in estimates. It says if you add a buffer here, output should rise by this much.
You audit an MES record. You validate a twin, and you expect it to be somewhat wrong, because every model is a simplification.
| MES | Production twin | |
|---|---|---|
| Answers | What is happening, and what happened? | What would happen if? |
| Measured by | Complete, correct records | How close predictions come to reality |
| If it fails | Dispatch and recording suffer today | A study or a decision gets delayed |
| Main users | Operators and supervisors | Engineers and planners |
| Acts on floor | Yes, in daily work | Usually through people or other systems |
How the data flows
Order mix, cycle times, stops by cause, changeovers and shift patterns.
Replay a past week and compare modelled output with what the MES recorded.
A new buffer, another crew or a different sequence, with no risk to the line.
The MES shows the real result. If it differs, fix the model.
Some twins send results back into the plant, such as a recommended sequence to a scheduler. Others just inform a decision made in a meeting, which is still a valid use.
Letting a production twin control machines directly is a big step. If a twin writes to a machine, it should go through the same limits and reviews as any other Level 3 system in ISA-95, and the ISA-95 levels guide shows where that sits.
Connections are the usual ones, such as MES APIs, OPC UA for machine signals, and historians for process data. MES system architecture shows where those sit.
Example: why a model needs measured data
Example: three stations in series with a nominal cycle of 30 seconds. A model built from nameplate data predicts 120 parts an hour.
The MES shows the line actually delivers 94 parts an hour. Short stops, restart losses and starved stations explain the gap, as blocked and starved describes.
When the team feeds in measured cycle times and stop patterns, the model gives 96 parts an hour. That is close enough to test ideas with.
The team then tests a buffer before the second station, and the model predicts 101 parts an hour. After the real change, the MES measures 99, so the estimate was slightly high but pointed the right way.
The check on that number is (120 - 94) / 94 = 27.7% too high for the nameplate model, and (96 - 94) / 94 = 2.1% for the calibrated one. Without measured data, the first model looks just as convincing on screen.
Cycle variation matters as well. A model using an average only will overstate flow, as cycle time variation explains.
What gets called a twin, and should not
Marketing blurs this more than the technology does. A 3D view of the plant coloured by live MES states is often sold as a twin.
That view can be a useful display, but it predicts nothing. Ask one question of anything called a twin: what can it tell me that has not happened yet?
Likewise, an MES dashboard with a trend line is reporting and not a model. See MES dashboards for what those should show.
Readiness: should you build one?
A twin pays off when decisions are expensive and hard to test on the real line. A new line layout, a costly buffer or a new shift model are good cases.
Before committing, run through this short list.
- Do you have at least a few months of event data? Stops, cycle times and the order mix are the minimum.
- Do you trust it? If downtime data accuracy is poor, the model inherits it.
- Is there one decision worth the work? Start from a real question, and give it a date and a cost.
- Who maintains it? The line changes, and an unmaintained twin is an old simulation.
When a twin can come first
Knowing where the shift goes usually gives you plenty to fix with no model at all. So the record comes first for most plants, and many are not there yet.
Two cases reverse that order. A line not yet built has only design data to work from, and an asset twin can feed straight from sensors, with production records playing a small part.
Running both pays best at large or capital-heavy sites with stable MES data and repeated decisions about layout, capacity or schedule.
Small plants can skip the model. A spreadsheet and a calculator such as production capacity answer most capacity questions at this size.
Where twin budgets go wrong
- Starting with the 3D picture. It impresses visitors and answers few questions, so build the logic first.
- Calibrating to nameplate values. As the example shows, this overstates output by a wide margin.
- No validation step. If nobody replays a past week, nobody knows how wrong the model is.
- Buying a twin to find out why the line stops. That question needs measurements, as OEE data collection methods covers.
Any model of your line is only as good as the measured data underneath. Fabrico is a light MES for OEE: it collects stops, speed and counts live from PLCs or IoT sensors on older machines, with add-on AI cameras counting output where a machine has no usable signal, giving you OEE per line and shift. See how Fabrico connects to machines.
What people ask next
Is a digital twin the same as a simulation?
Not quite. A simulation can run once on fixed data. A twin is kept in step with its real counterpart through data, so the model stays current. Many production twins use simulation inside them.
Does a digital twin replace an MES?
No. A twin predicts, and it does not dispatch orders, hold lots or keep the production record. It usually relies on an MES or similar for data, as described in what is MES.
What data does a production digital twin need?
Cycle times by product, stops by cause and duration, changeovers, order mix, shift patterns and buffer sizes. The shop floor data collection guide shows how to capture them.
Is OEE useful for a digital twin?
Yes, as a check. The factor breakdown shows if the model loses the right amount to stops, speed and scrap, not only the right total. Use the six big losses calculator to set out the measured losses.
How is a digital twin different from a digital thread?
The thread is the linked data trail of a product from design through production and service. The twin is a working model that uses some of that data. MES vs PLM covers where the thread begins.
Fabrico is a light MES for OEE. It takes stops, speed and counts from PLCs or IoT sensors, with add-on AI cameras counting output where a machine has no usable signal, and operators tag the cause at the line.