# The digital twin: a living copy of the plant

> A model of the plant, fed continuously by its real data, to simulate, optimise and anticipate. The digital twin is appealing — but its value rests entirely on the quality of the model and the data.

- Source : https://industryhub.site/en/learn/data-ai/jumeau-numerique
- Track : Data & AI
- Level : Practitioner
- Reading time : 6 min
- Language : English

## More than a 3D mock-up

The term "digital twin" often evokes a nice 3D image of the plant. That is a mistake. A mock-up is **frozen**; a digital twin is **living**: it is a model of the plant fed continuously by its [real data](/en/learn/data-ai/donnee-industrielle). It does not show what the plant *could* be, but what it *is* right now — and what it will become if you change a given parameter.

## The three degrees: model, shadow, twin

Not all "twins" are equal. Three degrees of connection to reality are distinguished:

- the **digital model**: a standalone simulation, with no automatic link to the plant. You test scenarios "offline".
- the **digital shadow**: reality sends its data to the model continuously, but with no return. The model reflects the current state.
- the **true twin**: the loop is closed both ways. The model receives data *and* can, in return, influence the plant's operation.

| Degree | Reality → model | Model → reality |
|---|---|---|
| Model | manual | none |
| Shadow | automatic, continuous | none |
| Twin | automatic, continuous | automatic (can act) |

Many projects sold as "twins" are in fact models or shadows — which is already very useful.

## What it is for in practice

A well-built twin lets you **simulate before acting**: test a setting, a recipe change, a bottleneck, without risking real production. It also serves to **optimise** continuously (energy, throughput, quality), to **anticipate** a drift in connection with [predictive maintenance](/en/learn/data-ai/maintenance-predictive), and to **train** operators on a risk-free copy. It is one of the most cited industrial use cases of [AI in industry](/en/articles/ia-industrie-enjeux-defis).

## What feeds it: model + data

A twin rests on two pillars. On one side a **model**: physical (process equations), statistical ([machine learning](/en/glossary#term-ml)) or hybrid. On the other a reliable, contextualised **data flow**, rising from the field through the controllers and a protocol such as [OPC UA](/en/glossary#term-opc-ua), often enriched by the production context of the [MES](/en/glossary#term-mes) and [IIoT](/en/glossary#term-iiot) connectivity. One without the other produces nothing usable.

## The limits and the trap

The classic trap is to confuse the render with the value. A magnificent 3D model fed by false data, or an over-rough physical model, gives elegant but wrong answers — and so dangerous ones if you follow them. Building a twin is expensive in modelling and data integration; it is only justified where simulation truly pays. As for the whole data chain, the rule fits in one sentence: **the twin is never worth more than the model and the data feeding it**.

## Key takeaway

A digital twin is a model of the plant fed continuously by its real data. Depending on the degree of connection, you speak of a model, a shadow or a true twin. It serves to simulate, optimise and anticipate — but is only worth the quality of the model and data feeding it.