Digital Twin Series · Level 1 and 2 · Accuracy
Digital Twin Fidelity and Model Accuracy
How detailed must the model be, and how do you know it is accurate enough for the decision?
Fidelity is the level of detail with which the virtual model represents the real system: its geometry, physics and behaviour. More fidelity usually means better accuracy, but also more cost, data and computing time. The right level is the lowest one that is good enough for the decision you want to make, and it must be checked against real measurements. This page walks up five fidelity levels using a conveyor positioning system, helps you pick a level, lets you break and then calibrate a model to see the error metrics, and shows how errors combine.
| Level | What the model holds | Typical accuracy | Use for |
|---|
Accuracy figures are typical values for a conveyor positioning example. Real accuracy depends on the machine, the model quality and the measurements behind it.
| Decision | Required accuracy (typical) | Recommended level |
|---|
Choose the model by the decision you want to make. Start with the lowest level that can answer it, and raise the fidelity only where the answer is not good enough.
A sketch of the trade-off, not measured data. Higher levels give finer accuracy and take longer to compute or need more data. The real-time twin runs continuously instead of in a single run. A well-calibrated simple model can beat an uncalibrated detailed one.
The real machine is a position axis that follows step commands with a lag, a small delay and sensor noise. The model has the same structure but wrong parameters. Calibration searches for the parameters that make the model match the measurements. What remains is mostly the sensor noise, which no model can remove. MAE is the mean absolute error, RMSE the root mean square error, and the error percentage here is the MAE divided by the full range of 8 mm.
Independent errors combine as the square root of the sum of squares (RSS), which is smaller than adding them. Adding them up is the worst case. Errors that are not independent, such as a shared reference, need more careful treatment.
Figures are indicative and the examples are simplified. Validate any model against your own measurements before you rely on it.