Real-Time Model Updating
A twin must revise not only its state but its own parameters as the real machine drifts, wears, and reveals mismodeled physics.
Two kinds of update
A twin updates its state continuously, but a machine also changes in ways the fixed model does not capture: surfaces erode, calibrations drift, seals age, and physics that was approximated turns out to matter. Real-time model updating is the slower revision of the model's own parameters and structure to keep it faithful over the asset's life, distinct from the fast state estimate.
Parameter estimation on the fly
The simplest form treats uncertain model parameters as extra states and lets the filter estimate them alongside the fast dynamics, a technique called state augmentation. Slowly varying quantities, such as a heat-transfer coefficient that changes as a surface fouls, are tracked this way. The filter's uncertainty on the parameter reveals when a change is real rather than noise.
Detecting structural mismatch
When the residuals, the persistent gap between predicted and measured outputs, show a pattern the current model cannot explain, the model structure itself is wrong. This is a signal to bring in higher-fidelity physics, add a missing effect, or retrain a surrogate. Distinguishing a parameter drift from a structural error is a core skill of a maintained twin.
Guardrails
- Constrain updates to physically plausible ranges so the twin never fits itself into nonsense
- Update parameters slowly relative to state, separating the two timescales
- Log every change so a prediction can be traced to the model version that made it
In the Kronos program
Before hardware exists, model updating is exercised against simulated drift to prove the machinery works. After the first-of-a-kind breeder Hyperion operates near 2030, real erosion and calibration drift will drive genuine updates. Because both machines run in fleets, the burner as Aegis and MetroVolt units, updates learned on one unit can inform priors for others while each unit keeps its own twin. See calibration and predictive maintenance.