Control-Oriented Modeling
Control-oriented models capture just enough physics to predict the input-output response controllers need, fast enough to run in real time.
A different goal than first-principles models
A high-fidelity plasma simulation aims to reproduce the physics as faithfully as possible and may take hours per case. A control-oriented model has a different goal: predict how the controlled outputs respond to the actuator inputs, accurately over the control bandwidth, fast enough to run inside a loop or an optimizer. Fidelity is spent only where control needs it.
Reduced models
Control-oriented models are built by reduction: linearizing around an operating point, keeping only the dominant dynamics, or fitting a low-order model to data from a high-fidelity code. The result might be a handful of states describing how current, shape, and stored energy respond, rather than millions of grid cells.
Data-driven identification
Models can also be identified directly from operating data using system-identification methods, which fit a model to the observed response to deliberate actuator perturbations. This captures the real machine's behavior, including effects a physics model missed, at the cost of validity only near the conditions where the data were taken.
Uses
- Feedforward computation in trajectory optimization
- Prediction inside model-predictive controllers
- The plant model in state estimators
- The core of flight simulators and digital twins
In the Kronos program
Kronos builds control-oriented models of the Hyperion breeder and the burner generators from its high-fidelity simulations, since the machines are not yet built. These reduced models are shared across the flight simulator, the digital twin, and the trajectory optimizer, so a scenario proven in one is consistent everywhere. After first plasma, system identification will refine them from operating data.