Reduced-Order Models
A reduced-order model keeps the essential behavior of a heavy simulation while running fast enough for control and rapid analysis.
Why reduce
High-fidelity models are accurate but too slow for real-time control or for the many evaluations an optimizer needs. A reduced-order model captures the dominant dynamics of the full system in a much smaller, faster form, trading some accuracy for the speed that certain tasks require.
How reduction is done
- Projecting the full system onto a few dominant modes.
- Fitting simplified physics that reproduce the key responses.
- Learning a compact model from high-fidelity data.
- Combining physics structure with data-driven corrections.
Physics versus data
A purely data-driven reduced model can be fast but may violate physics outside its training range. A physics-based reduction respects conservation laws but may miss effects it did not include. The strongest reduced models blend both, keeping physical structure while learning corrections from data.
Where they run
Reduced-order models sit inside real-time control, equilibrium reconstruction, scenario optimization, and the fast layers of the digital twin. They are one rung on the fidelity ladder.
Both machines
The Hyperion breeder and the tandem-mirror burner each have reduced models tuned to their physics, used wherever full fidelity is too slow to be useful.
Honesty about range
A reduced model is valid only where it was built to be, so its accuracy is characterized against the full model and its results flagged when it is pushed beyond that range, connecting to verification and validation.