Surrogate Models for Twins
Surrogates are fast approximations of expensive physics codes, letting a twin run inside real-time and assimilation loops.
Fast stand-ins for heavy physics
High-fidelity fusion codes, for plasma equilibrium, neutron transport, or structural response, can take minutes to days per run. A twin that must react within a control cycle or run hundreds of ensemble members cannot wait. A surrogate model is a fast approximation, trained on runs of the expensive code, that reproduces its outputs in milliseconds while carrying an estimate of its own error.
How surrogates are built
- Sample the expensive model across the operating envelope using a space-filling design
- Fit a fast approximator to the input-output pairs
- Test on held-out points to measure accuracy and flag where the fit is weak
Common families
Reduced-order models compress the physics onto a small number of dominant modes and remain interpretable. Gaussian processes give a prediction with calibrated uncertainty and suit smaller datasets. Neural networks handle large, high-dimensional data. Each is covered in its own page: reduced-order models, Gaussian processes, and neural networks.
The uncertainty obligation
A surrogate is only safe in a twin if it knows where it is unreliable. Outside the region it was trained on, a surrogate can be confidently wrong. A trustworthy surrogate reports an error estimate and refuses to extrapolate silently, handing back to the full physics code when the query leaves its valid domain.
Role in the Kronos twins
Surrogates are what make a physically rich twin of the Hyperion breeder or the burner run fast enough to matter: to reconstruct equilibrium inside the control cycle, to power hundreds of ensemble members for uncertainty, and to sweep thousands of what-if scenarios overnight. They are trained today on the same validated simulations that anchor the three-dimensional model, and they will be retrained against real data once the machines operate. See twin-in-the-loop control.