Surrogates for Fusion Simulation
Fusion design leans on expensive multiphysics codes; surrogates make the scans, optimization, and uncertainty studies those designs require affordable.
The computational burden
Designing a fusion machine couples plasma transport, magnetohydrodynamic equilibrium, electromagnetics, neutronics, and thermal-structural mechanics. Each of these can be a multi-hour high-fidelity computation. Design exploration, optimization, sensitivity analysis, and uncertainty quantification each demand many such evaluations, which surrogates make tractable.
Where surrogates enter
- Emulating transport codes to map confinement across operating parameters
- Fast equilibrium reconstruction for control and scenario studies
- Neutronics emulation for shielding and breeding estimates
- Structural response surfaces for electromagnetic and thermal loads
- Real-time twins for scenario planning
Field outputs
Fusion quantities of interest are often fields - profiles across the plasma, flux maps, stress distributions - not single numbers. Reduced-order approaches compress these fields with proper orthogonal decomposition or autoencoders, and a surrogate predicts the reduced coefficients. Neural operators are an emerging option for learning field-to-field maps directly.
Uncertainty is central
Fusion physics carries genuine uncertainty in transport coefficients, atomic data, and model form. Surrogates that report error bars - Gaussian processes and Bayesian methods - let designers propagate this uncertainty and see it in the predicted performance envelope rather than a single optimistic figure. This is why uncertainty-aware surrogates dominate credible fusion design work.
The honesty constraint
A surrogate emulates a model, and a model approximates reality. A surrogate prediction of machine performance is therefore two steps removed from a measured result and must be labeled as such. Simulated ranges with stated conditions are not hardware demonstrations, and no surrogate output substitutes for an experiment.
In Kronos work
For the Hyperion breeder and the burner designs, surrogates of the underlying multiphysics codes enable parameter scans, optimization, and uncertainty studies in interactive time. Frozen design points - such as the breeder's peak field and plasma current, or the burner's plug field and neutron fraction - come from full-fidelity computation; surrogates accelerate exploration around them and are always validated against those full-fidelity runs before informing a decision. These machines are, at this stage, design and simulation - not built hardware.