Surrogate Models for Turbulence
Emulating gyrokinetic turbulence results, the most expensive step in transport modeling, with learned models.
Gyrokinetics is expensive
Gyrokinetic simulation resolves the small-scale drift turbulence that drives cross-field transport. A single nonlinear run can consume large amounts of supercomputer time. Running one at every point of a profile, at every time step, inside an optimizer is not feasible, so learned emulators stand in.
Quasilinear versus nonlinear
Quasilinear models estimate saturated flux from linear growth rates times a saturation rule, and are far cheaper than nonlinear runs. Surrogates are built at both levels: some emulate the quasilinear model, others learn corrections toward nonlinear results, capturing effects the quasilinear rule misses.
- Emulate linear growth-rate spectra
- Emulate saturated flux and its spectrum
- Learn nonlinear corrections to a quasilinear baseline
High-dimensional inputs
Turbulence depends on many normalized parameters and on flux-surface geometry. Neural networks handle this dimensionality, but need training sets that cover it. Because each training point is expensive, active learning and careful design of the sampled space are essential.
Physical constraints
Fluxes should be non-negative in the expected direction, vanish below threshold, and respect known symmetries. Building these into the surrogate improves behavior in sparsely sampled regions and prevents unphysical outputs from corrupting downstream models.
Honest limits
A turbulence surrogate is only as good as the physics model it learned from and the region it was trained on. It cannot discover physics absent from its training data, and its predictions in new regimes must be checked against fresh simulations before being trusted.