QLKNN-Style Neural Transport Models
Neural networks trained on quasilinear gyrokinetic flux databases predict turbulent transport in milliseconds, replacing runtime calls to slow physics codes.
From quasilinear theory to a network
Quasilinear gyrokinetic codes estimate turbulent fluxes by combining linear growth rates with a saturation rule, far cheaper than full nonlinear simulation but still too slow for integrated modeling that must evaluate transport at every radius and timestep. The QLKNN approach trains a neural network on a large precomputed database of quasilinear flux outputs so the transport model becomes a fast function call.
Training database
The database spans a grid of dimensionless inputs: normalized temperature and density gradients, temperature ratios, safety factor, magnetic shear, and collisionality. For each point the quasilinear code records the ion and electron heat fluxes, the particle flux, and often the dominant instability type. The network regresses these outputs from the inputs.
Why it works
- Dimensionless inputs keep the learned map compact and transferable
- Separate networks or heads per transport channel (ion heat, electron heat, particles)
- Classification of the leading mode helps enforce the correct instability regime
- Inference is fast enough to sit inside a real-time-capable transport loop
A central engineering concern is preserving the critical gradient behavior. Because turbulent flux is near zero below threshold and rises steeply above it, training must sample the threshold region densely and the network must produce a clean turn-on to avoid unphysical residual transport.
Use in integrated modeling
Embedded in a core transport solver, a QLKNN-style model lets engineers predict evolving temperature and density profiles across a full discharge in seconds. This supports scenario exploration for machines still in design, including spherical-tokamak breeder concepts, where many candidate profiles must be screened.
The model inherits the assumptions of the underlying quasilinear theory and the coverage of its database. Outside that envelope it should not be trusted, so predictions are bounded and cross-checked against nonlinear gyrokinetics for key operating points.