Machine-Learned Reduced Transport Models
Reduced transport surrogates learn the mapping from plasma profiles to transport coefficients, enabling whole-discharge simulation at interactive speed.
What a reduced model does
Integrated tokamak modeling evolves the plasma temperature, density, and current profiles by solving 1D transport equations with source and transport terms. The transport coefficients are the bottleneck: computing them from first principles is far too slow for a full discharge. A reduced transport model supplies these coefficients cheaply, and machine learning provides a flexible way to build one.
Surrogate structure
The surrogate takes local plasma parameters and their gradients and returns diffusivities and pinch velocities for heat, particles, and momentum. It is trained on outputs from higher-fidelity turbulence codes and then queried at every radial grid point and timestep inside the transport solver.
Numerical demands
- Smoothness: discontinuous outputs destabilize the implicit transport solve
- Monotonicity where physics requires, so gradients drive the expected flux direction
- Bounded outputs to prevent runaway diffusivities
- Fast derivatives for Jacobian assembly in implicit time stepping
Because the transport solver couples the surrogate back to the evolving profiles, prediction errors feed forward and can compound. This makes robustness more important than pointwise accuracy: a slightly biased but smooth model often yields more trustworthy discharge simulations than a sharper but noisier one.
Applications
Reduced models let designers scan operating scenarios, test control strategies, and study profile stiffness for machines under development. For the Hyperion breeder concept they help estimate whether a target confinement can be reached at the intended plasma current of 9.86 MA, informing the operating point before any hardware exists.
These are computational tools. Every predicted profile is a simulation result conditioned on the fidelity of the training data and the validity of the transport assumptions, and is reconciled with dedicated turbulence and equilibrium codes before being used in design decisions.