Machine-Learned Profile Prediction
Models predict steady-state or evolving temperature, density, and current profiles from engineering settings, accelerating scenario studies.
Profiles as the target
Plasma performance is encoded in radial profiles of temperature, density, and current. Predicting these profiles from engineering settings, without a full transport simulation for every case, accelerates scenario design and control. Machine learning provides fast profile predictors trained on simulations or archived discharges.
Two prediction modes
- Steady-state: map engineering parameters to the settled profile shape
- Time-dependent: forecast how profiles evolve given actuator trajectories
- Boundary-conditioned: predict core profiles given a pedestal boundary
Steady-state predictors are simplest and useful for screening operating points. Time-dependent predictors are harder but support control, since they anticipate how the plasma responds to actuator changes over the next interval.
Representing profiles
Profiles can be output as values on a fixed radial grid, as coefficients of a basis expansion, or as parameters of a shape function. Basis and parametric outputs enforce smoothness and reduce dimensionality, while grid outputs are flexible but need smoothness regularization to remain physical.
Coupling to physics
Pure data-driven profile predictors risk violating physics, so they are often combined with, or constrained by, transport models. A predictor gives a fast estimate; a reduced transport solve confirms consistency. This keeps profile predictions grounded rather than merely plausible-looking.
Profile prediction supports finding and holding the design point of concepts like the Hyperion breeder, including operation at 9.86 MA plasma current. These are computational estimates for a machine under design, cross-checked against transport and equilibrium codes, with no performance claimed for hardware that is not yet built.