Plasma Profile Prediction
Predicting radial profiles of density, temperature, and pressure faster than transport simulation allows.
What a profile is
Key plasma quantities vary across the minor radius: electron and ion temperature, density, rotation, and current. These radial profiles determine performance and stability. Computing them from first principles requires transport simulation that couples turbulence, heating, and sources.
The prediction problem
Given machine settings and actuator waveforms (heating power, gas puffing, current, shape), predict the resulting profiles, either the final steady state or the time evolution. ML models learn this mapping from experimental or simulation data, running orders of magnitude faster than the physics solver.
- Inputs: engineering parameters and actuator time series
- Outputs: profiles on a radial grid, possibly over time
- Uses: scenario planning, control design, real-time estimation
Model choices
Feedforward networks predict steady-state profiles from scalar inputs. Recurrent or state-space models predict evolution. Physics-informed variants embed conservation or transport structure to improve extrapolation and reduce data needs.
Uncertainty and limits
Profiles predicted for machine settings far from the training set are unreliable. A trustworthy profile predictor reports uncertainty that grows with distance from the data, so downstream control does not act on confident-looking but unsupported predictions.
Design use
For a design-stage device such as the Kronos breeder, profile predictors are trained on simulation output to explore operating scenarios cheaply. They complement, and are checked against, full transport runs. The predicted quantities are model results for a design, not measurements of hardware.