Real-Time State Estimation with ML
Inferring the current plasma state from partial, noisy diagnostics fast enough to feed control.
The problem
A controller needs to know the plasma state now: shape, position, current profile, key temperatures. Diagnostics measure only some of these, indirectly and with noise, and often too slowly for a fast control loop. State estimation reconstructs the full state from what is measured, within the loop's latency budget.
Classical versus learned
Classical estimators (Kalman filters, observers) combine a model of the dynamics with measurements, and are well understood but assume a known, often linear model. Learned estimators map measurements to state directly, capturing nonlinearity and sensor quirks from data, at the cost of needing training data and validation.
- Learned observers trained on simulation or reconstructed data
- Recurrent networks that carry state between time steps
- Hybrid filters where ML supplies the model or the measurement map
Latency and reliability
A real-time estimator must produce output within a fixed time and must degrade gracefully when a sensor drops out. Networks give constant-time inference, an advantage, but must be tested against sensor faults and out-of-distribution states so they do not confidently output nonsense.
Uncertainty
A state estimate without uncertainty is dangerous for control. Good estimators report confidence, so the controller can be cautious when the state is poorly known, for example after a sensor fault or during an unfamiliar transient.
Design context
For tokamak designs including the Kronos breeder, learned state estimators are developed and validated in simulation as part of control studies. They inform how a future controller might close the loop; they are not measurements of operating hardware.