Bayesian Inference of Plasma State
Bayesian methods combine diagnostics and priors into a probability distribution over the plasma state, giving best estimates with calibrated uncertainty.
State estimation with uncertainty
The full plasma state, including temperature, density, and current profiles, is never measured directly. Bayesian inference treats the state as unknown, specifies a prior over plausible states, uses forward models to predict what each diagnostic would read, and updates to a posterior distribution given the actual measurements.
The machinery
Bayes' rule gives the posterior as proportional to the likelihood times the prior. The likelihood encodes each diagnostic's forward model and noise; the prior encodes physics constraints such as smoothness and positivity. The result is not a single profile but a distribution expressing what the data do and do not determine.
- Priors enforce smoothness, positivity, and known scalings
- Forward models link state to each diagnostic's raw signal
- Posterior yields best estimates and honest error bars
- Multiple diagnostics are fused consistently in one framework
Where ML enters
Full Bayesian inference can be slow because it requires many forward-model evaluations. Machine learning accelerates it in two ways: fast surrogates replace expensive forward models, and amortized inference networks learn to output the posterior directly from measurements, turning a costly optimization into a forward pass.
Value
Calibrated uncertainty is the point. Downstream control and analysis need to know not just the best estimate but how trustworthy it is, especially near operational limits where errors are costly. Integrated data-analysis frameworks built this way have become a standard for principled plasma state estimation.
For machines under design, such as the Hyperion breeder concept, Bayesian pipelines are constructed and tested on simulated diagnostics before construction. They operate on modeled data until hardware exists, and their posteriors quantify computational uncertainty, not measured performance of an unbuilt device.