ML-Assisted Interferometry Inference
Learned inversions turn line-integrated interferometer phase measurements into density profiles, handling fringe ambiguity and limited views.
What interferometry measures
A plasma interferometer measures the phase shift of a beam passing through the plasma, which is proportional to the line-integrated electron density along the beam path. Recovering a spatial density profile from a few such chords is an ill-posed inversion, since many profiles can produce the same line integrals.
The inversion challenge
With only a handful of viewing chords, classical inversion relies on assumptions about profile shape or magnetic flux surfaces. Machine learning provides a data-driven prior: a network trained on realistic density profiles learns which reconstructions are plausible, regularizing the inversion without hand-tuned assumptions.
- Inputs: line-integrated phase from each chord, plus magnetic-geometry context
- Outputs: a density profile or 2D density map
- Prior: learned from a library of physically realistic profiles
Fringe jumps
Interferometers can lose count of interference fringes during fast density changes or disruptions, producing discontinuous fringe jumps in the phase signal. Sequence models can detect and correct these jumps by recognizing that the underlying density evolves smoothly, improving signal reliability in real time.
Combining diagnostics
Interferometry is most powerful when fused with other diagnostics. A Bayesian framework can combine interferometer chords, Thomson scattering, and reflectometry into a single consistent density estimate, with machine learning accelerating the forward models that make such fusion tractable in real time.
Reliable density inference underpins fueling control and disruption avoidance for any device. In Kronos design planning these inversion tools are developed against simulated diagnostics for the breeder concept; they process modeled signals because the machine has not been built, and results are cross-checked against independent density measurements once hardware exists.