ML Tomography for Bolometry
Learned tomographic inversion reconstructs 2D radiated-power maps from sparse bolometer lines of sight, aiding impurity and radiation control.
Radiated-power imaging
Bolometers measure the total radiated power along lines of sight through the plasma. Reconstructing the 2D emissivity map from these line integrals is a tomography problem, made hard by the small number of views and the noise in each measurement.
Why learning helps
Classical tomographic inversion needs strong regularization to stay stable with few views, and the choice of regularizer biases the result. A neural network trained on realistic emissivity patterns learns a data-driven prior, producing reconstructions that respect the structures actually seen in plasmas, such as edge radiation belts and impurity-driven hot spots.
- Inputs: line-integrated bolometer signals and viewing geometry
- Outputs: 2D emissivity map on the poloidal cross-section
- Prior: learned from simulated or archived radiation distributions
Speed for control
Fast reconstruction matters because radiated-power distribution signals impurity accumulation and radiative collapse, both precursors to trouble. A learned inversion running each cycle lets a controller detect a growing radiating region and respond with impurity or fueling adjustments.
Limits
With few views the inversion remains fundamentally underdetermined, so the learned prior does real work and can bias reconstruction toward familiar patterns. Novel radiation structures may be smoothed away. Uncertainty maps and cross-checks with spectroscopy guard against over-confidence.
Radiation control is essential for protecting plasma-facing components in any device. In Kronos design studies for the breeder concept, tomography pipelines are validated on synthetic radiation fields before hardware exists, and their outputs are treated as estimates supporting the design rather than measured behavior of a built machine.