ML-Assisted Thomson Scattering Inference
Neural models infer electron temperature and density from Thomson scattering spectra faster and more robustly than fitting each spectrum by hand.
The measurement
Thomson scattering measures electron temperature and density by firing a laser into the plasma and analyzing the Doppler-broadened spectrum of scattered light. Broader spectra mean hotter electrons; total scattered intensity relates to density. Converting raw spectra into profiles is an inference problem repeated for many spatial channels and time points.
Classical fitting versus learning
The standard method fits a scattering model to each spectrum by nonlinear least squares, which is accurate but slow and can fail on low-signal or noisy channels. A neural network trained on simulated spectra learns the inverse map from spectrum to temperature and density directly, returning estimates in a single pass.
- Fast enough for between-shot or near-real-time profile estimates
- Robust to noise when trained with realistic instrument models
- Provides a strong initial guess that a classical fit can refine
- Flags low-quality channels via predicted uncertainty
Training data
Because ground-truth temperature is not directly observable, training uses a forward instrument model that maps known plasma conditions to synthetic spectra, including detector response, background light, and noise. The network learns to invert this forward model, and its accuracy depends on how faithfully the forward model represents the real diagnostic.
Uncertainty and validation
Physically meaningful error bars are essential because profiles feed downstream analysis. Bayesian or ensemble methods provide calibrated uncertainty, and results are validated against classical fits on well-conditioned spectra. Systematic errors in the forward model are the main risk and are checked with calibration data.
Fast, reliable profile inference supports control and analysis for any tokamak program. In Kronos design work such inference pipelines are prepared in simulation for the breeder diagnostics suite ahead of construction; they operate on modeled data because the hardware is not yet built.