Time Series Extrinsic Regression of Ion Cyclotron Emission Spectra Trained on Particle-In-Cell Simulations

TL;DR

This paper demonstrates that Time-Series Extrinsic Regression models trained on synthetic ion cyclotron emission spectra can accurately and rapidly infer plasma parameters from experimental spectra, aiding fusion diagnostics.

Contribution

It introduces a novel application of TSER models trained on simulated spectra to solve the inverse problem of plasma parameter estimation from ICE data.

Findings

TSER models accurately recover plasma parameters from synthetic spectra.

Models operate with near real-time speed.

Synthetic training enables effective inference on experimental data.

Abstract

Ion Cyclotron Emission (ICE) is a ubiquitous magnetised plasma phenomenon previously detected on virtually all large magnetic fusion devices and whose diagnostic potential for future power plants rests upon an accurate mapping of plasma parameters to spectra. This work demonstrates that the inverse problem is solved by training Time-Series Extrinsic Regression (TSER) models on synthetic ICE spectra from oblique propagation angle sweeps of nonlinear fully kinetic 1D3V particle-in-cell simulations of the magnetoacoustic cyclotron instability. Using datasets from a systematically constructed scan over reactor-relevant ranges of background magnetic field strength, density, and alpha-particle velocity pitch ($v_\parallel/|v|$) and concentration, we show that these bulk and fast ion parameters may be recovered from a JET ICE spectrum via TSER models with near real-time capability.