Physics-Informed Machine Learning Approach to Modeling Line Emission from Helium-Containing Plasmas

TL;DR

This paper explores a hybrid machine learning approach combining collisional-radiative models and experimental data to improve the measurement of electron density and temperature in helium-containing plasmas, especially under data-limited conditions.

Contribution

It introduces a hybrid neural network model that integrates CRM and experimental data, demonstrating improved prediction accuracy for plasma parameters in fusion-relevant conditions.

Findings

Ensemble model modestly improved $T_e$ prediction accuracy.

CRM-based model outperformed others in data-limited scenarios.

Hybrid approach shows potential for plasma diagnostics with constrained data.

Abstract

The helium I line intensity ratio (LIR) method is used to measure the electron density ($n_e$) and temperature ($T_e$) of fusion-relevant plasmas. Although the collisional-radiative model (CRM) has been used to predict $n_e$ and $T_e$, recent studies have shown that machine learning approaches can provide better measurements if a sufficient dataset for training is available. This study investigates a hybrid neural network approach that combines CRM- and experiment-based models. Although the CRM-based model alone exhibited negative transfer in most cases, the ensemble model modestly improved the prediction accuracy of $T_e$. Notably, in data-limited scenarios, the CRM-based model outperformed the others for $T_e$ prediction, highlighting its potential for applications with constrained diagnostic access.