Description
urate reconstruction of plasma parameters is essential for the interpretation of edge and divertor plasma physics and the development of real-time control strategies. Classical techniques, such as Bayesian inference, provide reliable inferences but are often computationally expensive, limiting their applicability for fast analysis and real-time applications.
In this work, we present a machine-learning inverse framework that employs deep neural networks to reconstruct plasma parameters directly from spectroscopic emissivities. A synthetic training dataset is generated using ADAS [1] and GOTO [2] collisional-radiative models, that includes the contribution from plasma-molecule interactions through the $H_2^+$ channel in ADAS [3]. The dataset links plasma parameters such as the electron temperature and density, neutral deuterium and helium densities, and molecular fraction contribution to line emissivities. To mimic experimental conditions, Gaussian noise is applied to the synthetic emissivities prior to training.
The neural network is trained to infer plasma parameters from the perturbed emissivities. To enforce physical consistency, custom differentiable layers are introduced to compute derived quantities, such as hydrogen ionisation and recombination rates, from the reconstructed parameters. The cost function is formulated by computing the emissivities from the reconstructed plasma parameters using the same collisional-radiative models and comparing them directly to the unperturbed input emissivities. This approach ensures an equal contribution of all reconstructed parameters to the optimisation process and embeds the underlying plasma physics into the training procedure.
The method is applied to a selected TCV plasma configuration with its performance assessed through comparison with classical Bayesian inference methods and independent Thomson scattering measurements. The results demonstrate that the deep-learning-based approach can reliably reconstruct plasma parameters while achieving inference times on the order of 1 ms. This is sufficient for both real-time applications, for plasma control, and rapid post-shot analysis for control room development, highlighting the potential of these kinds of physics-informed machine learning techniques for advanced plasma diagnostics.\\\
$[1]$ M. O'Mullane H. Summers. Atomic data and analysis structure. Technical Report JET-P9735, 2000.\
$[2]$ M. Goto et al., 2003, J. Quant. Spectrosc. Radiat. Transfer 76 331-44\
$[3]$D. Greenhouse et al., 2025, Plasma Phys. Control. Fusion 67 035006