Description
Magnetization has proven to be an effective strategy for enhancing α-particle confinement, reducing thermal conduction losses, and achieving higher fusion yields compared with conventional inertial confinement fusion (ICF) implosions. Amplification of an externally applied ~10 T magnetic field up to the kT level during the implosion leads to the development of strong spatial gradients in density and temperature within the compressed core near stagnation. Resolving this spatial structure is essential for a deeper understanding of the implosion dynamics and the associated extended-MHD phenomena. In this work, we present a multi-zone spectroscopic model based on artificial neural networks (ANNs) to extract the spatial distribution of core plasma conditions in Ar-doped cylindrical magnetized implosions performed at OMEGA. The model was trained using an extensive synthetic database of Ar K-shell spectra covering electron temperatures from 0.4 to 7.0 keV and mass densities from 0.1 to 5 g/cm³, providing an initial estimate of the plasma conditions. A subsequent constrained exhaustive search combined with Bayesian inference is then employed to refine the extracted parameters and quantify their associated uncertainties. The synthetic spectra generated from the inferred spatial gradients are consistent with testing datasets after a relatively short training period. Moreover, the inferred spatial profiles and corresponding emission spectra show good agreement with those previously obtained using the more computationally expensive multi-zone random-search technique, revealing, for example, a 40% increase in electron temperature and a 30% decrease in mass density in the magnetized scenario. The multi-zone analysis indicates a relatively low contribution to the total Ar K-shell emission from the inner core regions, suggesting the potential use of a second dopant, such as Kr, which is more sensitive to higher temperatures, to unambiguously resolve core conditions through an independent spectroscopic diagnosis. Finally, we discuss the potential of this ANN-based framework for the analysis of data from recent and upcoming campaigns at OMEGA, NIF, and LMJ.
Work supported by NNSA/NLUF Grant DE-NA0003940, DOE Office of Science Grant No. DE-SC0022250, Grant No. PID2022-137632OB-I00 (Spanish Ministry of Science, Innovation and Universities), ANR HeapHop Project No. ANR-22-CE30-0044 (France) and EUROfusion Consortium under Grant Agreement No. 101052200.