Enhancing Temporal Resolution in Fusion Diagnostics through Multimodal Neural Networks

Diag2Diag introduces a multimodal neural network model designed to enhance resolution by leveraging inter-diagnostic correlations within complex systems. As the use case, we demonstrate that the Thomson Scattering diagnostic at DIII-D can be reconstructed from others and its resolution can significantly be increased to 500kHz, while preserving intrinsic physics. The suite of input diagnostics includes Interferometer, ECE, Magnetic probes (Magnetics), Charge Exchange Recombination (CER), and MSE. The Super Resolution TS signals (SRTS) match the original TS signals whenever the measurements exist, while capturing events missed by the measured TS, enabling more precise analysis for new discoveries.

We investigate whether synthetic super-resolution TS can verify hypotheses in fusion plasma physics. Two phenomena are studied: ELM cycle analysis [1] and plasma response to external field perturbations [2]. For ELM cycles, the SRTS follows the measured TS when available and indicates ELM cycles often missed by the measured TS, even when it fired at “bunch mode” [3]. For external field perturbations, the SRTS reveals the experimental island effect induced by Resonant Magnetic Perturbation (RMP), providing the first diagnostic evidence of profile flattening at magnetic islands.

Given the successful proof of concept of this methodology and as future work, we will expand this study to spatial resolution enhancement as well as enhancing other diagnostics such as CER which are critical for monitoring the plasma profile.

[1] A.O. Nelson, et al, Nuclear Materials and Energy, 26, 100883 (2021)

[2] Q.M. Hu et al, Nucl. Fusion, 61, 106006 (2021)

[3] Z. HE et al, Plasma Sci. Technol., 21, 105603 (2019)