Description
The alkali beam emission spectroscopy (A-BES) system at W7-X is routinely used to reconstruct edge plasma electron density profiles from measured light emission in post-experiment analysis. While the diagnostic is capable to resolve high-frequency fluctuations in the plasma edge, the current data processing chain is not suitable for real-time use: obtaining a density profile from the measurement is too slow to be integrated into control schemes. The main computational bottleneck is the inversion from light profiles to density profiles, which is presently performed using a linearized Bayesian method [1]. Neural network surrogates offer an attractive alternative by shifting the computational cost to an offline training procedure, enabling sub-millisecond online inference even without optimization. We have previously shown that simple neural networks can reconstruct electron density profiles from synthetic A-BES signals [2]. In this work, we extend this approach to measured signals from OP2.2 and OP2.3 and demonstrate, as proof of concept, that such surrogates can meet real-time constraints. To encourage learning physically meaningful representations and to reduce sensitivity to measurement noise, we investigate advanced architectures, including operator learning methods [3]. Model performance is evaluated in terms of prediction accuracy and runtime on OP2.2 and OP2.3 test data previously unseen by the model.
References
[1] Vecsei, M. et al. Review of Scientific Instruments, 2021, 92 (11), 113501.
[2] Molnar, B. et al. Proceedings of the 51st EPS Conference on Plasma Physics, 2025
[3] Lu, L. et al. Nature Machine Intelligence, 2021, 3(3), pp. 218–229.