Speaker
Description
In future fusion power plants, to ensure safe and reliable operation, it will be necessary to be able to accurately and efficiently measure the distribution of energetic particles[1]. Since the fast-ion distribution cannot be measured directly, we need to reconstruct it from diagnostic measurements[2]. The problem of reconstructing the fast-ion distribution given measurements is an ill-conditioned inverse problem[3]. The set of diagnostic measurements used to reconstruct the fast-ion distribution is often sparse and noisy[4], leading to uncertain reconstructions that can be plagued by artifacts[5]. Improving both the quality and efficiency of such reconstructions is paramount to ensure a successful operation of future tokamaks such as ITER[6] and SPARC[7]. In this work, we discuss a possible solution in the form of variational auto encoders (VAE). VAEs are a type of deep neural network that have been used to solve problems such as de-noising[8], feature extraction[9] and image classification. We show that VAEs can be used to post-process tomographic reconstructions of the fast-ion distribution to achieve identification and removal of artifacts. We also show that VAEs can be used to reduce the dependence of tomographic reconstructions on the choice of the regularization parameter $\lambda$.
[1] M. Salewski, Fast-ion diagnostic in fusion plasmas by velocity-space tomography, Technical University of Denmark, 2019
[2] M. Salewski, et al, 2019, JINST 14 C05019
[3] M. Rud et al, 2026, Plasma Phys. Control. Fusion 68 015035
[4] B. Madsen et al, 2020, Plasma Phys. Control. Fusion 62 115019
[5] H. Järleblad et al, 2025, Nucl. Fusion 65 016060
[6] B. Bigot et al, 2019, Nucl. Fusion 59 112001
[7] A.J. Creely et al, 2020, Plasma Phys. Control. Fusion 86 865860502
[8] S. Ruipérez-Campillo et al, 2025, Expert Systems With Applications 300 130185
[9] H. Wang et al, 2025, Energy 330 136716