SHRX: a neural surrogate of turbulent transport in tokamak plasma and its comparison with TORAX against experimental data
Year: 2026
Type: preprint
Abstract: Integrated tokamak modelling needs fast surrogates of turbulent transport, but existing ones depend on JAX. We built SHRX, a neural network with 29,283 parameters that runs on NumPy alone, by distilling the QuaLiKiz neural network (QLKNN). A “bridge” converts the state of our one-and-a-half-dimensional simulator SINTEZ into the ten dimensionless QLKNN inputs. For stiff transport we introduce the error of the critical-gradient factor, \(\:{\delta\:}_{f}\), as a metric, and an ensemble of five networks gives an uncertainty estimate. We compare SINTEZ with SHRX against the TORAX code at three levels: identical local plasma states, an integrated ITER discharge with identical inputs, and 206 conventional and 41 spherical records of the H-mode database of the International Tokamak Physics Activity (ITPA). On machines held out of training, the median error of the heat diffusivity \(\:\chi\:\) is 9% and the median \(\:{\delta\:}_{f}\) is 2.3%; the ensemble lowers the \(\:\chi\:\) error to 5.9–6.5%. The ITER fusion power is 492 MW with SINTEZ and SHRX, 478 MW with SINTEZ and QLKNN, and 532 MW with TORAX. On experimental records the two codes agree with each other (\(\:r=0.995\) in \(\:\text{l}\text{n}W\)) better than with the measurements, and both are less accurate than the IPB98(y,2) regression (\(\:{\sigma\:}_{\text{l}\text{n}}\) of 0.55 and 0.56 against 0.21). On spherical tokamaks SHRX overestimates the stored energy by a median factor of 28. SHRX is a faster, JAX-free copy of QLKNN, not a better model than TORAX: most of the experimental error comes from inputs that both codes share.
Source: Research Square
Author: Beibarys Ermekuly
Institution: Ahmet Yesevi University
Language: English
Subfield: Nuclear and High Energy Physics
Field: Physics and Astronomy
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7225985751
