Predictive applications of the QuaLiKiz neural network within integrated modelling for JET scenarios

Recent neural network (NN) surrogates of the reduced gyrokinetic turbulent transport model, QuaLiKiz [Bourdelle PPCF 2016], have accelerated its evaluation from 10 s per radial point to 1 ms. The resulting model, named QLKNN, achieves this while also being accurate enough to match previous tokamak modelling simulations [Ho PoP 2021]. This enables the rapid iteration of transport solvers, such as JINTRAC [Romanelli PFR 2014], both for post-discharge analysis and predictive modelling.

This work applies JINTRAC+QLKNN on the ohmic ramp-up phase of a recent JET D hybrid scenario discharge, representing data not used to train the NN. This scenario has a hollow T_e profile, which is due to impurity accumulation and increases with isotope mass [Challis NF 2020]. Its effect on the q-profile can lead to the formation of locked MHD modes and disruptions [Pucella NF 2021]. The dynamic simulation evolves j, n_e, T_e, and T_i for 7.25 s, along with the equilibrium and impurities (Be, Ni, W). This reveals the impurity impact on the T_e hollowing, along with local Z_eff and Q_rad contributions. In addition, the q evolution was qualitatively validated by comparing it with the observed sawteeth onset.

The isotope effect on this validated simulation and density boundary condition sensitivities were also studied. This assisted the JET teams in their T campaign preparations, by affirming trends from previous H experiments and setting engineering targets to recreate the D q-profile in T experiments. This demonstrates the potential of using NN surrogates for fast routine analysis and discharge design, while stressing the inclusion of edge plasma considerations in these models.