Nuclear fusion using magnetic confinement, in particular in the tokamak configuration, is a promising path towards sustainable energy. A core challenge is to shape and maintain a high-temperature plasma within the tokamak vessel. This requires high-dimensional, high-frequency, closed-loop control using magnetic actuator coils, further complicated by the diverse requirements across a wide range of plasma configurations. In this work, we introduce a previously undescribed architecture for tokamak magnetic controller design that autonomously learns to command the full set of control coils. This architecture meets control objectives specified at a high level, at the same time satisfying physical and operational constraints. This approach has unprecedented flexibility and generality in problem specification and yields a notable reduction in design effort to produce new plasma configurations. We successfully produce and control a diverse set of plasma configurations on the Tokamak à Configuration Variable1,2, including elongated, conventional shapes, as well as advanced configurations, such as negative triangularity and ‘snowflake’ configurations. Our approach achieves accurate tracking of the location, current and shape for these configurations. We also demonstrate sustained ‘droplets’ on TCV, in which two separate plasmas are maintained simultaneously within the vessel. This represents a notable advance for tokamak feedback control, showing the potential of reinforcement learning to accelerate research in the fusion domain, and is one of the most challenging real-world systems to which reinforcement learning has been applied.

The role of the pedestal position on the pedestal performance has been investigated in AUG, JET-ILW and TCV. When the pedestal is peeling-ballooning (PB) limited, the three machines show a similar behaviour. The outward shift of the pedestal density leads to the outward shift of the pedestal pressure which, in turns, reduces the PB stability, degrades the pedestal confinement and reduces the pedestal width. Once the experimental density position is considered, the EPED model is able to correctly predict the pedestal height. An estimate of the impact of the density position on a ITER baseline scenario shows that the maximum reduction in the pedestal height is 10% while the reduction in the fusion power is between 10% and 40% depending on the assumptions for the core transport model used.When the pedestal is not PB limited, a different behaviour is observed. The outward shift of the density is still empirically correlated with the pedestal degradation but no change in the pressure position is observed and the PB model is not able to correctly predict the pedestal height. On the other hand, the outward shift of the density leads to a significant increase of η e (where η e is the ratio of density to temperature scale lengths, η e = L ne /L Te ) which leads to the increase of the growth rate of microinstabilities (mainly ETG and ITG) by 50%. This suggests that, when the pedestal is not PB limited, the increase in the turbulent transport due to the outward shift of the density might play an important role in the decrease of the pedestal performance.

We present an ultrafast neural network (NN) model, QLKNN, which predicts core tokamak transport heat and particle fluxes. QLKNN is a surrogate model based on a database of 300 million flux calculations of the quasilinear gyrokinetic transport model QuaLiKiz. The database covers a wide range of realistic tokamak core parameters. Physical features such as the existence of a critical gradient for the onset of turbulent transport were integrated into the neural network training methodology. We have coupled QLKNN to the tokamak modelling framework JINTRAC and rapid control-oriented tokamak transport solver RAPTOR. The coupled frameworks are demonstrated and validated through application to three JET shots covering a representative spread of H-mode operating space, predicting turbulent transport of energy and particles in the plasma core. JINTRAC-QLKNN and RAPTOR-QLKNN are able to accurately reproduce JINTRAC-QuaLiKiz T i,e and n e profiles, but 3 to 5 orders of magnitude faster. Simulations which take hours are reduced down to only a few tens of seconds. The discrepancy in the final source-driven predicted profiles between QLKNN and QuaLiKiz is on the order 1%-15%. Also the dynamic behaviour was well captured by QLKNN, with differences of only 4%-10% compared to JINTRAC-QuaLiKiz observed at mid-radius, for a study of density buildup following the L-H transition. Deployment of neural network surrogate models in multi-physics integrated tokamak modelling is a promising route towards enabling accurate and fast tokamak scenario optimization, Uncertainty Quantification, and control applications.

Transport analyses using first-principle turbulence codes and 112-D transport codes usually study radial transport properties between the tokamak plasma magnetic axis and a normalized minor radius around 0.8. In this region, heat transport shows significantly stiff properties resulting in temperature scalelength values (R∕LT) that are relatively independent of the level of the radial heat flux. We have studied experimentally in the tokamak à configuration variable [F. Hofmann et al., Plasma Phys. Controlled Fusion 36, B277 (1994)] the radial electron transport properties of the edge region, close to the last closed flux surface, namely, between ρV=V/Vedge=0.8 to 1. It is shown that electron transport is not stiff in this region and high R∕LTe values (∼20–40) can be attained even for L-mode confinement. We can define a “pedestal” location, already in L-mode regimes, where the transport characteristics change from constant logarithmic gradient, inside ρV = 0.8, to constant gradient between 0.8 and 1.0. In particular, we demonstrate, with well resolved Te and ne profiles, that the confinement improvement with plasma current Ip, with or without auxiliary heating, is due to this non-stiff edge region. This new result is used to explain the significant confinement improvement observed with negative triangularity, which could not be explained by theory to date. Preliminary local gyrokinetic simulations are now consistent with an edge, less stiff, region that is more sensitive to triangularity than further inside. We also show that increasing the electron cyclotron heating power increases the edge temperature inverse scalelength, in contrast to the value in the main plasma region. The dependence of confinement on density in ohmic plasmas is also studied and brings new insight in the understanding of the transition between linear and saturated confinement regimes, as well as of the density limit and appearance of a 2/1 tearing mode. The results presented in this paper provide an important new perspective with regards to radial transport in tokamak plasmas which goes beyond L-mode plasmas and explains some previous puzzling results. It is proposed that understanding the transport properties in this edge non-stiff region will also help in understanding the improved and high confinement edge properties.

Reducing the uncertainty on physical input parameters derived from experimental measurements is essential towards improving the reliability of gyrokinetic turbulence simulations. This can be achieved by introducing physical constraints. Amongst them, the zero particle flux condition is considered here. A first attempt is also made to match as well the experimental ion/electron heat flux ratio. This procedure is applied to the analysis of a particular Tokamak à Configuration Variable discharge. A detailed reconstruction of the zero particle flux hyper-surface in the multi-dimensional physical parameter space at fixed time of the discharge is presented, including the effect of carbon as the main impurity. Both collisionless and collisional regimes are considered. Hyper-surface points within the experimental error bars are found. The analysis is done performing gyrokinetic simulations with the local version of the GENE code, computing the fluxes with a Quasi-Linear (QL) model and validating the QL results with non-linear simulations in a subset of cases.