Teobaldo Luda, a graduate student from Politecnico di Torino in Italy, has been working with the GA Theory group to develop accelerated tokamak transport simulations via neural network (NN) based regressions of turbulent energy, particle and momentum fluxes predicted by the TGLF transport model. In this project a multilayer NN is trained and tested on a large database of TGLF runs, that use input parameters based on the experimental conditions of existing DIII-D shots. The NN has been integrated into TGLF. The accuracy of the NN is known to become worse outside of the training domain, so to solve this range-of-validity problem a new remote data acquisition technology has been developed. When the TGLF-NN model is run, the accuracy of the NN prediction is evaluated. If the accuracy condition is verified the result will be quickly computed with the NN, otherwise the full TGLF calculation is performed and the simulation data are stored in the remote database. The NN is constantly trained with the new TGLF results, allowing the NN to be used on an ever-growing range of parameters. Compared with TGLF results, it is found that the NN can accurately predict particle, energy, and momentum fluxes for both electrons and ions, while providing a computational speedup of over 5 orders of magnitude compared to the original calculation, which makes it ideal for scenario development simulations and real-time plasma control.
NIMROD modeling of massive gas injection (MGI) in DIII-D with pre-seeded 2/1 islands show that for identical island amplitude, the phase of the 2/1 island relative to the gas jet can significantly affect the spatial and temporal localization of the peak radiated power, producing nearly a factor of two variation in the maximum local radiation. This is consistent with previous simulations showing the importance of the n=1 mode phase, as well as experimental results showing that the n=1 phase can be decoupled from the injector phase with pre-existing n=1 perturbations. The simulations also show that after the radiated power peak, but before the thermal quench is completed, magnetic energy is dissipated at a rate comparable to thermal energy. This may in part explain observations of >100% radiated energy fraction—defined as the radiated energy over the pre-disruption thermal energy—seen in some DIII-D experiments. MGI when locked modes are present is part of the 2016 Joint Research Target on disruptions and has been investigated in recent DIII-D and C-Mod experiments. Comparison of these experiments with the modeling results is ongoing.
Joint GA Theory – PPPL Highlight: Orso Meneghini traveled to Princeton Plasma Physics Laboratory the first week of February to present new developments of the OMFIT framework and its modules, and provide support to PPPL scientists that are using the framework. As part of this collaboration, the physics modules in OMFIT were extended to enable time-dependent kinetic equilibrium reconstructions of the NSTX tokamak. In this workflow, in addition to the magnetic and MSE measurements, the EFIT equilibrium reconstruction is constrained with the plasma pressure (including the fast-ion contribution) and the plasma current which are inferred based on time-dependent TRANSP simulations. M. Podesta, H. Yuh, A. Diallo, F. Poli provided the support needed to fetch the experimental data that is used in OMFIT to automatically setup the TRANSP simulation. Discussion with S. Kaye, A. Diallo and G. Canal spurred the development of a new procedure to account for the iso-thermal flux-surfaces assumption within the OMFIT kinetic equilibrium reconstruction. OMFIT and the new capabilities were tested by A. Diallo, S. Sabbagh, J.Berkery, W. Wehner, and G. Canal, which provided valuable user's feedback. Work was also done in support of O. Izacard development of a UEDGE module in OMFIT. The visit was concluded with a presentation and a live demo showcasing an NSTX kinetic equilibrium reconstruction within OMFIT.
Nonlinear studies in GYRO have largely verified the critical gradient model (CGM) of energetic particle (EP) transport, with some subtle corrections that appear to improve agreement with experiment. The CGM places the critical EP density gradient, above which unstable Alfven eigenmodes (AEs) drive EP transport to reduce the gradient to the critical level, where the AE growth rate equals the low microturbulent growth rate. Nonlinear simulations also identify this condition as the point of transport runaway across multiple cases, thus verifying the model. In a recently published validation study against a DIII-D discharge, the growth-rate intersection point was determined by isolating AE and microturbulent drives in separate linear simulations. However, such a scheme can get the true AE critical gradient (runaway transport onset) point wrong. Linear simulations must apparently keep both AE and mictoturbulent drives on simultaneously. Under this prescription, the flattening in the core is stronger which brings validation results more in line with experiment. The reported nonlinear simulations also show that equilibrium flow shear pushes the critical gradient to a higher value, well characterized by a simple modification to the linear prescription above up to at least five times the experimental flow shear level. This flow shear modification to the critical gradient is insignificant at the level of flow shear in the experiment.
These highlights are reports of research work in progress and are accordingly subject to change or modification