Sterling Smith and Orso Meneghini participated in the 2nd ITER code camp, which focused on development of physics workflows in IMAS (Integrated Modeling and Analysis Suite). Over the course of two weeks, two key workflows that are part of the OMFIT framework were adapted to work with IMAS: the execution of the GACODE suite of codes ©GYRO, TGLF, NEO, TGYRO; and the ability to generate experimental profile fits from experimental data within the OMFITprofiles module. Both adaptations relied on the OMAS (Ordered Multidimensional Array Structure) [gafusion.github.io/omas] library to generate the interfaces with IMAS. As a result of this effort, OMFITprofiles can be now be leveraged by any device that stores its data in IMAS format, as was demonstrated at the code-camp for the WEST tokamak. Furthermore, any device that was already supported by OMFITprofiles (DIII-D, C-Mod, NSTX, COMPASS) can now write both its core profile raw data and fits into IMAS. JET support in OMFITprofiles was added during the code-camp.

Dr. Zhirui Wang from PPPL visited GA to collaborate with Y. Liu and lead DIII-D experiments (03/20/2018 and 03/23/2018) to study the Neoclassical Toroidal Viscosity (NTV) contributed by NBI beam driven energetic particles. The effect of beam ions on the NTV torque was previously predicted theoretically by drift kinetic computations using the MARS-K code. In the experiments, a beam voltage scan with matched beam power and torque was performed to vary the precession frequency of fast ions shot by shot. In each discharge, resonant and non-resonant magnetic perturbations were applied to study the coil phasing dependence of the NTV torque. The fraction of trapped fast ions was also varied by applying different beam combinations. The existence of an energetic particle driven NTV, if verified and validated in further analysis, has important implication for ITER given the presence of a large fraction of both fusion alphas and beam driven energetic particles.

Emeritus Professor Chuan Liu of the University of Maryland, was welcomed back to GA this week for a brief visit to discuss a fusion book he is writing. Professor Liu was formerly the head of the GA Theory Group from 1981 to 1985. He was also the President of the Taiwan National Central University from 2003 to 2006 and more recently served as the Founding Master of a residential college at the University of Macau from 2014 to 2017.

Klaus Hallatschek from the Max-Planck-Institute for Plasma Physics in Garching visited GA from Feb 12th to March 8th to work with the theory group on multiple topics. He developed a new self-adjoint algorithm to evaluate the gyrokinetic collision operator, and this new formulation was implemented in CGYRO. Using CGYRO, Klaus was subsequently able to successfully recover the results from 2-fluid resistive ballooning turbulence simulations (with the NLET code). Differences between the 2-fluid and gyrokinetic results were only 3% and 8% for the particle and energy diffusivities, respectively. He also worked with members of the theory group on new fluid moment hierarchies for the Fokker-Planck equation, in support of the development of a new high performance computing framework at GA.

Past implementation of neural-network accelerated models for core turbulence (TGLF-NN) and pedestal (EPED1-NN) have been done using the FANN (Fast Artificial Neural Network) C library. However, the FANN library has not kept up with recent developments in machine learning where the introduction of new minimization algorithms and activation functions promise the ability to train better models, faster. This aspect is particularly important as the size of our training datasets is rapidly increasing. Furthermore, second generation machine learning libraries are natively capable of taking full advantage of modern computers increased parallelism, Graphical Processing Units (GPUs), or Tensor Processing Units (TPUs). A new implementation of TGLF-NN and EPED1-NN with state-of-the-art Google TENSORFLOW library is indeed very encouraging, well exceeding the accuracy and training speed of the existing neural-network implementations done with FANN. The speedup provided by GPU-enabled computing nodes on the GA clusters can be significant, depending on the complexity of the network topology and the size of the training dataset. These recent developments provide a solid foundation for development of future machine-learning based model.