News and Events

U.S. Burning Plasma Organization eNews

Nov 30, 2018 (Issue 135)


USBPO Mission Statement: Advance the scientific understanding of burning plasmas and ensure the greatest benefit from a burning plasma experiment by coordinating relevant U.S. fusion research with broad community participation.

CONTENTS

Announcements  
Director’s Corner
C.M. Greenfield
Research Highlight
M.D. Boyer
Contact and Contribution Information  

Announcements

10th ITER International School

The next ITER International School will be held January 21-25, 2019 in Daejeon, Korea with a focus on “The Physics and Technology of Power Flux Handling in Tokamaks” http://www.iterschool2019.kr. Students and early career researchers are encouraged to attend.

Director’s Corner By C.M. Greenfield

Events during the APS Division of Plasma Physics Conference in Portland, OR

Many of you attended the Research in Support of ITER contributed oral session and the Town Meeting on the ITER Research Plan. Tim Luce’s talk at the latter has been posted on the USBPO web page at https://burningplasma.org. We will also post all of the slides from the contributed oral session – just waiting for clearance from some of the speakers’ institutes.

The 10th ITER International School will be held in Daejon, Korea in January

The ITER International School aims to prepare young scientists/engineers for work in the field of nuclear fusion and in research applications associated with the ITER Project. The tenth edition will be held from January 21-25, 2019 in Daejeon, Korea. This year’s school will focus on a very important topic, “The Physics and Technology of Power Flux Handling in Tokamaks.” More information can be found at http://www.iterschool2019.kr.

Although the deadline has now passed for application for a USBPO scholarship, I urge students and early-career researchers, even without a scholarship, to consider attending if at all possible.

Progress at ITER

I visited ITER for the ITER Council meeting this month, and even in the five short months since my last visit, there has been major and highly visible progress. The ITER Organization reports that the project has now exceeded 60% completion (toward first plasma).

(Left) The tokamak complex continues to take shape. The tokamak bioshield is now hidden behind the walls of the tokamak building. (Right) The tokamak pit is being prepared for the first tokamak components.

Soon after my visit, on November 26, the first machine component - a cryostat feedthrough for poloidal field coil #4 - entered the tokamak pit. This is no small feat; the component is 10 meters long and weighs in at 6.6 tons. This auspicious occasion marks the beginning of five years of tokamak assembly activities.

https://www.iter.org/doc/www/content/com/Lists/Stories/Attachments/3170/feeder_installation_3.jpg

On the night of November 26, the first machine component was delicately lowered 30 meters down onto the tokamak pit floor, marking the beginning of five years of intense assembly activities. ãITER Organization.

Research Highlight

Operations & Control (Leaders: Eugenio Schuster & Dan Boyer)

Neural networks for real-time modeling of neutral beam injection on NSTX-U

M.D. Boyer, S.M. Kaye, K.G. Erickson

Princeton Plasma Physics Laboratory

Author e-mail: mboyer@pppl.gov

For tokamak reactors to become a viable approach to energy production, they must be capable of operating at a high fusion gain, ideally reaching steady-state operation, while avoiding conditions that could lead to instabilities and potentially damaging disruptions. Consistently achieving and maintaining high performance conditions while avoiding disruptions requires careful tailoring of the plasma equilibrium, including the spatial distribution of toroidal plasma current, temperature, density, and plasma rotation. A great deal of recent work has focused on developing and demonstrating techniques to actively control the evolution of profiles [1-9], as well as to detect and predict conditions that could lead to disruptions [10, 11].

Most recent profile control approaches make use of model-based control techniques, incorporating data-driven (e.g. [3]) or first-principles-driven models (e.g. [9]) of the plasma response into the design and tuning of the algorithm. This enables systematic handling of the high dimensionality of the problem, nonlinear coupling between magnetic and kinetic profiles and enables transferring approaches from device to device with minimal empirical tuning. Controllers developed using models driven by a first-principles description of the underlying physics can enable improved system performance and allow for operation over a wider range of conditions than schemes based on linear data-driven models. In practice, many first-principles models are not amenable to use in control-design or in real-time applications, either due to the complexity of the model or the computational time required, and it is necessary to sacrifice accuracy (and potentially control performance) by using simplified models.

For example, the NUBEAM code [12] used in TRANSP [13] for calculating the effect of neutral beams on the plasma (heating, current drive, torque, etc.) is a Monte Carlo code that can take several seconds to minutes per calculation, even when parallelized over many cores. While recent work has enabled TRANSP and NUBEAM to be used as a high-fidelity simulator of control algorithm performance [14, 15], the direct use of these models in control designs or real-time implementations is not tractable. Recent profile control approaches (e.g. [9]) have instead modeled the spatially-dependent beam effects as fixed spatial profile shapes with empirical scaling expressions for their magnitudes. While these approaches have been experimentally successful, the use of higher fidelity models could enable improved performance over a wider operating range.

Neural networks have recently been used to develop approximate models for computationally demanding codes in magnetic fusion, including TGLF and EPED [16], and provide a path to generating real-time capable high-fidelity predictive models. In order to take advantage of the high-fidelity results of NUBEAM calculations in real-time applications, NubeamNet, a neural network model for approximating NUBEAM, has been developed [17]. For the initial version of the model, the training and testing database was generated from the NUBEAM calculations output from interpretive TRANSP analysis of shots from the 2016 NSTX-U campaign [18] and augmented with scans of key inputs to NUBEAM, including Zeff, anomalous fast ion diffusivity, beam energy, and beam modulation patterns. A set of roughly 1,000 TRANSP runs based on approximately 250 NSTX-U shots was used to create a database of 100,000 NUBEAM calculations. Eighty percent of the shots in the dataset were randomly assigned to be used for model training, ten percent were assigned to validation, and the final ten percent were reserved for testing. No NUBEAM results from the discharges assigned to the testing dataset were used to train models, while validation data was used to assess accuracy and generalization during hyper parameter tuning. Inputs to the model were chosen to be: plasma boundary shaping parameters, beam powers, edge neutral density, Zeff, electron temperature and density profiles, q profile, and fast ion diffusivity. The outputs to be predicted by the model were chosen to be the neutron rate, shine through, charge-exchange and orbit loss, and profiles of beam heating to ions/electrons, beam current drive and torque, and fast ion pressure.

Example comparisons of the time history of NUBEAM calculations and NubeamNet estimations (with 3 layers of 125 nodes) for the total neutron rate, beam drive current at r=0.053, and fast ion pressure at r=0.053 are shown in Figure 1 for TRANSP run 204991S28. A comparison of predicted current drive, fast ion pressure, and ion heating profiles to those calculated by NUBEAM at t=0.787 are shown in Fig. 2. The results show that the neural network is able to closely approximate both the time behavior of quantities of interest as well as profile shapes.

Figure 1: Comparison of neural network model predictions to NUBEAM calculated values through run 204991S28. (left) neutron rate, (middle) beam driven current at r=0.053, and (right) fast ion pressure at r=0.053.

The neural network was implemented in the NSTX-U real-time computer and a scan of model topology was conducted to assess the scaling of calculation time with model complexity. Results, shown in Figure 3, show that models with complexity required to optimize the model fit (3 layers, ~125 nodes) can be run within the typical 200 microsecond cycle time of the NSTX-U control system. The neural network modeling approach enables a systematic way to optimize the trade-off between accuracy and execution time based on the requirements of specific applications. Recent advances in real-time PCI-based internode communication in the NSTX-U control system [19] will enable offloading calculations to a dedicated computer with enough cores to simultaneously calculate the models for uncertainty quantification as well as calculation of the sensitivity of outputs to changes in inputs needed by real-time control and optimization algorithms.

Figure 2: Comparison of neural network model predicted profiles to NUBEAM calculated profiles at t=0.787s in run 205018S56.

The accuracy and speed of the neural network model makes it well-suited to many real-time applications, including equilibrium reconstruction, dynamic observers and predictors, and actuator trajectory optimization. Initial application development is underway, including development of a nonlinear Kalman filter for real-time estimation of the current profile evolution, Zeff, and anomalous fast ion diffusivity from limited, noisy real-time measurements [20]. This application incorporates the neural network into a reduced model of the magnetic diffusion equation to predict the current drive and neutron rates, as well as their sensitivity to changes in the estimated parameters. Future model development will include creation of training sets and models based on predicted discharges to make the model useful for the planned operation space of future NSTX-U campaigns. Model development for other devices, including DIII-D and KSTAR is also underway.

https://lh4.googleusercontent.com/XWZlkSNqs9g5dO56jRaoLQVkrJ9ah2b4K3qNunYpoGOI68-Mj9h4lxsUYxo8SI3o3onlCu4NYK8qXwVN6t7vM52nXVZ095YntSvJ5_3Z6ABVtmiJQX1tU9ntM-VWdwTKobOX4COs

Figure 3: Comparison of model execution time as a function of number of neural network layers and nodes per layer.

A neural network model for evaluating the beam heating, current drive, torque, and other effects of the NSTX-U neutral beam system on the plasma has been developed. The model was trained on NUBEAM results calculated for the discharges in the first NSTX-U campaign. The speed and accuracy of the neural network model makes it well-suited for many real-time applications on NSTX-U, including dynamic observers for profiles [20], equilibrium reconstruction, and profile control.

Acknowledgements

This work was supported by the US Department of Energy Grant under contract number DE-AC02-09CH11466.

[1] Barton, J., et al., Nucler Fusion 52, 12, 123018 (2012).

[2] Boyer, M.D., et al., Plasma Physics and Controlled Fusion 55, 10, 105007 (2013).

[3] Moreau, D., et al., Nuclear Fusion, 53, 6, 063020 (2013).

[4] Boyer, M.D., et al., IEEE Transactions on Control Systems Technology 22, 5, 1725-1739 (2014).

[5] Maljaars, E., et al., Nuclear Fusion 55, 023001 (2015).

[6] Ilhan, Z., et al., “Model Predictive Control with Integral Action for the Rotational Transform Profile Tracking in NSTX-U”, Proceedings of the 2016 IEEE Multi-Conference on Systems and Control, Buenos Aires, Argentina (2016).

[7] Wehner, W., et al., “Combined Rotation Profile and Plasma Stored Energy Control for the DIII-D Tokamak via MPC”, Proceedings of the 2017 IEEE American Control Conference, Seattle, WA, USA (2017).

[8] Goumiri, I., et al., Physics of Plasmas 24, 5, 056101 (2017).

[9] Schuster, E., et al., Nuclear Fusion 57, 11, 116026 (2017).

[10] Rea, C., et al., Plasma Physics and Controlled Fusion 60, 084004 (2018).

[11] Berkery, J., et al., Physics of Plasmas 24, 5, 056103 (2017).

[12] Pankin, A., et al., Computer Physics Communications, 159, (2004).

[13] Poli, F., et al., TRANSP. Computer Software. 27 Jun. 2018. Web. doi:10.11578/dc.20180627.4.

[14] Boyer, M.D., et al., Nuclear Fusion 55, 053033 (2015).

[15] Boyer, M.D., et al., Nuclear Fusion 57, 066017 (2017).

[16] Meneghini, O., et al., Nuclear Fusion 57, 086034 (2017).

[17] Boyer, M.D., et al., “Real-time capable neural network approximation of NUBEAM for use in the NSTX-U control system”, Proceedings of the 2018 EPS Conference on Plasma Physics, Prague, Czech Republic (2018).

[18] Menard, J., et al., Nuclear Fusion 57, 102006 (2017).

[19] Erickson, K., et al., Fusion Engineering and Design, 133, 104, (2018).

[20] Boyer, M.D., et al., submitted to Nuclear Fusion (2018).


Calendar of Burning Plasma Events

2018

December 4-5

Fusion Power Associates 39th Annual Meeting & Symposium, Fusion Energy: Strategies & Expectations through the 2020s

Washington, DC

December 6-7

Meeting of the Fusion Energy Sciences Advisory Committee (FESAC)

Washington, DC

2019

January 15-17

ITPA Coordinating Committee & CTP ExComm

ITER HQ, France

March 18-21

US-EU Transport Task Force (TTF) meeting

Austin, TX

April 8-12

ITPA Diagnostics Topical Group meeting

Canberra, Australia

April 15-17

Sherwood Theory conference

Princeton, NJ

June 2-6

28th IEEE/NPSS Symposium on Fusion Engineering (SOFE)

Jacksonville, FL

July 8-12

46th European Physical Society Conference on Plasma Physics (EPS)

Milan, Italy

August 19-21

17th International Workshop on Plasma Edge Theory in Fusion Devices

UCSD, CA

Sept 20 – Oct 3

6th International Symposium on Liquid Metals Applications for Fusion (ISLA-6)

University of Illinois at Urbana-Champagne, IL

October 21-25

61st Annual Meeting of the APS Division of Plasma Physics

Fort Lauderdale, Florida

2020

JET DT-campaign (https://www.euro-fusion.org/newsletter/jet-full-throttle/)

JT60-SA First Plasma (http://jt60sa.org/)

Contact and Contribution Information

This newsletter provides a monthly update on U.S. Burning Plasma Organization activities. The USBPO operates under the auspices of the U.S. Department of Energy, Fusion Energy Sciences (FES) division. All comments, including suggestions for content, may be sent to the Editor. Correspondence may also be submitted through the USBPO Website Feedback Form.

Become a member of the U.S. Burning Plasma Organization by signing up for a topical group.

Editor: Walter Guttenfelder (wgutten@pppl.gov)

 

Privacy Policy
Who's in the USBPO News and Events References and Links Media Forums Contact Us