Title: Kinetic Simulation of Sheared Flows in Tokamaks
1Kinetic Simulation of Sheared Flows in Tokamaks
- J.A. Heikkinen2, S.J. Janhunen1, T.P. Kiviniemi1,
S.Leerink1, M. Nora1, and F. Ogando1,3, - 1 Teknillinen Korkeakoulu (TKK), Finland
- 2 Valtion Teknillinen Tutkimuskeskus (VTT),
Finland - 3 Universidad Nacional de Educación a Distancia
(UNED), Spain
2Outline
- Background
- Gyrokinetic full f approach for toroidal fusion
plasmas. - Examples of Vlasov codes.
- Description of a PIC full f code.
- Testing
- Comparison to neoclassical theory.
- Linear and nonlinear benchmarks of unstable
modes. - Influence of noise on results.
- Transport simulations in toroidal plasmas.
- Future challenges of full f
3Section IBackground
4Turbulence and zonal flows
- Self-organization and zonal flow generation with
turbulence is important not only in fusion
plasmas but also in other fluids like planetary
atmoshpheres and solar plasma - Transport transients and related transport
barriers are important, e.g., for ITER fusion
reactor design - Empirical evidence from a number of tokamak
facilities indicates a complex scaling law for
transport transients - The scaling laws for transport barriers in
tokamaks are today to a large extent not
understood by physics principles
5Sheared flows and turbulence appear in toroidal
plasma simulation
6Kinetic simulation required
- All trials up to date around the world to see the
spontaneous confinement transition in an
experimental-like first principles plasma
simulation have failed so far. - Reason to this may be that not all relevant
physics like kinetic details of particle motion
have been included in the simulation codes, lack
of self-consistency with background evolution, or
model simplifications like using fluid codes have
been too crude.
7Section IIGyrokinetic full f approach for
simulation of toroidal plasmas
8Full f simulation of tokamak plasmas
- In full f simulation, one solves for the whole
particle distribution function in phase space and
in time either non-perturbatively, with
gyrokinetics or drift-kinetically - Pioneering non-perturbative 5D particle-in-cell
(PIC) simulation of magnetically confined
toroidal plasmas (unrealistic me/mi) - Cheng Okuda, 78 Sydora, 86 LeBrun, 93,
Kishimoto, 94 - Eulerian and semi-Lagrangian Vlasov 3D, 4D, and
5D simulation - Cheng Knorr, 76 Manfredi, 96
Sonnendruecker, 99 Xu, 06 Scott, 05
Grandgirard, 06 - Gyrokinetic 5D particle simulation
- Furnish, 99 Heikkinen, 00, 03 Chang, 03.
-
9Calculation flux in particle codes
Calculation of forces from fields and velocity
Acceleration and increment of velocity
Computation of electric field. Magnetic is given.
Initial step with ?0
Displacements and new positions. Boundary
conditions.
Calculation of density. Current profile fixed.
Resolution of Poisson equation for the
electrostatic potential.
10Delta f vs. full f
- Delta f calculates perturbations from an assumed
background distribution f0. - Powerful for small f-f0
- Linear mode analysis
- Snapshot transport analysis
- Path-breaking global transport studies for large
toroidal installations
- Full f calculates the whole particle
distribution. - Fitting processes that perturbate strongly the
particle distribution - Strong transient or long time scale transport in
core or edge plasmas - Strong particle/energy sources
- Full neoclassical equilibrium
- Edge plasma (sheaths, wall losses, recycling,
separatrix, flows) - Large MHD (sawteeth, ELMs)
11Delta f vs. full f
- Automatic extension of Delta f to full f
calculation may be possible - Non-conservative effects collisions, sources
- Loading optimization
- Automatic increase of the number of markers when
required - Few particles (10-100) per cell are needed for
good results with small f-f0.
- Full f can be realized both in PIC and Vlasov
(Eulerian, semi-Lagrangian) - Vlasov simulation is free of noise and has
controlled numerical resolution in phase space. - PIC is numerically flexible and straightforward
to implement - Needs many particles per cell (1000) or fine
grid discretization of velocity space for an
acceptable noise level or accuracy.
12Section IIIExamples of full f Vlasov codes
13Full f 5D gyrokinetic code GYSELA Grandgirard,
06 (CEA, LPMIA-Univ, IRMA-Univ, LSIIT-Illkirch)
- Semi-Lagrangian GYSELA code
- - fixed grid, follows trajectories backwards,
- - global code
- - 5D GAM and ITG Cyclone tests successful
14Full f 5D gyrokinetic code TEMPEST Xu 06 (LLNL,
Calif-Univ, GA, LBNL, PPPL)
- - Based on modified gyrokinetics valid for large
long-wavelength and small short-wavelength
variations (Qin, 06) - - Fixed grid, equations solved via a
Method-of-Lines approach and an implicit
backward-differencing scheme using iteration
advances the system in time - - Developed for circular core or divertor edge
geometry - - 4D neoclassical tests successful. 5D drift wave
and ITG benchmarking ongoing -
-
15Section IVThe ELMFIRE code an example of a full
f PIC code
16The ELMFIRE group
Founded in 2000 at Euratom-Tekes Contributors
from Finland Spain Holland Main
affiliations VTT TKK ... but also ... CSC Ã…bo
Akademi UNED (Spain)
17ELMFIRE code
- Full f nonlinear gyrokinetic particle-in-cell
approach for global plasma simulation (present
version electrostatic). - Magnetic coordinates (?,?,?) Boozer 81.
- Guiding-center Hamiltonian White Chance, 84.
- Gyrokinetics is based on Krylov-Boholiubov
averaging method in description of FLR effects
(P. Sosenko, 01). - Adiabatic or kinetic electrons with impurities.
- Parallelized using MPI with very good
scalability. - Based on free software PETSc and GSL for math
calc.
18ELMFIRE full f features
- An initial canonical distribution function avoids
the onset of unphysical large scale ExB flows
(Heikkinen, 01) - Direct implicit ion polarization (DIP) and
electron acceleration (DEP) sampling of
coefficients in the gyrokinetic equation - Quasi-ballooning coordinates to solve the
gyrokinetic Poisson equation
- Versatile heat (RF, NBI, Ohmic) sources and
particle sources/ recycling - Full binary collision operator
19Poisson equation
- W.W. Lee proposed standard model with
polarization drift included in equation operator. - Ion density evaluated from ion motion without
polarization drift - P. Sosenko proposes including polarization in the
ion density. - Ion density evaluated from ion motion with
polarization drift. Circular gyro-orbits.
20Implementation into ELMFIRE
- Solve ? by isolating ion polarization drift
contribution to density. - That contribution is calculated implicitely every
timestep using also previous values of ?. - The gyroaveraged electric field is interpolated
from grid potential values for the ion
polarization drift.
Larmor circle
ith subparticle cloud
y
k
i
?py
?px0
xp,yp
?px-
?px
?py0
ds
k
?py-
x
pth point on the Larmor orbit of the kth ion
gyro-center of the kth ion
21Section VBenchmarking of ELMFIRE
22Testing ELMFIRE
- Comparison to neoclassical theory in the presence
of turbulence. - Frequency of GAM and Rosenbluth residual.
- Neoclassical radial electric field.
- Comparison to other codes has been done in the
Cyclone Base cases. - Linear growth of unstable modes and their phase.
- Nonlinear saturation of transport in both
adiabatic and kinetic-electron case. - Comparison to experimental results.
- Collaboration with IOFFE Institute and St.
Petersburg Polytechnic working with the FT-2
tokamak.
23Available resources
- Gyrokinetic full f computation is very demanding
computationally. Parallel computation is a need. - CSC (The Finnish IT Center for Science) provides
shared use of high-end parallel computers.
- IBM eServer cluster 1600. 512 processors with
2.2TFlops, 384GB RAM and High Performance Switch
communication. - Cluster of 768 AMD OpteronTM processors up to
3.2TFlops, 1600GB RAM, Infiniband network. - Cray XT4 (Hood) 70TFlop, 70TB RAM HP ProLiant
Supercluster, 10.6 Tflop, 100 TB
24Geodesic Acoustic Modes
- Neoclassical theory predicts GAM frequency and
Rosenbluth residual. - Results show good wide agreement with theory.
- Simulations done on a plasma annulus.
- R0.3-0.9 m, a0.08 m, B0.6-2.45 T, q1.28-2.91,
Ti90-360 eV, ni5.11019m-3 (r/a0.75)
25Neoclassical radial electric field
- Neoclassical radial electric field is well
followed in conventional (L-mode) turbulent
simulations both in radius and in time
- R1.1 m, a0.08 m, B2.1 T, I22 kA, parabolic
ion heated n,Ti,e profiles (r0.04m).
26Neoclassical benchmarking
- Parameters and boundary conditions
- FT-2 like parameters, R00.55 m and a0.08 m,
- Itot 55 kA, T,n,j (1-(r/a)2)? ,
- n051019 1/m3
- T0300 eV in high T case, T0120 eV in low T
- no heating, no loop voltage
- relaxing profiles, cooling by recycling on
outer radius
27Neoclassical benchmarking
Number of particles must be sufficient as Er may
depend on noise Er radial dependence fairly well
predicted by standard neoclassical theory.
However, Reynolds stress and poloidal Mach number
can be important.
28Linear growth of unstable modes
- Test based on adiabatic Cyclone Base Case (Dimits
PoP '00) - Red points from ELMFIRE, blue line fit from
article. - Figures show growth rates and typical time
evolution for a mode with k??i0.3
29Linear growth of unstable modes
- Test based on adiabatic Cyclone Base Case (Dimits
PoP '00) - Red points from ELMFIRE, blue line fit from
article. - Figures show growth rates and typical time
evolution for a mode with k ?? i0.3
Region of linear growth
30Linear growth of unstable modes
- Test based on kinetic Cyclone Base Case (Chen NF
'03) - Filled circles and squares from ELMFIRE
- Dashed line fit for the growth rate ? from Chen
NF 03.
31Evolution of thermal conductivity
- Evolution of ?i is studied with nonlinear runs of
Cyclone Base. - Measured at ra/2 (q1.4). Using kinetic
electrons. R/LT10. Weak collisionality
T(a/2)2000 eV, n51017 m-3. - Convergence requires a large number of particles
per cell.
32Section VIInfluence of noise on results
33Influence of noise on results
- Particle simulation not only covers physical
density fluctuations, but it produces undesirable
noise with fluxes that perturbate the solution. - Associated diffusivity can be estimated from the
radial particle shift during decorrelation time. - Physical radial ion heat conductivity can be
estimated from mixing-length estimate of the
physical level of fluctuations, being also
proportional to T3/2.
34Effects on calculated conductivity
- Image shows influence of strong collisionality
and potential averaging on ion radial heat
conductivity. - Collisionless cases show residual noise
conductivity. - Noise is filtered out by averaging potential over
flux surface. - So far noise is reduced by brute force ...
higher N!
Scaled Cyclone Base Case with kinetic electrons
T100 eV, n4.51019 m-3
35Contribution of noise to the heat flux
The convective noise flux prediction gives a
fairly good estimate of the unphysical ion heat
conductivity in the simulations
36From noise to turbulence spectrum
Wave number spectrum from different time
instants demonstrates how one moves from white
noise to physical spectrum in modes.
37Problematic levels of noise
Fluctuations _at_ ra/2
- Figure show exceptionally bad case regarding
noise effects. - It is a kinetic cyclone base case with scaled
parameters and low T100eV and n4.51017 m-3. - Density fluctuations remain almost constant in
time. - Regression shows almost perfect N-1/2 scaling,
indicating that results are dominated by noise.
38but not always problematic...
- In FT-2, fluctuation levels are much higher
(10-40) than in scaled Cyclone Base Case (1). - So high perturbation level warrants the use of a
full f scheme. - Image shows density fluctuations relative to flux
surface average. - Relative importance of noise values can be seen
in videos of both cases.
39Probability distribution functions
- Fluctuation levels can be represented by the PDF
graphs. - PDFs show fluctuation distributions over a middle
flux surface. - Density values are averaged over time.
- Turbulence in FT-2 takes fluctuations up to 40
in start (up) and 15 after 50µs (down).
40Section VIITransport simulations
41Case 1 under study LH heated FT-2
- Parameters from 100 kW LH heated 22 kA FT-2
tokamak. - Case shows a rapid growth of potential and
electric field and reduction of thermal
conductivity interpreted as ITB formation. - Top figure diagram (r,t) of flux surface average
of potential. - Bottom figure ion diffusivity at mid radius.
42Evolution of profiles
- Strong ExB flow shear is created at r0.05 m,
where a knee-point in Ti profile is found
43Calculated spectra
- Parameters of the FT-2 case, devised to cause the
formation of Internal Transport Barrier. - B2.2T, I22kA, n3.51019m-3, Ti250eV, Te300eV
S(k) at r5.1cm
? vs. time
S(k) at r7.5cm
44Case 2 under study LH heated FT-2
- Heating phase for 100 kW LH heated 22 kA FT-2
tokamak (O8 impurities included).
45Evolution of large scale fluctuations
- Density fluctuations plots show the formation and
further destruction of macroscopic structures
46Evolution of diffusivity
- Both particle diffusivity and heat conductivity
drop drastically when poloidal flow shear
destroys the turbulent structures - The figures show values from the middle radius
47Case 3 under study Ohmic FT-2
- Reproduce the FT-2 reflectometer signal I(t)
with ELMFIRE - I(t)? w(r,?) dn(r,?,t) r dr d?
- W(r, ? ) Weighting function calculated by beam
tracing code using exp. data - dn(r,?,t) Density fluctuations simulated by
ELMFIRE code - Results
- Poloidal velocity calculated by the shift of the
power spectrum of I(t) is in reasonable agreement
with the poloidal velocity measured by the
Doppler reflectometer for an Ohmic discharge - Width of the power spectrum is too narrow
compared to the experimental results.
48Frequency broadening is too narrow in the
simulation
49Section VIIIFuture challenges
50Unresolved issues in full f
- How to ensure a sufficient number of particles in
all grid cells when density varies strongly in
global PIC simulation - Adaptation of good grid resolution for strongly
varying f in Vlasov simulations - Full collision operator in Vlasov simulations
- SOL plasma simulation with sheath boundaries
- Gyrokinetics with strong fluctuations
51Scalability
IB-new Sepeli with InfiniBand and optimized
parallelization IB Sepeli with InfiniBand LAM
Sepeli with Gigabit Ethernet IBM-new IBM with
optimized parellelization
52Resource scenario
53Conclusions
- Successful 5D gyrokinetic full f PIC and Vlasov
simulations of tokamak electrostatic turbulence
and transport for core plasma. - Careful benchmarking of the codes is performed in
appropriate limits for the turbulence saturation
and neoclassical characteristics. - Linear and nonlinear benchmarking.
- Vlasov approach has less noise and better control
of resolution PIC code is more flexible to
implement - Most urgently needed for edge plasma simulations
further development of nonlinear terms in
gyrokinetics may be needed
ACKNOWLEDGEMENTS
This project receives funding from the European
Commission
CSC The Finnish IT Center for Science for its
computing facilities