knosos: a fast orbit-averaged neoclassical code for arbitrary...

29
KNOSOS: a fast orbit-averaged neoclassical code for arbitrary stellarator geometry Jos´ e Luis Velasco Laboratorio Nacional de Fusi´ on, CIEMAT Developed in collaboration with Iv´ an Calvo, F´ elix Parra (U. Oxford, CCFE) Jos´ e Manuel Garc´ ıa-Rega˜ na and J Arturo Alonso Kinetic Theory Working Group Meeting 2018, Madrid Laboratorio Nacional de Fusión Aknowledgements: A. Moll´ en, C. Nuehrenberg (IPP), S. Satake, S. Morita, XL Huang (NIFS)...

Upload: others

Post on 10-Jul-2020

7 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

KNOSOS: a fast orbit-averaged neoclassicalcode for arbitrary stellarator geometry

Jose Luis Velasco

Laboratorio Nacional de Fusion, CIEMAT

Developed in collaboration with Ivan Calvo, Felix Parra (U. Oxford,CCFE) Jose Manuel Garcıa-Regana and J Arturo Alonso

Kinetic Theory Working Group Meeting 2018, Madrid

Laboratorio Nacional de Fusión

Aknowledgements: A. Mollen, C. Nuehrenberg (IPP),S. Satake, S. Morita, XL Huang (NIFS)...

Page 2: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Motivation and goal

In stellarators, neoclassical transport:

Tends to dominate radial particle and energy fluxes in the coreof reactor-relevant plasmas [Dinklage, NF (2013)].

Explains impurity accumulation ob-served in many stellarator plasmas[Burhenn, NF (2009)].

Energy transport simulations Equations Algorithms First results Ongoing work 1/26

Page 3: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Motivation and goal

Goal of this work: write a neoclassical code (i.e., a code thatsolves the Drift Kinetic Equation together with the quasineutralityequation) that is, at the same time, accurate and fast:

Accurate so that:

it gets the simulation quantitatively closer to the experiment;

Fast so that:

it can be included in an optimization loop;it allows to perform comprehensive parameter scans;

i.e., generally speaking, so that it improves confidence in ourpredictions and that these predictive capabilities can be fullyexploited.

Energy transport simulations Equations Algorithms First results Ongoing work 2/26

Page 4: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Outline

Energy transport simulations (short overview).

Equations.

Algorithms.

First simulations for arbitrary geometry.

Examples of other applications.

Energy transport simulations Equations Algorithms First results Ongoing work 3/26

Page 5: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Stellarator-specific regimes, εeff

At low collisionality, stellarator neoclassical transport is verydifferent from (and larger than in) tokamaks [Helander, PPCF (2012)].

Several strategies for characterizing magnetic configurations w.r.t.neoclassical energy transport:

1 Calculation of the level of 1/ν transport.

One scalar, effective ripple, gives level of 1/ν transport.

Fast calculation of εeff provided by NEO [Nemov, PoP (1999)].

Fails to explain configuration dependence of the experimentalτE stellarator scaling [Yamada, NF (2005)].

Energy transport simulations Equations Algorithms First results Ongoing work 4/26

Page 6: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Neoclassical τE [Pedersen, NF (2015)]

2 Calculation of τE solving the energybalance for every species b:

3

2

〈∂nbTb〉∂t

+1

V ′∂

∂rV ′〈Qb〉 = 〈Pb〉 .

Qb calculated with DKES [Hirshman, PoF (1986)]1.

τE ∼∫dV

∑nbTb∫

dV∑Pb

, gives level of transport for given configuration and

sources [Turkin, PoP (2011), Geiger, PPCF (2015)].

Monoenergetic approach by DKES (and others [Beidler, NF (2011)])saves some time by calculating a database of transport coefficientsinstead of solving the DKE multiple times.

Still, for each new magnetic configuration, creating databasetakes time (∼ 1− 2 days [Beidler, W7-X Workshop (2016)]).See also [Landreman, APS (2017)].

It oversimplifies the DKE (next slides).1

The net energy source is provided by a suite of codes, and there is an ad-hoc anomalous model at the edge.

Energy transport simulations Equations Algorithms First results Ongoing work 5/26

Page 7: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Improved calculations of neoclassical energy flux

3 Accurate calculation of Qb for fixed kinetic profiles.

Recently, there have been refinements of some aspects of neo-classical theory that have lead to more complete neoclassicalcodes (EUTERPE, SFINCS, FORTEC-3D...):

electric field tangent to magnetic surface (aka ϕ1);magnetic drift tangent to magnetic surface;radially global effects;complete collision operator.

Calculations with all the ingredients can get extremely time-consuming and memory-demanding, specially at low collisionalities.

⇒ Comprehensive parameter scans may become impossible.

Energy transport simulations Equations Algorithms First results Ongoing work 6/26

Page 8: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Specific goal

We have outlined three strategies for characterizing the energytransport of a magnetic configuration. Our goal is to create a codethat can contribute to the three strategies by:

1 Giving figures of merit for neoclassical transport of collisionalityregimes other than the 1/ν (i.e.

√ν and sb-p).

2 Employing a different ”database of coefficients” that allows3

to be very fast at low collisionalities and, at the same time,to retain all the necessary ingredients of neoclassical theory.

KNOSOS solves the bounce-averaged drift kinetic and quasineutralityequations discussed in [Calvo, PPCF (2017)] for arbitrary geometry.

Energy transport simulations Equations Algorithms First results Ongoing work 7/26

Page 9: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Coordinates on phase space

We are interested in calculating the ion2 distribution functionF (ψ(x), α(x), l(x), E , µ, σ), where:

ψ is the radial coordinate.We choose ψ = ψt, where 2πψt is the toroidal magnetic flux.

α is an angle that labels magnetic field lines on the surface.We choose α ≡ θ − ιζ, where θ and ζ are Boozer angles and ιis the rotational transform.

l is the arc length along the magnetic field line.

B = ∇ψ ×∇α ,b = B/B = B/|B| .

E = v2 + Ze/mϕ is the total energy per mass unit, v = |v|,and ϕ(ψ, α, l) is the electrostatic potential.

µ is the magnetic moment.

σ = v‖/v is the sign of the parallel velocity, v‖ = v · b.2

We will take a mass ratio expansion√me/m � 1; all quantities will refer to ions and Ze, m, etc, will

have the usual meaning.

Energy transport simulations Equations Algorithms First results Ongoing work 8/26

Page 10: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Drift kinetic equation

In these coordinates, the equation for the ion distribution func-tion F can be written as

(v‖b + vd) · ∇F = C[F, F ] ,

where vd = vM +vE is the sum of magnetic and E×B drifts,and C[F, F ] is the (ion-ion) collision operator.

We know that the motion of particles along the magnetic fieldis much faster than the motion perpendicular to it.

At low collisionalities (ν∗ ≡ νiiR0/vT � ε3/2), the parallelstreaming term also dominates over the collision term for bothpassing and trapped trajectories.

In an expansion in the normalized Larmor radius, the first termis:

v‖∂lF = 0 ,

and, to next order, F is determined by integrals over l of thedrift-kinetic equation.

Energy transport simulations Equations Algorithms First results Ongoing work 9/26

Page 11: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Bounce-averaged drift kinetic equation

We have:

Passing particles :

∫ 2π

0dα

∫ lmax(ψ,α)

0

1

|v‖|C[F, F ]dl = 0 .

Trapped particles : −∂ψJ∂αF+∂αJ∂ψF =∑σ

Ze

m

∫ lb2

lb1

1

|v‖|C[F, F ]dl .

Everything has become expressed in terms ofderivatives of the second adiabatic invariant:

J(ψ, α, E , µ) ≡ 2

∫ lb2

lb1

|v‖|dl .

At the bounce points, lb1 and lb2 , v‖ = 0.

Energy transport simulations Equations Algorithms First results Ongoing work 10/26

Page 12: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Bounce-averaged drift kinetic equation (II)

−∂ψJ∂αF + ∂αJ∂ψF =∑σ

Ze

m

∫ lb2

lb1

1

|v‖|C[F, F ]dl .

These derivatives of the second adiabatic invariant are connectedto the drifts perpendicular to the magnetic field line:

∂αJ =2Ze

m

∫ lb2

lb1

1

|v‖|vd·∇ψ dl , ∂ψJ = −2Ze

m

∫ lb2

lb1

1

|v‖|vd·∇α dl .

∂αJ contains the radial magnetic and E × B drift caused bythe inhomogeneity of the magnetic field strength and of theelectrostatic potential on the flux surface respectively.

Can model 1/ν, global effects and transport caused by ϕ1.

∂ψJ contains the precession tangential to the flux surface causedby the radial variation of the magnetic field strength and of theelectrostatic potential (i.e. Er) respectively.

Eqs. contain the√ν and superbanana-plateau regimes.

Energy transport simulations Equations Algorithms First results Ongoing work 11/26

Page 13: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Local bounce-averaged drift kinetic equation

We would like to solve a drift-kinetic equation that

is linear and radially local and,

rigorously includes the tangential magnetic drift.

It can be proven that [Calvo, PPCF (2017)] this can be done rigorouslyif the magnetic configuration is optimized enough. Then3

|∂αJ∂ψg| � |∂ψJ∂αg| ,where g(ψ, α, v, λ) is the dominant piece of the non-adiabaticcomponent of the deviation of the ion distribution function from aMaxwellian FM :

g = F − FM [1− (Zeϕ1/T )] ,

and we have changed coordinates to:

v = 2√E − Zeϕ0/m) and λ = µ/(E − Zeϕ0/m) ,

and used ϕ(ψ, α, l) = ϕ0(ψ) + ϕ1(ψ, α, l) with |ϕ1| � |ϕ0| .3

This also happens in large aspect-ratio stellarator for large values of the radial electric field.

Energy transport simulations Equations Algorithms First results Ongoing work 12/26

Page 14: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Local bounce-averaged drift kinetic equation (II)

The equations that determine g(ψ, α, v, λ) are∫ lb2

lb1

dl

|v‖|vd · ∇α ∂αg +

∫ lb2

lb1

dl

|v‖|vd · ∇ψΥFM =

∫ lb2

lb1

dl

|v‖|C lin[g]

with

∫ 2π

0

g dα = 0 for trapped particles and g = 0 for passing

and

(Z

T+

1

Te

)ϕ1 =

en

∫ ∞

0

dv

∫ B−1

B−1max

dλv3B

|v‖|g

Υ =∂ψnn +

∂ψTT

(mv2

2T −32

)+

Ze∂ψϕ0

T .

C lin[g] is the linearized collision operator. We will use pitch-angle-scattering for the moment.

vE · ∇α ≈ ∂ψϕ0, so the system is linear in ϕ1.

This is the set of equations solved by KNOSOS4.4

We do not use explicitly the expansion of [Calvo, PPCF (2017)] of the magnetic configuration around aperfectly optimized stellarator as in [Velasco, PPCF (2018)] but, for small |∂ψϕ0|, if the stellarator is not optimizedenough, it is not justified to use the local DKE.

Energy transport simulations Equations Algorithms First results Ongoing work 13/26

Page 15: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Explicit expressions

Using the expressions of the drift in Boozer coordinates5,everything gets written in terms of a few bounce integrals:(IvM,α(α, λ) +

1

vdIvE ,α(α, λ)

)∂αg +

(IvM,ψ (α, λ) +

1

vdIvE,ψ (α, λ)

)FMΥ +

=vν

vd∂λIν(α, λ)∂λg ,

with vd ≡ mv2

Ze and

IvE,α = ∂ψϕ0

∫ lb2

lb1

dl√1− λB

, IvM,α =1− λ

2

∫ lb2

lb1

dl√1− λB

∂ψB

B,

IvE,ψ = −∫ lb2

lb1

dl√1− λB

∂θϕ1 , IvM,α = −1− λ2

∫ lb2

lb1

dl√1− λB

∂θB

B,

Iν =

∫ lb2

lb1

dl

√1− λBB

.

5For the sake of simplicity, in this talk we assume vaccum magnetic field, B = Bζ∇ζ.

Energy transport simulations Equations Algorithms First results Ongoing work 14/26

Page 16: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Coefficients

Integrals in l done using extended midpoint rule:

Number of points tripled until the result converges.Note that integrals like:

I =

∫ l2

l1

dlf(l)√

1− λB(l)

diverge logarithmically close to the passing/trapped boundary. Onecan ease the convergence (and thus speed up the calculation) byremoving the divergence and solving it analytically:

I = I0 +∑j

Ij ,

I0 =

∫ l2

l1

dl

f(l)√1− λB(l)

−∑j

f(lj)√−λ(l− lj)[∂lB|lj + 1

2∂2lB|lj (l− lj)]

,

Ij =

√√√√ −2

λ∂2lB|lj

ln

√√√√ ∂2lB|lj−2

√−∂lB|lj x−

1

2∂2lB|lj x

2 − λ(∂2l B|l1x + ∂lB|lj )

l2−l1

0

.

The same can be done at other boundaries (e.g. from trapped inone period to trapped in two periods).

Energy transport simulations Equations Algorithms First results Ongoing work 15/26

Page 17: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Coefficients (II)

When integrating along the field line using fixed step ∆ζ,∆θ,instead of evaluating

B(θ, ζ) =∑m,n

Bm,n cos[mθ + nζ] ,

B(θ + ∆θ, ζ + ∆ζ) =∑m,n

Bm,n cos[m(θ + ∆θ) + n(ζ + ∆ζ)] ,

B(θ + 2∆θ, ζ + 2∆ζ) =∑m,n

Bm,n cos[m(θ + 2∆θ) + n(ζ + 2∆ζ)] ,

...

we can precalculate a few sines and cosines cos(mθ + nζ),sin(mθ + nζ), cos(m∆θ + n∆ζ) and sin(m∆θ + n∆ζ) and usetrigonometric identities to iterate:

B(θ + ∆θ, ζ + ∆ζ) =∑m,n

Bm,n[cos(mθ + nζ) cos(m∆θ + n∆ζ)−

sin(mθ + nζ) sin(m∆θ + n∆ζ)]

Energy transport simulations Equations Algorithms First results Ongoing work 16/26

Page 18: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Discretization

We are left with two variables:

Bounce-averages in l.

Pitch-angle scattering collision operator ⇒ v is a parameter.Calculate ∂vQi for 28 values of v and integrate Gauss-Laguerre.

Radially local equations ⇒ ψ is a parameter.

(IvM,α(α, λ) +

1

vdIvE ,α(α, λ)

)∂αg +

(IvM,ψ (α, λ) +

1

vdIvE,ψ (α, λ)

)FMΥ +

=vν

vd∂λIν(α, λ)∂λg .

Discretize gi,j ≡ g(αi, λj), with uniform grid in λ and α.Second order central differences.

g = 0 at boundary between passing and trapped.

v‖∂λg = 0 at well bottoms.

Subtlety: at bifurcations, g is continuous in α and λ, but ∂λg is notcontinuous in λ.

Energy transport simulations Equations Algorithms First results Ongoing work 17/26

Page 19: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Discretization (II) gi,j ≡ g(αi, λj)(IvM,α(α, λ) +

1

vdIvE ,α(α, λ)

)∂αg +

(IvM,ψ (α, λ) +

1

vdIvE,ψ (α, λ)

)FMΥ +

=vν

vd∂λIν(α, λ)∂λg ,

Top: gi,1 = 0 .

Arbitrary point:

∂αg|i,j =gi+1,j − gi−1,j

2∆α, ∂λg|i,j =

gi,j+1 − gi,j−1

2∆λ,

∂2λg|i,j =

gi,j+1 + gi,j−1 − 2gi,j(∆λ)2

.

Bifurcations (W = I, II...):

∂αg|i,j =gi+1,j,W − gi−1,j

2∆α,

rhs =∑w

Iν ,i,j+1,w−gi,j+3,w + 2gi,j+2,w − 3gi,j+1,w

4(∆λ)2

− Iν ,i,j−1gi,j − gi,j−2

4(∆λ)2.

Bottom (j = N + 1): rhs = −Iν ,i,N−1gi,N−gi,N−2

4(∆λ)2.

Energy transport simulations Equations Algorithms First results Ongoing work 18/26

Page 20: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Linear problem

(IvM,α(α, λ) + 1

vdIvE,α(α, λ)

)∂α

[IvM,ψ(α, λ) +

+vνvd∂λν(α, λ)∂λ g 1

vdIvE,ψ(α, λ)

]FMΥ

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

. . . . . . . . . . . . . . .

·. . .. . .. . .. . .

=

. . .. . .. . .. . .

Square matrix, Nλ ×Nα elements per row.

Sparse, ∼ 5 non-zero elements per row.Non-zero elements are always at the same position for a givenflux-surface, but relative weight varies with v, Er...

Direct solver, based on LU factorization.

Will be useful for solving quasineutrality.

Energy transport simulations Equations Algorithms First results Ongoing work 19/26

Page 21: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Benchmark with DKES (vM · ∇α = 0, ϕ1 = 0)

W7-X (standard configuration, EIM,low ripple) at:

ψ/ψLCMS = 0.3 (top right).

ψ/ψLCMS = 0.5 (bottom left).

ψ/ψLCMS = 0.7 (bottom right).

10-4

10-3

10-2

10-1

100

10-6 10-5 10-4 10-3 10-2 10-1

Γ11

[A.U

.]

ν/v [m-1]

Er/v=0.0Er/v=1x10-5

Er/v=3x10-5

Er/v=1x10-4

Er/v=3x10-4

Er/v=1x10-3

Er/v=3x10-3

10-4

10-3

10-2

10-1

100

10-6 10-5 10-4 10-3 10-2 10-1

Γ11

[A.U

.]

ν/v [m-1]

Er/v=0.0Er/v=1x10-5

Er/v=3x10-5

Er/v=1x10-4

Er/v=3x10-4

Er/v=1x10-3

Er/v=3x10-3

10-4

10-3

10-2

10-1

100

10-6 10-5 10-4 10-3 10-2 10-1

Γ11

[A.U

.]

ν/v [m-1]

Er/v=0.0Er/v=1x10-5

Er/v=3x10-5

Er/v=1x10-4

Er/v=3x10-4

Er/v=1x10-3

Er/v=3x10-3

Energy transport simulations Equations Algorithms First results Ongoing work 20/26

Page 22: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Benchmark with DKES (vM · ∇α = 0, ϕ1 = 0)

LHD (R = 3.75) at ψ/ψLCMS = 0.5(top right).

QIPC at:

ψ/ψLCMS = 0.1 (bottom left).

ψ/ψLCMS = 0.25 (bott. right).

10-6

10-5

10-4

10-3

10-2

10-1

100

101

10-6 10-5 10-4 10-3 10-2 10-1

Γ11

[A

.U.]

ν/v [m-1]

Er/v=0.0Er/v=1x10-5

Er/v=3x10-5

Er/v=1x10-4

Er/v=3x10-4

Er/v=1x10-3

Er/v=3x10-3

10-3

10-2

10-1

100

101

102

10-6 10-5 10-4 10-3 10-2 10-1

Γ11

[A

.U.]

ν/v [m-1]

Er/v=0.0Er/v=1x10-5

Er/v=3x10-5

Er/v=1x10-4

Er/v=3x10-4

Er/v=1x10-3

Er/v=3x10-3

10-5

10-4

10-3

10-2

10-1

100

101

10-6 10-5 10-4 10-3 10-2 10-1

Γ11

[A

.U.]

ν/v [m-1]

Er/v=0.0Er/v=1x10-5

Er/v=3x10-5

Er/v=1x10-4

Er/v=3x10-4

Er/v=1x10-3

Er/v=3x10-3

Energy transport simulations Equations Algorithms First results Ongoing work 21/26

Page 23: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Benchmark with DKES (vM · ∇α = 0, ϕ1 = 0)

49 simulations at low collisionalities

Nα=32, Nλ=256.

Total time: 2.4 s (0.05 s/point)at one single core of Dual Xeonquad-core 3.0 GHz (∼ severaltimes slower than e.g. Marconi).

Relatively large overhead:

Grid in α and λ: 0.2 s.Bounce integrals 0.9 s.Integral of g in λ: 0.2 s.

Easy to parallelize: v, ψ, bounceintegrals, ∂ψϕ0, ϕ1 (next slide)...

10-4

10-3

10-2

10-1

100

10-6 10-5 10-4 10-3 10-2 10-1

Γ11

[A.U

.]

ν/v [m-1]

Er/v=0.0Er/v=1x10-5

Er/v=3x10-5

Er/v=1x10-4

Er/v=3x10-4

Er/v=1x10-3

Er/v=3x10-3

* One could think that anything below (say) 1 s is equally fine, butnote that solving ambipolarity, quasineutrality, transport balance...requires evaluating the drift kinetic equation MANY times.

Energy transport simulations Equations Algorithms First results Ongoing work 22/26

Page 24: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Solving quasineutrality (matrix response approach)

[IvM,α(α, λ) +

1

vdIvE,α(α, λ)

]∂αg(α, λ)− vν

vd∂λIν(α, λ)∂λg(α, λ) =

=

(IvM,ψ (α, λ)− 1

vd

∫ lb2

lb1

dl√1− λB

∂θϕ1

)FMΥ , (1)

(Z

T+

1

Te

)ϕ1 =

en

∫ ∞0

dv

∫ B−1

B−1max

dλv3B

|v‖|g . (2)

If we parametrize ϕ1 using N vectors uk and N coefficients, thenthe linear system can be esquematically written as

ϕ1 = ϕ01 +Aϕ1 (3)

where A is a N ×N matrix:

ϕ01 obtained by solving (1) with ϕ1 = 0 and then using (2).

The kth row of A is obtained by solving (1) with ϕ1 = uk,

solving (2) and substracting ϕ01.

DKE is solved N + 1 times, but LU factorization is done once(e.g., in prev. case, for N = 1000, CPU time is 40 times larger).

Once ϕ01 and A are known, the linear system (3) can be solved.

Energy transport simulations Equations Algorithms First results Ongoing work 23/26

Page 25: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Example: effect of plasma profiles on transport ofcollisional impurities [Velasco et al., in preparation]

Besides, depending on the plasma parameters, affecting Qi, weknow that ϕ1 caused by the bulk species may have an importanteffect on radial transport of collisional impurities.

Yet, only a total of ∼ 100 calculations of ϕ1 (and of ΓZ) havebeen published in papers [Garcıa-Regana, NF (2013), Garcıa-Regana,

NF (2017), Garcıa-Regana, NF (2018), Mollen, PPCF (2018), Velasco,

PPCF (2018)].

Roughly speaking, this covers: ∼ 5 magnetic configurations ×5 radial positions × 5 different plasma profiles.

More comprehensive study (dependence on the configuration,colllisionality, bulk plasma profiles...) remains to be done.

We will do so by combining ϕ1 calculated with KNOSOS and an-alytical formulas derived the flux of collisional impurities [Calvo,

arXiv (2018), TTF (last week)].

Energy transport simulations Equations Algorithms First results Ongoing work 24/26

Page 26: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Example (II): ΓW44 vs. bulk profile shape

Same plasma parameters as in [Calvo, arXiv (2018), TTF (last week)].

LHD6 (R0 = 3.67 m) at√ψ/ψLCMS = 0.8.

n = 1.2× 1020 m3, T = 1.3 keV.

Er = T2e

(∂rnn + 3.37∂rTT

)(ion root limit, 1/ν regime).

1008 calculations of ϕ1 andΓW44.

For ϕ1 = 0, outward pinch(V > 0) if ∂rT/T > 2∂rn/n[Helander, PRL (2017)].

With ϕ1 6= 0, outward pinchfor hollow density only.

Experimental evidence:[Huang, NF (2017)].

-5

-4

-3

-2

-1

0

-5 -4 -3 -2 -1 0 1 2

T’/

T [

m-1

]

ni’/ni [m-1]

Er<0 Er>0 | |

|

|||

V(ϕ1=0)<0 V(ϕ1=0)>0

Er<0 Er>0 | |

|

|||

V(ϕ1=0)<0 V(ϕ1=0)>0

-5

-4

-3

-2

-1

0

1

2

V[m

/s]

6For this plot, we used a preliminary version of KNOSOS that makes certain assumptions on the magnetic

configuration [Velasco, PPCF (2018)].

Energy transport simulations Equations Algorithms First results Ongoing work 25/26

Page 27: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Conclusions

We have developed the code KNOSOS. It can provide very fastcalculations of:

Level of transport in several low-collisionality regimes (1/ν,√ν

and superbanana-plateau) for arbitrary geometry.DKES-like radial fluxes including the effect of vM · ∇α.ϕ1 and Er (ongoing).

Benchmarking and experimental validation:

Radial fluxes agains DKES and FORTEC-3D.ϕ1 against EUTERPE, SFINCS, FORTEC-3D and analytical ex-pressions of [Calvo, JPP (2018)].Comparison with measurements (e.g. Qi and Er in W7-X [Alonso,

EPS (2017), TTF (2018)]).

Next steps:

Transition between 1/ν and plateau.Solve DKE for electrons.Improve collision operator.

Energy transport simulations Equations Algorithms First results Ongoing work 26/26

Page 28: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Backup

Energy transport simulations Equations Algorithms First results Ongoing work 27/26

Page 29: KNOSOS: a fast orbit-averaged neoclassical code for arbitrary …fusionsites.ciemat.es/kineticsmadrid2018/files/2018/09/... · 2018-09-19 · KNOSOS: a fast orbit-averaged neoclassical

Energy transport simulations Equations Algorithms First results Ongoing work 27/26