RESEARCH ARTICLE | JANUARY 19 2023

Linearized frequency domain Landau-Lifshitz-Gilbert

equation formulation
Special Collection: 67th Annual Conference on Magnetism and Magnetic Materials

Zhuonan Lin; Vitaliy Lomakin 

AIP Advances 13, 015216 (2023)

https://doi.org/10.1063/9.0000609
 CHORUS


View Export
Online Citation

Articles You May Be Interested In

Optimal control of magnetization dynamics in ferromagnetic heterostructures by spin-polarized currents

J. Appl. Phys. (November 2010)

Midpoint numerical technique for stochastic Landau-Lifshitz-Gilbert dynamics

J. Appl. Phys. (April 2006)

Numerical integration of Landau–Lifshitz–Gilbert equation based on the midpoint rule

14 October 2024 15:27:01

J. Appl. Phys. (May 2005)
 AIP Advances ARTICLE scitation.org/journal/adv

Linearized frequency domain

Landau-Lifshitz-Gilbert equation formulation
Cite as: AIP Advances 13, 015216 (2023); doi: 10.1063/9.0000609
Submitted: 4 October 2022 • Accepted: 31 October 2022 •
Published Online: 19 January 2023

Zhuonan Lin1, 2 , 3 and Vitaliy Lomakin1, 2,a)

AFFILIATIONS
1
Department of Electrical and Computer Engineering, University of California San Diego, La Jolla, California 92093, USA
2
Center for Memory and Recording Research, La Jolla, California 92093, USA
3
Materials Science and Engineering Program, University of California San Diego, La Jolla, California 92093, USA

Note: This paper was presented at the 67th Annual Conference on Magnetism and Magnetic Materials.
a)
Author to whom correspondence should be addressed: [email protected]

ABSTRACT
We present a general finite element linearized Landau-Lifshitz-Gilbert equation (LLGE) solver for magnetic systems under weak time-
harmonic excitation field. The linearized LLGE is obtained by assuming a small deviation around the equilibrium state of the magnetic

14 October 2024 15:27:01

system. Inserting such expansion into LLGE and keeping only first order terms gives the linearized LLGE, which gives a frequency domain
solution for the complex magnetization amplitudes under an external time-harmonic applied field of a given frequency. We solve the linear
system with an iterative solver using generalized minimal residual method. We construct a preconditioner matrix to effectively solve the linear
system. The validity, effectiveness, speed, and scalability of the linear solver are demonstrated via numerical examples.
© 2023 Author(s). All article content, except where otherwise noted, is licensed under a Creative Commons Attribution (CC BY) license
(http://creativecommons.org/licenses/by/4.0/). https://doi.org/10.1063/9.0000609

I. INTRODUCTION for a complex magnetization amplitude, which can, then, be used

The magnetization dynamics is described by the Landau- to provide linear time domain solutions. The FD-LLGE solver is
Lifshitz-Gilbert equation (LLGE). LLGE is a time dependent non- developed based on the finite element based micromagnetic simula-
linear equation, and it describes the magnetization dynamics in both tor FastMag.7 The linear system is solved effectively with an iterative
linear and non-linear regimes. Solving LLGE can be computation- solver. The number of iterations is significantly reduced using a lin-
ally expensive as it requires obtaining solutions at many time steps ear preconditioner. Solving the FD-LLGE is much more efficient and
and it is often numerically stiff, which may require either small provides a more physically insight than solving the original LLGE.
time steps or assisting linear solvers. Under weak dynamics excita-
tions, however, the magnetization dynamic response may be linear. II. FORMULATION
Examples of such systems are spin wave excitations under weak The magnetization dynamics is described by the LLGE, which
time-dependent applied fields or the initial dynamics under spin is, in its implicit form, is written as
torque excitations.1,2 In such cases, LLGE can be linearized in terms
of a weak magnetization deviation around the equilibrium state. ∂m ∂m
= −γm × (Heff + Ha ) + αm × , (1)
Such linearization has been used to obtain solutions in terms of the ∂t ∂t
eigenstate representations3 and it can be used to obtain solutions where m is the normalized magnetization, γ is the gyromagnetic
even for non-linear problems.4–6 ratio, α is the damping constant. The term Ha = H0a + ha is the
Here, we present a linear frequency domain LLGE (FD-LLGE) applied field containing a static component H0a and a dynamic
solver. The FD-LLGE solver provides a linearized magnetization component ha , which is a time harmonic excitation at a circular
solution as a response to a dynamic excitation, e.g., the applied field, frequency ω, i.e., it can be written as
which is a characterized by a given frequency. Considering a given
frequency allows formulating a time independent linear equation ha = Re{h̃a (r)e jωt }, (2)

AIP Advances 13, 015216 (2023); doi: 10.1063/9.0000609 13, 015216-1

© Author(s) 2023
 AIP Advances ARTICLE scitation.org/journal/adv

where h̃a is the complex amplitude of the exciting magnetic field, ∂v ∂v

= −γ(Λ(H0 )v + Λ(m0 )h + Λ(m0 )ha ) + αΛ(m0 ) . (7)
which represents its magnitude and phase. The term Heff in LLGE ∂t ∂t
(1) is the effective field. The effective field is a function of m, and it is
composed of several components, including the magnetostatic field Since the v is normal to m0 , we can project every vector
Hms , exchange field Hex , and anisotropy field (assumed uniaxial) and operator onto the tangent space TM(m0 ) of m0 , by using the
Han : projection operator9

Heff = Hms + Hex + Han ≡ Cm, Pm0 = (I − m0 ⊗ m0 ), (8)

∇⋅m ′
Hms = Ms ∇ ∫ dr ,
∣r − r′ ∣ where ⊗ is the dyadic Kronecker tensor product and I is the identity
(3) matrix. It can be shown that when restricted to the vector fields in
2A 2
Hex = ∇ m, TM(m0 ), the operator Λ(m0 ) is linear and anti-symmetric and it is
Ms also invertible,9 i.e.,
Han = HK (k̂ ⋅ m)k̂.
Λ(m0 )Λ(m0 ) = −I. (9)
Here, M S is the saturation magnetization, A is the exchange con-
stant, H K is the anisotropy field, and k̂ is the uniaxial anisotropy axis With Eqs. (8) and (9), we can simplify Eq. (7) by multiplying by
direction. The effective field is linear in m and, therefore, C is the Λ(m0 ) at both sides:
linear field operator that is independent of m.
The LLGE (1) is non-linear in m due to the presence of the ∂v
cross products and it describes the magnetization dynamics in a B– + γA– v = γha , (10)
broad range of situations, including linear and non-linear effects. ∂t
In various cases, however, the general LLGE can be linearized. Such
a linearization is allowed when the magnetization varies insignifi- where
cantly around its equilibrium state. Weak magnetization variations
can be due to weak excitations, e.g., by weak applied fields or by spin A– = Pm0 ((H0 ⋅ m0 )I − C)

14 October 2024 15:27:01

transfer torque (STT). Even in the general non-linear cases, the ini- (11)
B– = Pm0 (Λ(m0 ) + αI).
tial dynamics that contains important information about the system
behavior can be characterized by the linearization.3,4
We present a framework that uses a linearized LLGE to study Assuming the linearity of the response, the small deviation v can be
the magnetization dynamics. We first present a linearized time expressed as
domain LLGE. Then, we construct a FD-LLGE to study small
oscillations around the equilibrium state driven by time-harmonic v = Re{ṽ(r)e jωt }, (12)
excitations.
We seek a solution for small magnetization deviation v around
the equilibrium state such that where ṽ is the complex amplitude of the magnetization deviation.
Inserting Eq. (12) into Eq. (10), we obtain a FD-LLGE
m = m0 + v. (4)
( jωB– + γA– )ṽ = γh̃a. (13)
Here, m0 is the equilibrium magnetization state for the sys-
tem without a dynamic excitation, which is given by the Brown
condition8 Here, h̃a in the right-hand side is a known function of spatial coor-
dinates and ṽ is an unknown, which is found by solving the linear
m0 × Heff (m0 ) = 0, (5) system of equations.
The linear system matrix of Eq. (13) is dense due to the pres-
which corresponds to ∂m0 /∂t = 0 in the LLGE. ence of the magnetostatic field operator Hms . Therefore, to enable
The magnetization deviation v is normal to m0 , so that the solving large problems, iterative methods, such as the conjugated
normalization of m is preserved. Because of the linearity of Heff in gradient (CG) or generalized minimal residual method (GMRES),10
terms of m, we can write Heff (m) = H0 + Cv, where H0 = Heff (m0 ). typically should be used. Micromagnetic problems often are stiff,
Inserting representation (4) into LLGE (1), and linearizing the equa- which results in a badly conditioned matrix and many linear iter-
tion by keeping only the terms linear in v, α and ha , we can write a ations required for the solution. Reducing the number of iterations
linearized LLGE for v: can be achieved by developing a proper preconditioner. It has been
∂v ∂v shown that the high condition number for micromagnetic systems is
= −γ(v × H0 + m0 × h + m0 × ha ) + αm0 × , (6) a result of the effects of the exchange field linear operator.11 To alle-
∂t ∂t
viate this problem, we use the projected sparse matrix Cex – = Pm0 Cex
where h = Cv is the dynamic effective field corresponding to v. of the exchange field operator Cex , which represents the discretiza-
Denoting the cross operator as Λ(u)v = u × v, we can express Eq. (6) tion of the (2A/M s )∇2 operator in Eq. (3), to construct a sparse
in its matrix form preconditioner matrix

AIP Advances 13, 015216 (2023); doi: 10.1063/9.0000609 13, 015216-2

© Author(s) 2023
 AIP Advances ARTICLE scitation.org/journal/adv

P = (jωB– + γCex – ). (14) confined to the region of 10 nm near the center of the strip at fre-
quency f = ω/(2π) = 20 GHz. The results were obtained on an Intel
The inverse matrix of preconditioner P−1 can be approximately Core i9-9900K CPU.
calculated by incomplete LU decomposition (ILU)10 or using the Figure 1 demonstrates the validity of the solver by comparing
block-inverse preconditioner.11 Multiplying by P−1 in both sides of the results obtained via the FD-LLGE (15) with (12) and via the orig-
the Eq. (13), we obtain a preconditioned linear problem inal LLGE (1). The figure shows the space averaged y-component
of v for L = 100 nm. The number of mesh vertices for this case
was 3,927. The number of iterations based on the ILU0 precondi-
P−1 (jωB– + γA– )ṽ = γP−1 h̃a. (15)
tioner for the FD-LLGE solver was 57 and the computational time
was 3 s. The initial, i.e., early time dynamics, is different, which is
The preconditioned linear system of Eq. (15) has a significantly because for the non-linear solver there are initial non-linear effects.
reduced condition number and can be solved in a much smaller The later time dynamics is described by the FD-LLGE solver accu-
number of iterations than the original problem of Eq. (13). rately. A snapshot of vy at t = 10 ns shown in the bottom inset of
Fig. 1 demonstrates a standing spin wave pattern.
III. NUMERICAL RESULTS Figure 2 demonstrates the computational performance of the
FD-LLGE solver by showing the number of iterations and computa-
We implemented the FD-LLGE solver as a part of the finite ele-
tional time as a function of the strip length L and the corresponding
ment method (FEM) based micromagnetic simulator FastMag with
number of tetrahedral mesh vertices N. The results are shown for
tetrahedral elements.7 The effective fields are computed as for the
the non-preconditioned formulation of Eq. (13) and preconditioned
general LLGE. We note that the computation of the used FEM based
formulation of (15). It is evident that the non-preconditioned solver
effective field operator, such as Fredkin-Koehler method,12 results
requires many iterations, whereas using preconditioners leads to a
in non-perfectly Hermitian linear system operator A– .13 Potentially,
significant reduction of the number of iterations. The ILUk precon-
it may lead to numerical challenges if a solver assumes Hermitian
ditioners perform better than ILU0 in terms of having a smaller
matrix properties. Here, we use a solver and preconditioner that do
number of iterations, but they have a higher cost per iterations.
not assume any special matrix properties. The linear system of equa-
Overall, we find that ILUk with k = 1 has the best performance in
tions (15) is solved using the GMRES algorithm with the relative
terms of the computational time. Overall, the solver performance is
error of 10−8 . The preconditioner is based on the ILU0 and ILUk fla-
good, and it allows addressing large scale computational problems.

14 October 2024 15:27:01

vors of the ILU preconditioner using the sparse matrix P of Eq. (14).
We note that the solving the linear system without using a precon-
The results are shown for a magnetic strip of width w = 30 nm, thick-
ditioner, i.e., Eq. (13), requires a large number of iterations, e.g.,
ness h = 1 nm, and length L > w ranging from 100 nm to 10 μm. The
1,952 iterations for the L = 100 nm case, which makes such a solver
material parameters are M s = 800 emu/cm3 , Aex = 1 μerg/cm, and
impractical. Therefore, using the preconditioned system of Eq. (15)
α ranges from 0.01 to 0.0001. The maximal mesh edge length was
is critical.
chosen
√ as 2 nm to be sufficiently smaller than the exchange length Finally, Fig. 3 shows the space averaged magnitude ⟨∣ṽy ∣⟩ as a
of A/Ms = 12.5 nm. The equilibrium magnetization is along the function of frequency f . One can see that the solution ⟨∣ṽy ∣⟩ exhibits
longest direction (see the top inset in Fig. 1). The magnetization a strong resonance response, with the resonant frequency related to
dynamics is excited by a weak applied magnetic field of h̃a = 50 Oe the resonant standing spin wave excitation. The figure also shows
the number of iterations with ILUk with k = 1 preconditioner. The
number of iterations increases at the resonant frequency, but this

FIG. 1. Time dynamics of the average magnetization obtained via the full LLGE
and FD-LLGE solvers for L = 100 nm, α = 0.01. The top inset shows the strip with
its magnetization equilibrium state. The bottom inset shows a snapshot of v y at FIG. 2. Size dependence of the linear solver iteration number and computational
t = 10 ns. time when using different preconditioners for α = 0.01.

AIP Advances 13, 015216 (2023); doi: 10.1063/9.0000609 13, 015216-3

© Author(s) 2023
 AIP Advances ARTICLE scitation.org/journal/adv

an Energy Frontier Research Center funded by the U.S. Department

of Energy, Office of Science, Basic Energy Sciences under Award No.
DESC0019273. This work used the XSEDE,14 which is supported
by NSF Grant #ACI-1548562, specifically, it used the Bridges and
Comet systems supported by NSF Grant #ACI-1445506.

AUTHOR DECLARATIONS
Conflict of Interest
The authors have no conflicts to disclose.

Author Contributions

FIG. 3. Frequency dependence of the magnetization deviation magnitude for Zhuonan Lin: Conceptualization (supporting); Data curation (lead);
L = 100 nm, α = 0.0001. Methodology (supporting); Writing – original draft (supporting);
Writing – review & editing (supporting). Vitaliy Lomakin: Concep-
tualization (lead); Data curation (supporting); Methodology (lead);
Writing – original draft (lead); Writing – review & editing (lead).
increase is modest, which demonstrates a good performance even
when the excitation frequency is close to the resonant frequency.
DATA AVAILABILITY
IV. SUMMARY AND DISCUSSION The data that support the findings of this study are available
We presented a FD-LLGE solver that allows obtaining the mag- from the corresponding author upon reasonable request.
netization dynamics solutions driven by weak time harmonic excita-
tions. The formulation is based on linearizing the original non-linear REFERENCES

14 October 2024 15:27:01

time domain LLGE and assuming a single frequency excitation 1
and solution, which results in a single linear system of equations A. V. Khvalkovskiy, D. Apalkov, S. Watts, R. Chepulskii, R. S. Beach, A. Ong,
X. Tang, A. Driskill-Smith, W. H. Butler, P. B. Visscher, D. Lottis, E. Chen,
for the complex magnetization deviation around the magnetiza- V. Nikitin, and M. Krounbi, J. Phys. D. Appl. Phys. 46, 074001 (2013).
tion equilibrium state. The linear system is preconditioned by a 2
Z. Lin and M. Kostylev, J. Appl. Phys. 117, 053908 (2015).
sparse preconditioner that allows significantly reducing the number 3
Z. Lin, I. Volvach, X. Wang, and V. Lomakin, Phys. Rev. Appl. 17, 34016 (2022).
of iterations and computational time required for its solution. 4
J. Fischbacher, A. Kovacs, H. Oezelt, T. Schrefl, L. Exl, J. Fidler, D. Suess,
We note that once the solutions for ṽ is obtained, the time N. Sakuma, M. Yano, A. Kato, T. Shoji, and A. Manabe, AIP Adv. 7, 045310 (2017).
5
domain solution v is found for all time via Eq. (12). M. D’Aquino, C. Serpico, G. Bertotti, T. Schrefl, and I. D. Mayergoyz, J. Appl.
Assuming a small number of linear iterations for solving Phys. 105, 07D540 (2009).
6
Eq. (15), the FD-LLGE solver provides a much more efficient S. Perna, F. Bruckner, C. Serpico, D. Suess, and M. d’Aquino, J. Magn. Magn.
approach than solving the original non-linear LLGE for finding solu- Mater. 546, 168683 (2022).
7
R. Chang, S. Li, M. V. Lubarda, B. Livshitz, and V. Lomakin, J. Appl. Phys. 109,
tions in the linear regime. If the excitation is given by multiple
07D358 (2011).
frequencies, e.g., by a pulse, then multiple frequency domain solu- 8
W. F. Brown, Micromagnetics (Interscience Publishers, 1963).
tions can be combined via the Fourier transform. The results are 9
M. d’Aquino, C. Serpico, G. Miano, and C. Forestiere, J. Comput. Phys. 228, 6130
shown for real-valued frequencies, but solutions can also be obtained (2009).
for complex-valued frequencies, which can provide insights into the 10
Y. Saad, Iterative Methods for Sparse Linear Systems (SIAM, 2003).
magnetization dynamics behavior. The FD-LLGE solver is imple- 11
S. Fu, R. Chang, I. Volvach, M. Kuteifan, M. Menarini, and V. Lomakin, IEEE
mented in the FEM framework, and it can also be extended to finite Trans. Magn. 55, 2910496 (2019).
12
difference implementations. D. R. Fredkin and T. R. Koehler, IEEE Trans. Magn. 26, 415 (1990).
13
F. Bruckner, M. d’Aquino, C. Serpico, C. Abert, C. Vogler, and D. Suess, J. Magn.
ACKNOWLEDGMENTS Magn. Mater. 475, 408 (2019).
14
J. Towns, T. Cockerill, M. Dahan, I. Foster, K. Gaither, A. Grimshaw,
This work was supported as part of Quantum Materi- V. Hazlewood, S. Lathrop, D. Lifka, G. D. Peterson, R. Roskies, J. R. Scott, and
als for Energy Efficient Neuromorphic-Computing (Q-MEEN-C), N. Wilkins-Diehr, Comput. Sci. Eng. 16, 62 (2014).

AIP Advances 13, 015216 (2023); doi: 10.1063/9.0000609 13, 015216-4

© Author(s) 2023