Compact Modeling of Magnetic Tunnel Junction

Morgan MADEC, Jean-Baptiste KAMMERER, Fabien PREGALDINY, Luc HEBRARD, Christophe LALLEMENT
Institut d’électronique du solide et des systèmes
UMR 7163 – Centre National de Recherche Scientifique / Université Louis Pasteur
23, rue du Loess, 67037 STRASBOURG CEDEX 02
Email: madec@iness.c-strasbourg.fr

Abstract—A new compact model of a magnetic tunnel junction magnetic electrodes. Both models are self-consistent and are
(MTJ) is presented in this paper. This model is intended to written to be reusable for other spintronic devices.
describe the behavior of a MTJ and to take the magnetic as well
as the non-linear electronic transport phenomena into account. Magnetization and tunneling conductance models are
It should be suitable for circuits simulation and thus, it must be presented in the next two sections. Then, some information
simple (no finite element approach, analytical current versus about the implementation of the model in VHDL-AMS are
voltage characteristic only). For this purpose, some assumptions given. Finally, simulation results are exposed and discussed in
are made. The MTJ model is separated in two entities. The first the two last sections of the paper.
one concerns the magnetization of a ferromagnetic thin film.
The other focuses on the electrical conduction of a MTJ. Both II. DYNAMIC MODEL OF A FERROMAGNETIC FILM
models are implemented and coupled in VHDL-AMS in order to An MTJ consists in two ferromagnetic layers (FM1 and
obtain the compact model of a MTJ, which is parameterized FM2) separated with a thin insulator (Figure 1). The tunneling
with 25 values (19 physical parameters and 6 semi-empirical current through an MTJ depends on the relative orientation of
ones). The first simulations are encouraging. They allow to the magnetization vector in the two ferromagnetic layers. The
retrieve classical results on MTJ and to predict interesting first modeling task thus focuses on a single ferromagnetic
behaviors. layer in its environment (i.e. under an external magnetic field
and close to another ferromagnetic film).
I. INTRODUCTION
The potential application range of spintronic devices, and
especially magnetic tunnel junction (MTJ) is large [1]. MTJ
(
R = f M1 ⋅ M 2 ) Ferromagnetic layer (FM1)
M1
are already used for memories [2], configurable circuits [3][,4]
and magnetic field sensors [5]. For those designs, complex Insulator
simulation processes are carried out. Huge equations sets used
in standard methods are very time-consuming and simulation Ferromagnetic layer (FM2)
are often reduced to small areas or reduced to a small number M2
of components. In opposition, the designers of circuits
including MTJ needs fast simulating models. In addition, the
design under consideration may contain more than one MTJ. Figure 1. A Magnetic tunnel junction in the antiparallel state
For this purpose, compact modeling of MTJ is a very
interesting way to explore. Although models of MTJ have A. Basic Compact Model of a Ferromagnetic Film
already been developed for design [4], to our knowledge, the
model presented in this paper is the first complete MTJ The magnetic behavior of a thin ferromagnetic layer has
compact model including both magnetic and electrical sides. been widely studied [6]. Many commercial, proprietary or
open-source software (e.g. OOMMF, Simulmag) [7] can be
The challenge of compact modeling is to reduce the found to treat such issue. Nevertheless, those simulators often
complex space-and-time-dependant set of equations that consist in a 2D-3D numerical solver which is very time and
describes finely each phenomenon into a set of 1D analytical memory consuming. Such sets of equations is not acceptable
equations which describes by a simple but complete way the for our purpose. To simplify the issue, the ferromagnetic layer
behavior of the device. This process involves some is assumed to be small enough to be considered as mono-
approximations which are noticed and justified in the domain. Under this consideration, the Stoner-Wolfarth
following. The model of the MTJ can be divided into two sub- formalism [8] can be applied and the dynamic of the
models: the magnetization of a single ferromagnetic layer and magnetization vector of the FM layer can be accurately
the conductance through a tunneling barrier composed with described by the Landau-Lifshitz-Gilbert equation [9][10].

978-1-4244-2332-3/08/$25.00 © 2008 IEEE 229

 Kammerer et al. [11] developed an extended expression of III. MODEL OF THE TUNNELING CURRENT IN MTJ
the LLG equation including four terms: the torque associated The second task consists in modeling the electrical
with the Zeeman energy, the one associated with the conduction in the junction which depends on the relative
anisotropy energy inside the layer, the classical Rayleigh orientation of the ferromagnetic layers, the voltage applied to
damping term introduced by Gilbert [10] and a “correction” the junction, the material and the geometry of the spacer. A
torque which is used to lock the magnetization norm. This MTJ conductance is generally characterized by a set of two
model describes accurately the magnetization reversal parameters 𝐺 and 𝐺 , which are respectively the
mechanism but two important effects miss in order to model conductance in the parallel and the anti-parallel state. Two
MTJ: the coupling and the spin transfer between FM layers.
alternative parameters are often used for this purpose: 𝐺̅ , the
B. Magnetization model mean conductance, and the tunnel magnetoresistance rate
The new version is based on Kammerer’s model [11] to (TMR) defined as following: 𝐺̅ = and 𝑇𝑀𝑅 = .
which the related phenomena cited above is added.
A. The principle of magnetoresistance
1) Introduction of Coupling Effects: The first one is The energy band diagram of the MTJ around the spacer is
introduced to take the coupling between the two layers into given on the Figure 2. The difference of electrical behavior
account. There are many coupling paths between two close between parallel and anti-parallel states can be explained by
ferromagnetic layers, each having a different behaviour : the the shift of the energy band between spin-up and spin-down
direct pinhole coupling [12], the exchange coupling [13], the electron due to the internal molecular field when the FM is
Néel orange-peel coupling [14], dipolar coupling [15] etc. magnetized. The conduction through the spacer can be
The characterization of each coupling path is a very hard task separated in two channels, one per spin direction. For the
and the results may depend on materials, technologies and parallel state (P), there is a large among of free electrons both
side and one spin channel is favored, leading to an higher
fabrication processes. In this compact model, the coupling is
conductance.
modeled by a single coupling coefficient which depends on
Energy
the spacer thickness through a decreasing exponential VL
FM1
VL
INSULATOR FM2

function. This model corresponds to the dominant behaviour Δφ VL

for the three first coupling path listed above. Under this Fermi level
CB
φ

condition, the effective field used in the LLG equation [10] CB↑
for FM1 is given by: CB↑
h2
CB
0 CB
CB↓ h1 VB
𝐻 ,
⃗ = 𝐻 ⃗ − 𝑁𝑀⃗ + 𝐽𝑀 ⃗ (1) CB↓
VB
VB

with 𝐽 = 𝐽 exp 2π√2 λM and where 𝐽 is the “non-physical” x

coupling coefficient without insulating barrier, and 𝜆 is the Figure 2. Band diagramof the MTJ around the spacer
characteristic thickness of the coupling function. 𝑁 is the
demagnetizing tensor [16]. B. State of the art
2) Introduction of Spin-Torque-Transfer: The second The calculation of the current through the barrier is a pure
related effect taken into account in the new model of the quantum mechanics problem that can be solved by using the
magnetization is the spin transfer between the two layers. Schrödinger equation and the Landauër-Büttiker formula.
Those equations are too complex to be implemented in a
Slonczewski [17] demonstrated that, when the ballistic
compact model.
transport of spin-polarized electron from one ferromagnetic
layer through a nonmagnetic barrier is strong enough, the Nevertheless, many simple models can be found in the
transversal component of the spin is absorbed by the second literature. The first one, developed in 1975 by Jullière [18],
layer and may be sufficient to reverse its magnetization. The states that the tunneling current in each spin channel is
spin torque transfer (STT) can be introduced as a new torque proportional to the product of the effective tunneling density
of states for the given spin channel for the metal on each side
term in LLG equation. This model remains accurate as long
of the barrier. This model is sufficient to explain qualitatively
as the two FM layers can be considered as ideal spin filters the magneroresistance effect but is not quantitative. In
(the spin of every electron emitted by a layer must be parallel particular, the shape of the potential barrier is not taken into
with the magnetization of the layer). For instance, if a current account whereas many works proved that this parameters
𝐼 → circulate from FM1 to FM2, FM2 is subjected to the influence the electrical behavior [19]. In 1989, Slonczewski
following spin torque: solved the “free electron model” to state two main results [13].
First, the conduction in one channel can be expressed as a
Γ ⃗ = −𝐾𝐼 → 𝑀 ⃗ × 𝑀 ⃗ × 𝑀⃗ (2) function of the electron wave vector inside and outside the
barrier with the following expression.
where 𝐾 is a constant. ′
′
𝐺 =
ℏ
e ′ ′
(3)

230
 where 𝜅 is the wave vector on the evanescent wave in the
insulator are 𝑘 is the wave vector of electron in the spin 𝐺(𝑉, 𝜃) = 𝐺 ∙ 1 − 𝑞𝑉 + (𝑞𝑉) ∙ 1−
direction 𝜎 in the layer FM𝑖. Second, Slonczewski
demonstrated that the variation of the conductance in function cos 𝜃 (8)
of 𝜃, the angle formed by the magnetization direction in FM1
and FM2 is given by:
By integration over V, one finds the following expression
𝐺(𝜃) = 𝐺̅ (1 + 𝑃 𝑃 𝑐𝑜𝑠𝜃) (4) for the 𝐼(𝑉) function which is used in the MTJ model.
where 𝑃 is the polarization of the FM𝑖 layer defined in the ⎡
Jullière model [18]. Slonczewski model gives accurate results ⎢
but is based on a zero-bias assumption, which means that the 𝐼(𝑉, 𝜃) = ∫ 𝐺(𝑢, 𝜃)𝑑𝑢 = ⎢𝑞 −𝑞 /
+
variation of 𝐺 with the voltage applied on the MTJ [20] can’t ⎢
be modeled by the previous equations. ⎣

The G-V dependence in a tunneling junction has been

studied in 1975 by Brinkman which approximated the 𝑉 1+ + / −𝑞 +
observed 𝐺(𝑉) characteristics by a parabolic function [20].
The coefficient of the polynomial depends on the barrier ⎤
height and dissymmetry. /
⎥ (9)
Δ ⎥
𝐺(𝑉) = 𝐺 1− /
𝑒𝑉 + (𝑒𝑉) (5) ⎦
where 𝜑 and Δ𝜑 are respectively the barrier mean height
and dissymmetry as depicted on Figure 2. Brinkman’s IV. RESULTS
equation is only available in the low-voltage range. With this The model of the MTJ has been implemented in VHDL-
expression, TMR should not vary with voltage. This is in AMS. It is composed with 2 electrical terminals, 3 input
conflict with observations [19]. In 2004, Li et al. [21] model quantities (the 3D external magnetic field vector) and 25
permits to explain this effect, but this model is, as usual, too generic parameters: 5 geometrical parameters, 6 physical
complex for compact model. parameters per FM layer, 2 physical parameters for the spacer
C. Compact Model of the Tunneling Current and 6 semi-empirical parameters that can be extract from
characterization measurement. The model has been tested with
The compact model of the tunneling current is based on Advanced MS, the Anacad/Mentor Graphics simulator. Some
these works. First, Li’s equation is solved by using results are given in the following.
Slonczewski’s assumptions and methods in order to calculate
an accurate approximation for 𝐺 and 𝐺 . Second, with (4) First, an example of magnetization reversal on a single
and Jullière’s result [18], one can write: 1µm² area 100 nm thick iron layer over its easy axis (x-axis) is
( )
given (Fig. 3). The reversal is achieved in 200 ps. On the right
𝐺(𝑉, 𝜃) = 𝐺(𝑉) 1 −
( )
cos 𝜃 (6) figure, the precession of the magnetization vector in the y-z
plane can be observed.
For 𝐺̅ (𝑉), Brinkman’s model is used. For 𝑇𝑀𝑅(𝑉), a
simple fit function should be found. Gaussian or sinus x
interpolation used in previous work [22][23] are dismissed
because those functions can’t be integrated. The fit function
used in our model is the following:
𝑇𝑀𝑅(𝑉) = 𝑇𝑀𝑅 (7)

(a) (b)
where 𝑇𝑀𝑅 is the TMR at zero-bias calculated using Li’s
model and 𝑉 is a fit parameter corresponding to the voltage at Figure 3. Example of magnetization reversal on an iron layer with Rayleigh
which the TMR is divided by 2. Again, this equation is only damping coefficient of 0.1. (a) gives the time evolution of the three
components 𝐻 , 𝐻 and 𝐻 whereas (b) gives the precession 𝐻 = 𝑓(𝐻 ) .
available in the low-voltage range. In particular, high-voltage
TMR inversion can’t be modeled by this equation [19][21].
For composite MTJ, dissymmetry appears in the 𝑇𝑀𝑅(𝑉)
characteristic. This is introduced in the model by
differentiating the half-TMR voltage for positive and negative
voltages. Finally, the expression of the conductance is given
by:

Figure 4. Simulated 𝐼 − 𝐻 characteristic for a well-aligned Fe/Al2O3/Co

MTJ (black) and a misaligned one (grey)

231
 The second figure (Fig. 4, black plot) is the simulated tendencies. It is a powerful tool for a first dimensioning of
𝐼 − 𝐻 characteristic for Fe (100 nm) / Al2O3 (2 nm) / Co (100 MTJ-based circuit design.
nm) MTJ. There are two critical magnetic field values. The
first one corresponds to the reversal of softer material (i.e. Fe) REFERENCES
and the second one corresponds to the reversal of the harder [1] S. S. Parkin, “Spintronic materials and devices: past, present and
(i.e. Co). The values of critical magnetic field are equal to the future” in Electron device meeting, pp. 903-906, 2004.
coercive field of both material (i.e. 40 kA/m² and 600 kA/m² [2] J.-B. Kammerer et al., “A two-axis magnetometer using a single
for Co) within about the coupling coefficient which can be magnetic tunnel junction” in IEEE Sensors Journal, vol. 4(3), pp. 313-
321, 2004.
extracted from this measurement. The grey plot corresponds to
[3] S. Tehrani et al., “Magnetoresistive random access memory using
the same simulation than the black one but for a misaligned magnetic tunnel junctions” in Proc. of IEEE, vol. 91(5), 2003.
MTJ: the easy axis of Co and Fe layer are shifted by 10°. An [4] N. Bruchon, G. Cambon, L. Torres and G. Sassatelli, “Magnetic
important asymmetry can be observed. This figure illustrates remanent memory structures for dynamically reconfigurable FPGA”, in
the potential of our compact model, which is predictive. The International Conference on Field Programmable Logic and
last result (Fig. 5) presented in this paper is the variation of Applications, pp. 687-690, 2005.
current density, conductance and TMR in function of the [5] W. Zhao, E. Belhaire, V. Javerliac, C. Chapert and B. Dieny,
voltage applied on the Fe / Al2O3 / Co MTJ. As it is a “Evaluation of a non-volatile FPGA based on MRAM technology”, in
2006 IEEE International Conference on Integrated Circuit Design and
composite MTJ, curves are not symmetric. Technology, article number 1669369, 2006.
[6] H. Krönmuller and M. Fähnle, “Micromagnetism and Microstructures
1,8 1

0,75
Conductance (µS)

1,6 0,5
in Ferromagnetic Solids”, ed. Cambridge University Press, 2003.
Current (µA)

0,25
1,4
-500 -250
0
0 250 500
[7] OOMMF Website: math.nist.gov/oommf.
-0,25
1,2
-0,5 [8] E. Stoner and E. Wohlfarth, “A mechanism of magnetic hysteresis in
-500 -250
1
0 250 500
-0,75

-1
heterogeneous alloys,” IEEE Transactions on Magnetics, vol. 27, no. 4,
Voltage (mV)
5
Voltage (mV)
pp. 3475–3518, 1991.
4 [9] L. D. Landau and E. M. Lifshitz, “Theory of dispension of magnetic
3 permeability in ferromagnetic bodies” is Phys. Z. Sowietunion, vol. 8,
TMR (%)

2 p. 153, 1935.
1
[10] T. L. Gilbert, “A Lagrangian formulation of the gyromagnetic equation
-500 -250
0
0
Voltage (mV)
250 500 of the magnetic field” in Physical Review, vol. 100, p. 1243, 1955.
[11] J. B. Kammerer, L. Hébrard, “Modèle dynamique d’aimantation de
Figure 5. Conductance of the MTJ (top-left), current through the MTJ (top-
couches minces : application à la modélisation compact de Jonction
right) and TMR (bottom) versus the applied voltage. Magnétique à Effet Tunnel” in the Proceedings of TAISA’06, pp. 85-88.
[12] J. Dubowik, B. Szymanski, F. Stobiecki and K. Roll, “Interlayer
All the results presented in this section are qualitatively in exchange coupling across Cu/Ti/Cu spacer layer” in Phys. Stat. Sol.,
accordance with measurements on real MTJ that can be found vol. 196(1), pp. 86-89, 2003.
in literature [17][20][21]. A quantitative validation of the [13] J. C. Slonczewski, “Conductance and exchange coupling of two
model is under consideration. ferromagnets separated by a tunneling barrier” in Phys. Review B, vol.
39(10), PP. 6995-7002, 1989.
V. DISCUSSION AND CONCLUSION [14] J. C. S. Kools, W. Kula, D. Mauri and T. Lin, “Effect of finite magnetic
The MTJ compact model presented in this paper, are based film thickness on Néel coupling in spin valves” in Journal of Applied
Physics, vol. 85(8), pp. 4466-4468, 1999.
on some assumptions (mono-domain ferromagnetic layer, low
voltage, independency of the physical parameters and the [15] M. Hehn et al., “360° domain wall generation in the soft layer of
magnetic tunnel junction” in Applied Physics Letter, vol. 92, article
tunneling current with temperature, etc). These assumptions number 072501, 2008.
are necessary to reduce the simulation time, which is [16] E. du Trémolet de Lacheisserie, D. Gignoux and M. Schlenker,
mandatory to speed up the simulation process during the “Magnetism: Fundamentals”, ed. Springer, 2004.
design of large circuit (MTJ and its signal conditioning, [17] J. C. Slonczewski, “Current-driven excitation of magnetic multilayer”
circuits with many MTJ). The results obtained in simulation in Journal of Magnetism and Magnetic, vol. 159, pp. L1-L7, 1996.
matches very well with experimental results found in [18] M. Jullière, “Tunneling between ferromagnetic film” in Physics Letter,
literature. The model returns the 𝐼 − 𝑉 characteristic of a MTJ vol. 54(A), pp. 225, 1975.
as a function of the external magnetic field. It is parameterized [19] F. Montaigne, M. Hehn and A. Schuhl, “Tunnel barrier parameters and
through 19 values which correspond to physical constants of magnetoresistance in the parabolic band model” in Physics Review B,
vol. 64(14), n°144402, 2001.
the material and to the geometry of the component. The model
can be adjusted through 6 other parameters that should be [20] W. F. Brinkman, R. C. Dynes and J. M. Rowell, “Tunneling
conductance of assymetrical barrier” in Journal of Applied Physics, vol.
extracted in order to characterize the model for a given 41(5), pp. 1915-1921, 1970.
technology. [21] F. F. Li et al., “Bias dependent tunneling in ferromagnetic junctions and
inversion of the tunneling magnetoresistance from a quantum
The validation of the model with one or more set of MTJ, mechanical point of view” in Journal of Applied Physics, vol. 95(11),
which is an important work, is still in progress. In addition, pp. 7243-7é46, 2004.
many “second order” phenomena (e.g. temperature effect) [22] S. Lee, S. Lee, H. Shina and D. Kim, “Advanced HSPICE macromodel
should be added in order to improve the model, especially for for magnetic tunnel junction” in Japanese Journal of Applied Physics,
very thin MTJ for which they become predominant. vol. 44, pp. 2696-2700, 2005.
Nevertheless, it is a good “first order” predictive model which [23] J. A. C. Bland and B. Heinrich, “Ultrathin Magnetic Structures III :
is sufficient to give qualitative results and important Fundamentals of Nanomagnetism”, ed. Springer-Verlag, 2004.

232