Switching dynamics in titanium dioxide

memristive devices
Cite as: J. Appl. Phys. 106, 074508 (2009); https://doi.org/10.1063/1.3236506
Submitted: 01 July 2009 • Accepted: 22 July 2009 • Published Online: 09 October 2009

Matthew D. Pickett, Dmitri B. Strukov, Julien L. Borghetti, et al.

ARTICLES YOU MAY BE INTERESTED IN

High switching endurance in memristive devices

Applied Physics Letters 97, 232102 (2010); https://doi.org/10.1063/1.3524521

Brain-inspired computing with memristors: Challenges in devices, circuits, and systems

Applied Physics Reviews 7, 011308 (2020); https://doi.org/10.1063/1.5124027

A comprehensive review on emerging artificial neuromorphic devices

Applied Physics Reviews 7, 011312 (2020); https://doi.org/10.1063/1.5118217

J. Appl. Phys. 106, 074508 (2009); https://doi.org/10.1063/1.3236506 106, 074508

© 2009 American Institute of Physics.

 JOURNAL OF APPLIED PHYSICS 106, 074508 共2009兲

Switching dynamics in titanium dioxide memristive devices

Matthew D. Pickett, Dmitri B. Strukov, Julien L. Borghetti, J. Joshua Yang,
Gregory S. Snider, Duncan R. Stewart,a兲 and R. Stanley Williamsb兲
Information and Quantum Systems Laboratory, Hewlett-Packard Laboratories, 1501 Page Mill Road, Palo
Alto, California 94304, USA
共Received 1 July 2009; accepted 22 July 2009; published online 9 October 2009兲
Memristive devices are promising components for nanoelectronics with applications in nonvolatile
memory and storage, defect-tolerant circuitry, and neuromorphic computing. Bipolar resistive
switches based on metal oxides such as TiO2 have been identified as memristive devices primarily
based on the “pinched hysteresis loop” that is observed in their current-voltage 共i-v兲 characteristics.
Here we show that the mathematical definition of a memristive device provides the framework for
understanding the physical processes involved in bipolar switching and also yields formulas that can
be used to compute and predict important electrical and dynamical properties of the device. We
applied an electrical characterization and state-evolution procedure in order to capture the switching
dynamics of a device and correlate the response with models for the drift diffusion of ionized
dopants 共vacancies兲 in the oxide film. The analysis revealed a notable property of nonlinear
memristors: the energy required to switch a metal-oxide device decreases exponentially with
increasing applied current. © 2009 American Institute of Physics. 关doi:10.1063/1.3236506兴

I. INTRODUCTION with time under an applied voltage or current. We see that

The reversible nonvolatile resistance switching of thin-
film oxide insulators and semiconductors has been exten-
sively studied and reviewed1–3 over more than 40 years. We
are especially interested in bipolar oxide switches4–6 since
they have been identified7 as being physical examples of
memristors8 and/or memristive devices9 that were first pre-
dicted by Chua in 1971. These devices are of both scientific
interest, since memristors represent a fourth fundamental
passive circuit element, and practical relevance, since Chua
provided the mathematical foundation for their utilization in
various circuit applications.10
A significant challenge to the widespread adoption of
memristive oxide switches by circuit designers has been the
lack of a predictive compact model that can be integrated
into a time domain simulation package like SPICE. The de-
sired circuit element model should abstract, parametrize, and
predict the behavior of the device under an arbitrary voltage
and current bias, at least under some restricted domain of
state and/or excitation. Such a model has not been available
for bipolar oxide switches despite recent progress in under-
standing their physical switching mechanism in the context
of coupled electronic and ionic conductors.2,3,11–14
In this paper, we present a nonlinear memristive model
of bipolar switching derived from the experimental results of
a dynamical testing protocol applied to a set of Pt– TiO2 – Pt
crosspoint devices. We analyze both the static electronic con-
duction behavior and the switching dynamics that arise from
ionic motion in the device. Our analysis quantitatively estab-
lishes that electronic conduction in these devices is domi-
nated by an effective tunneling barrier width w that varies
a兲
Present address: National Research Council of Canada, 100 Sussex Dr.,
Ottawa, Ontario K1A 0R6, Canada.
b兲
Electronic mail: stan.williams@hp.com.
the dynamical behavior for off and on switching is highly
nonlinear and asymmetric, and these observations can be un-
derstood in terms of an exponential dependence of the drift
velocity of ionized dopants on the applied current or
voltage15 and the competing or cooperative behavior of ionic
drift and diffusion, depending on the switching voltage
polarity.14
II. DEVICE DESCRIPTION AND CHARACTERISTICS
For our experiments, we fabricated single crosspoint
metal-oxide-metal switches on thermally oxidized silicon
substrates with the following vertical structure:
Si/ SiO2共100兲 / Ti共5兲 / Pt共15兲 / TiO2共50兲 / Pt共30兲, where the
numbers in parentheses are the layer thicknesses in nanom-
eters. TiO2 was grown by sputter deposition from a rutile
target onto a substrate held at 250 °C. X-ray diffraction
共XRD兲, transmission electron microscopy 共TEM兲, and Ruth-
erford backscattering spectrometry 共RBS兲 analyses on sister
films of TiO2 grown with the same process showed that the
layers are amorphous or nanocrystalline 共TEM/XRD兲 and the
as-grown stoichiometric ratio 共RBS兲 of oxygen to titanium is
2.00 ⫾ 0.01.16
In order to minimize the series resistance and the para-
sitic resistance-capacitance 共RC兲 time constant of the wires
and contact pads for time-sampled measurements, we present
here the results from a large-area 5 ⫻ 5 ␮m2 crosspoint de-
vice. We performed similar experiments and analyses on a
total of five devices with four different areas on five inde-
pendently fabricated samples, including 50⫻ 50 nm2 devices
defined by imprint lithography 共see supplemental online
materials17兲. The dynamical switching behavior was effec-
tively insensitive to the device size and can ultimately be

0021-8979/2009/106共7兲/074508/6/$25.00 106, 074508-1 © 2009 American Institute of Physics

074508-2 Pickett et al. J. Appl. Phys. 106, 074508 共2009兲

S Pt w III. MEMRISTIVE MODEL

(a)
In order to track the changing tunnel gap thickness,
TiO2 TiO2-x Rs V
which defines the state of the device, throughout the switch-
A Pt ing process and obtain quantitative dynamical information,
we have performed a quasistatic state-test protocol summa-
(b) rized schematically in Fig. 1 and described in more detail in
-4
10
the Methods section below. Using a two-step iterative pro-
i (A)

10
-6
cess, we first apply a low voltage amplitude 共too low to
ON OFF
-8
result in observable switching or hysteresis兲 triangular exter-
10
nal voltage sweep to interrogate the state of the device as
-1 0 1
v (V) defined by its i-v characteristic, and then apply a constant
1.0 voltage time-sampled pulse to change the state of the device
(c)
共Fig. 1兲. This process is iterated until the junction has fully
v (V)

0.5 switched or the maximum test time has been reached. The
0.0 state data conformed to the framework of a canonical mem-
ristive system,9 which comprises two coupled expressions of
40 the form
i (µA)

20 i = G共w, v兲v 共1兲

0 and
-1 0 1 2 3
t (s) ẇ = f共w,i兲. 共2兲

FIG. 1. 共Color兲 Device schematic and characterization protocol. 共a兲 Sche-
matic of the device cross section after electroforming and the four-wire
time-sampled current-voltage test setup used for the state-test protocol. For
the test setup S, A, and V represent a voltage source, ammeter, and voltme-
ter, respectively. In the device schematic w and Rs represent the tunneling
barrier width and the electroformed channel resistance, respectively. 共b兲 Ex-
ample switching i-v curve for the device studied in this work. Positive
polarity turns the device state to off, while negative polarity turns the device
on. The blue curve corresponds to the fit for a conducting channel series
resistance of Rs = 215⫾ 6 ⍀. 共c兲 Example time series data for the internal
voltage and current from the state-test protocol, with the interrogation v共t兲
and i共t兲 plots shown in blue and with the state-change pulses in red. The
pulse width increases exponentially in order to span a wide range of total
time at voltage.
described as an ⬃1 nm length modulation of a 30–300 nm
diameter electroformed conduction channel embedded in a
thin semiconducting film.16
We electroformed the TiO2 devices by applying a volt-
age sweep from 0 to ⫹6 V over 5 ms. Electroforming causes
localized heating and oxygen vacancy formation in these
materials16 and produced an irreversible decrease in the re-
sistance from the approximately gigaohms as-fabricated state
to the approximately kilohm/megaohm on-off-switching re-
gime. In the devices studied here, the result of electroform-
ing was a conducting channel with a resistance in hundreds
of ohms that shunts most of the oxide film.18 Electroforming
also produced a remnant tunnel gap between the conducting
channel and the opposite Pt electrode, the width w of which
can be modulated by applying a voltage across the device
and inducing the motion of ionized defects.7,14,15,18 The sche-
matic diagram and the switching i-v characteristic in Fig. 1
summarize this picture. The series resistance intrinsic to the
channel can be obtained directly from the switching i-v by
fitting the differential resistance at high bias.18 For the device
reported in detail here, the value obtained was Rs
= 215⫾ 6 ⍀.
Equation 共1兲 is the static transport relation in which G is a
generalized conductance that is a function of the internal
voltage v and the state variable w. Equation 共2兲 is the dy-
namical or rate equation, in which ẇ is the time derivative of
the state variable and f expresses the functional dependence
of ẇ on w and i. This formulation provides the mathematical
framework for revealing the device physics and using the
memristive devices in a circuit. The unusual mixture of volt-
age and current dependent functions in Eqs. 共1兲 and 共2兲 re-
spectively, was utilized to ensure that the compact model
presented below would contain analytical expressions. Given
the form of Eq. 共2兲, the instantaneous state of the system
must depend on the history of the current flow through the
device.
IV. STATIC TRANSPORT: TUNNELING GAP WIDTH w
AS THE STATE VARIABLE
We successfully reduced the i-v characteristic for each
state of the device to a single variable by fitting to the
equivalent circuit model of Fig. 1: an Ohmic resistor in series
with an electron tunnel barrier. We used the Simmons i-v
expression for a rectangular barrier with image forces19 since
it is the simplest model that fits our data well with physically
reasonable parameters and is consistent with Eq. 共1兲. The full
set of expressions that embody the equivalent circuit are in-
cluded in the supplementary information.17 Remarkably, a
single dominant state variable emerged from this analysis:
the Simmons tunnel barrier width, which we label w. We also
considered the tunneling barrier height ␸0 as a plausible
single state variable, but could not accurately reproduce the
experimental data by varying only the barrier height during
the fitting procedure. Thus, the switching effect is primarily
due to an effective tunneling distance modulation.
Figure 2 shows the current-voltage fits with barrier
width as the state variable using a regression procedure de-
scribed in more detail in Sec. VII. Invariant device param-
074508-3 Pickett et al. J. Appl. Phys. 106, 074508 共2009兲

10
-3 t=0 s 4.7ms 970 ms channel width in similar devices using the technique of
110 µs
790 µs
28 ms
160 ms
5.7 s
33 s pressure-modulated conductance mapping.16 This channel
10
-4
geometry yields a resistivity estimate of ⬃4 ⫻ 10−3 ⍀ cm,
i (A) which is significantly more resistive than metallic Pt or Ti
10
-5
but reasonable for reduced TiO2.20,21
t
We note the tunneling effective mass was assumed to be
-6
10 equal to the free electron mass because we do not have an
Model
independent means to measure this value. This could pro-
-1 -0.1 0.1 1
duce a systematic error in the above parameters and the re-
v (V)
ported barrier widths. For this reason, the numerical results
FIG. 2. 共Color兲 Current-voltage characteristics for a selection of times. Nine should be considered as effective values, although they none-
i-v tests taken during the 4.5 V off-switching state test on the 5 ⫻ 5 ␮m2 theless provide a good idea of the measurement precision.
device are shown. The lines are best fits to the series resistor plus Simmons Invariant parameter values were obtained from a first regres-
tunnel barrier equivalent circuit. The data points at high positive voltage are
from the state-change pulse immediately following the respective state-test
sion analysis, e.g., at a particular voltage pulse amplitude,
data and were not included in the fitting procedure, but the extrapolation of after which only w was allowed to vary in subsequent fits.
the tunneling equation to these measured values indicates that the conduc- This procedure enabled us to track the width of the tunnel
tion model works well even into the switching regime. The legend lists the barrier as a function of time and the applied voltage.
total time the device was under the applied voltage in seconds, after which
the i-v data were collected.
V. DYNAMIC BEHAVIOR: EXPONENTIAL
DEPENDENCE OF SWITCHING TIME ON THE DEVICE
eters that resulted were barrier height ␸0 = 0.95⫾ 0.03 eV, CURRENT
dielectric constant ␬ ⫽ 5 ⫾ 1, and channel area A
= 10 000⫾ 2500 nm2. The series channel resistance, Rs After reducing each state i-v to a value of w, we exam-
= 215 ⍀, was also kept fixed during the fitting procedure. ined the voltage-dependent dynamics of the effective barrier
The barrier width w varied from 1.1 to 1.9 nm, with typical width during switching. Figure 3共a兲 shows the time evolution
fitting uncertainty of ⫾0.1 nm. Typical relative errors in the of w during six off-switching tests with external voltages
current and voltage measurements were 1%, but for the sake ranging from 3.0 to 5.5 V. Figure 3共c兲 shows five on-
of clarity, error bars are omitted from Figs. 2 and 3. Assum- switching tests with voltages of ⫺1.25 and ⫺1.4 V. We
ing a cylindrical geometry for the conduction channel, the chose to study only two biases for on switching because the
best fit area A corresponds to a diameter of 113 ⫾ 7 nm, switching speed was much faster and is more sensitive to the
which is consistent with previous direct measurements of the magnitude of the applied bias and we had a RC-limited time
resolution 共⬃1 µs兲 in our test setup. On-switching repeatabil-
(a) (c)
ity was therefore demonstrated via multiple state tests at the
1.8 -1.4 V -1.25 V
Vext 1.8 -1.4 V -1.25 V
same voltage.
1.6
-1.25 V The device state evolved in a continuous but highly non-
w (nm)

w (nm)

1.6 linear manner between the limiting on and off states, and the
1.4 switching speed was strongly dependent on the current mag-
3.0 V 4.5 V Vext
1.2 3.5 V 5.0 V
1.4
nitude and polarity. In order to determine an analytical ex-
4.0 V 5.5 V
pression of the dynamical state evolution 关Eq. 共2兲兴 for off
-5 -3 -1 1 -5 -4 -3 -2 -1 0
10 10
t (s)
10 10 10 10 10 10
t (s)
10 10 and on switching, we again applied a regression technique
(b)
4
(d) described in Sec. VII. The end result was a set of consistent
10 t t
3
10 parameters for the full range of applied voltages studied,
-dw/dt (nm/s)
dw/dt (nm/s)

2
10
1 with the following analytical expressions:
10
0
10 off switching 共i ⬎ 0兲:

冉 冊 冋 冉 冊 册
-1
3.0 V 4.5 V 10
w − aoff 兩i兩
-2 -1.25 V
10
3.5 V 5.0 V -1.40 V -1.25 V i w
10
-4 4.0 V 5.5 V
10
-3 -1.40 V -1.25 V ẇ = f off sinh exp − exp − − 共3兲
1.2 1.4 1.6 1.8 1.4 1.5 1.6 1.7 1.8
ioff wc b wc
w (nm) w (nm)
and the fitting parameters f off = 3.5⫾ 1 ␮m / s, ioff
FIG. 3. 共Color兲 Dynamical behavior of the tunnel barrier width w. The state = 115⫾ 4 ␮A, aoff = 1.20⫾ 0.02 nm, b = 500⫾ 70 ␮A, and
variable w evolves as a function of time for different applied voltages for a wc = 107⫾ 4 pm;
series of 共a兲 off-switching and 共c兲 on-switching state tests on the same de-
on switching 共i ⬍ 0兲:

冉 冊 冋 冉 冊 册
vice. The legends indicate the applied external voltage. The lines are the
numerical solution to the respective switching differential equations de-
scribed in the text. 共b兲 and 共d兲 show the numerical derivative ẇ of the data
i w − aon 兩i兩 w
ẇ = f on sinh exp − exp − − − 共4兲
in 共a兲 and 共c兲 plotted as a function of w for the different applied voltages. ion wc b wc
The lines are calculated from the differential equations using the measured
values of w and i at each point in time. The irregularity of the calculated ẇ and the fitting parameters f on = 40⫾ 10 ␮m / s, ion = 8.9⫾
vs. w lines in the on-switching plots is caused by the changes in the current 0.3 ␮A, aon = 1.80⫾ 0.01 nm, b = 500⫾ 90 ␮A, and wc
that accompany the change in state 共ẇ is a function of two variables, w and
= 107⫾ 3 pm. These nonlinear functions of i and w were
i, and both are changing兲. The derivative of the state variable ẇ is inter-
preted as the speed of the oxygen vacancy front as the applied voltage determined through a combination of physical insight from
pushes it away from or attracts it toward the top electrode. theoretical analyses11,18 that involved the numerical solution
074508-4 Pickett et al. J. Appl. Phys. 106, 074508 共2009兲

1
10
-1
VI. DISCUSSION AND CONCLUSION
10 ON OFF -3
-1 10
10 These results confirm the picture that is emerging of the

tswitch (s)
10
10
-3
10
-5 10
-7 10
10 Switch time
Switch energy
10
-9 10
1 2 3
|i| (mA)
FIG. 4. 共Color兲 Calculations of switching time and energy using the mem-
ristive model. Constant-current magnitude switching times and energies are
presented for on- and off-switching events. The on and off states were cho-
sen as won = 1.2 nm and woff = 1.8 nm to reflect a resistance switching ratio of
500. The plots were evaluated by numerical integration of Eqs. 共5兲 and 共6兲.
These results demonstrate the asymmetry between the on and off switching
behavior and show that the switching energy decreases exponentially with
the applied current because of the highly nonlinear switching dynamics.
of coupled differential equations and trial-and-error modifi-
cations.
Figures 3共a兲 and 3共c兲 show the resulting time series fits
as solid curves, indicating that the phenomenological dy-
namical expressions represented the data accurately and thus
satisfied our desire for a compact predictive model. Figures
3共b兲 and 3共d兲 show the numerical time derivatives of the
time series plotted against the instantaneous values of the
barrier width and demonstrate the agreement between the
form of the model and the data set.
The memristive model is a predictive formulation for
gaining intuition into these strongly nonlinear dynamical sys-
tems. As examples, two quantities of interest, the time and
the energy required to switch a device between two arbitrary
states, can be computed from Eqs. 共3兲 and 共4兲 for times that
were not experimentally accessible. Figure 4 shows the
switching time and energy as a function of constant applied
current, calculated by numerical integration of the following
formal expressions between the limits won = 1.2 nm and woff
= 1.8 nm for both on and off switching:
Eswitch (J)
-5
structure of an electroformed metal-oxide memristive device,
-7
as illustrated in Fig. 1. The electroforming process created a
-9 localized conducting channel that extended most of the way
-11 across the 50 nm titanium dioxide film, leaving an ⬃2 nm
wide insulating gap. The state variable of the device w is the
effective width of the tunneling gap, and the electrical cur-
rent transport process is limited primarily by tunneling
through this gap, represented here by the Simmons equation
in series with the Ohmic conduction channel. Switching be-
tween selected off and on states with resistance ratios of
approximately 500 is accomplished by changing the width of
the tunneling gap by less than 1 nm. A completely indepen-
dent analysis of similar devices by the technique of pressure-
modulated conductance microscopy22 yielded similar results
in terms of the subnanometer change in the width of a tun-
neling gap during switching.
The width of the gap is changed by an applied voltage or
current, with a positive bias applied to the top electrode pic-
tured in Fig. 1, leading to an increase in the state variable w
and a corresponding exponential increase in the resistance of
the device. A negative bias leads to a decrease in w and a
corresponding exponential decrease in resistance. This bipo-
lar behavior is in agreement with the identification of the
mobile dopants in the gap being positively charged oxygen
vacancies.21
The switching behavior is complex but it is represented
well by the dynamical Eqs. 共3兲 and 共4兲 for the derivative
of w as a function of w and i as shown in parts 共b兲 and 共d兲
of Fig. 3, respectively. The physical basis of the phenomeno-
logical form can be understood qualitatively. The
sinh关i / i0兴exp关−w / wc兴 dependence of the switching rate may
have two contributions: a nonlinear drift at high electric
fields15,23 and local Joule heating of the junction that speeds
up the thermally activated drift of oxygen vacancies.11,18 A
similar exponential dependence of switching rate on the ap-

⌬t = 冕 dw
ẇ
共5兲
and
E= 冕 iv共w兲dw
ẇ
. 共6兲
The validity of the model shown in Fig. 4 for switching
times much shorter and longer than those measured explic-
itly during the state tests has been examined by switching the
devices with short pulses and confirming the stability of the
off state after it had been at rest for one year.17 The series
resistances of the wires and conducting channel were not
included in the calculation of the switching energy since it is,
in principle, possible to effectively eliminate them with short
wires in an integrated circuit. The switching time and energy
decrease exponentially with increasing applied current. Ad-
ditionally, the asymmetry between off and on switching is
clearly observed.
plied voltage has been reported by Tamura et al.24 in cation
migration based sulfide bipolar switches, which they attrib-
uted to device self-heating. Both effects produce similar be-
havior and are likely present simultaneously; the nonlinearity
of either effect could account for the observed storage to
switch time ratio of greater than 1013 共see supplementary
materials17兲.
The absolute rates of off- and on switching were ob-
served to be dramatically different, with the off switching
several of orders of magnitude slower than on switching for
an equivalent applied voltage. This behavior was previously
attributed to the interaction of diffusion and drift in a net
internal electric field.14 For a positive bias applied to the top
electrode in Fig. 1, positively charged oxygen vacancies are
repelled toward the conducting channel, and thus the effec-
tive tunneling gap w increases. This increases the concentra-
tion of the vacancies near the channel and results in both an
increased vacancy diffusion current acting in the opposite
direction to the drift by Fick’s law and also an internal elec-
tric field opposite to the applied field, which slows down the
vacancy drift. Eventually, the total drift velocity goes to zero,
074508-5 Pickett et al.
and there is a bias-dependent maximum value that w can
attain.14 However, when a negative bias is applied to the top
electrode in Fig. 1, the externally applied field, the internal
field of the concentrated vacancies, and the diffusion all act
in the same direction, thus dramatically speeding up on
switching compared to off switching for the same applied
voltage.
We have introduced and utilized a test protocol that ex-
plicitly maps the time evolution of state to analyze the be-
havior of a dynamical electronic device. We have determined
a compact memristive model for an electroformed TiO2 bi-
polar switch that utilizes the Simmons tunneling equation for
the static transport expression, the width of the tunnel gap in
the device as the state variable, and a phenomenological dif-
ferential equation for the dynamical expression. The model
provides a physical picture of the transport, imparts signifi-
cant insight into the device dynamics, and enables the
switching properties to be predicted over a wide range of
operating currents and times.
VII. METHODS
A. Electrical state-test protocol
See Fig. 1 for a schematic of the test setup and an ex-
ample of the voltage and current waveforms measured dur-
ing the state-test protocol. In order to study the switching
dynamics over six orders of magnitude in time, we fixed the
amplitude of the applied voltage during each pulse but expo-
nentially increased the pulse widths from one step to the
next. This protocol required that the state be invariant when
the device was not under a voltage pulse intended to change
the state, i.e., it must be nonvolatile and the applied test
waveform must be nonperturbative. This was a good as-
sumption for our devices and our measurement protocol,
since in our experience the devices stay at an arbitrary state
without changing for at least a year in air at room tempera-
ture and we did not observe any hysteresis in the state-
interrogation i-v characteristics.
We examined the applied voltage dependence of off-
switching dynamics for the device by applying a set of six
state tests with pulse heights ranging from 3.0 to 5.5 V. Ad-
ditionally, we examined the applied voltage dependence of
the on-switching behavior at ⫺1.25 and ⫺1.4 V. Each test
had a total of 40 voltage pulses, which gave a total time
under the applied voltage of 33 s. After each application of
the state test, we reset the device to its initial condition with
a quasi-dc triangular voltage sweep of the opposite polarity.
Although we held the applied or external voltage fixed
during the state tests, there was an appreciable voltage
dropped in the 2 k⍀ electrodes. The actual or internal volt-
age seen by the device was lower than the applied voltage
and evolved continuously throughout the switching process
because of the voltage divider it formed with the series re-
sistance of the electrodes. The current seen by the device
during a test, i, also changed throughout the switching pro-
cess. The electrode series resistance and the nonlinearity of
the device tunneling conductivity yield a complicated rela-
tionship between the current and the external and internal
voltages throughout the switching process. This behavior
J. Appl. Phys. 106, 074508 共2009兲
highlights the necessity of a four-wire time-sampled mea-
surement scheme to completely characterize the state evolu-
tion during memristive switching. All of the time series for
the state data discussed in this paper are included as supple-
mental information to demonstrate that the internal voltages
and the currents changed by less than a factor of 2 during all
of the fixed external voltage state steps. Although the exter-
nal voltages were held constant, it was the currents through
the devices that changed the least during the state tests, and
thus we expressed the dynamical equations in terms of cur-
rent.
B. Regression
For the i-v fitting we implemented a nested loop regres-
sion procedure based on ODRPACK.25 In the outer loop, the
parameters A, ␸0, and ␬ were optimized for the entire data
set simultaneously. During each step of the outer loop, an
inner nested regression was performed to determine w indi-
vidually for each of the 40 i-v characteristics. For the dy-
namics fitting, at each step of the outer regression, we nu-
merically solved Eqs. 共3兲 and 共4兲 using the Runge–Kutta
method as implemented in IGOR PRO software for a single w
versus time series. At each time step in the numerical inte-
gration, we explicitly used the experimentally measured cur-
rent from the four-point probe data for that state-test to ac-
count for the fact that the current changes during the
switching process.
ACKNOWLEDGMENTS
We thank Philip J. Kuekes, Gilberto Medeiros-Ribeiro,
R. Ramesh, and Warren Robinett for insightful discussions.
This research was supported by the SyNAPSE program of
the Defense Advanced Research Projects Agency under Con-
tract No. HR0011-09-3-0001.
1
G. Dearnaley, A. M. Stoneham, and D. V. Morgan, Rep. Prog. Phys. 33,
1129 共1970兲.
2
R. Waser and M. Aono, Nature Mater. 6, 833 共2007兲.
3
R. Waser, R. Dittmann, G. Staikov, and K. Szot, Adv. Mater. 共Weinheim,
Ger.兲 21, 2632 共2009兲.
4
K. Tsunoda, Y. Fukuzumi, J. R. Jameson, Z. Wang, P. B. Griffin, and Y.
Nishi, Appl. Phys. Lett. 90, 113501 共2007兲.
5
D. Jeong, H. Schroeder, and R. Waser, Electrochem. Solid-State Lett. 10,
G51 共2007兲.
6
A. Sawa, T. Fujii, M. Kawasaki, and Y. Tokura, Appl. Phys. Lett. 88,
232112 共2006兲.
7
D. B. Strukov, G. S. Snider, D. R. Stewart, and R. S. Williams, Nature
共London兲 453, 80 共2008兲.
8
L. O. Chua, IEEE Trans. Circuit Theory 18, 507 共1971兲.
9
L. O. Chua and S. Kang, Proc. IEEE 64, 209 共1976兲.
10
L. O. Chua, IEEE Trans. Circuits Syst. 27, 1014 共1980兲.
11
M. Janousch, G. I. Meijer, U. Staub, B. Delley, S. F. Karg, and B. P.
Andreasson, Adv. Mater. 共Weinheim, Ger.兲 19, 2232 共2007兲.
12
R. Meyer, L. Schloss, J. Brewer, R. Lambertson, W. Kinney, J. Sanchez,
and D. Rinerson, Proceedings of the Ninth Annual Non-Volatile Memory
Technology Symposium, 2008 共unpublished兲.
13
D. S. Jeong, H. Schroeder, and R. Waser, Phys. Rev. B 79, 195317 共2009兲.
14
D. B. Strukov, J. L. Borghetti, and R. S. Williams, Small 5, 1058 共2009兲.
15
D. B. Strukov and R. S. Williams, Appl. Phys. A: Mater. Sci. Process. 94,
515 共2009兲.
16
J. J. Yang, F. Miao, M. D. Pickett, D. A. A. Ohlberg, D. R. Stewart, C. N.
Lau, and R. S. Williams, Nanotechnology 20, 215201 共2009兲.
17
See EPAPS supplementary material at http://dx.doi.org/10.1063/
074508-6 Pickett et al. J. Appl. Phys. 106, 074508 共2009兲

22
1.3236506 for details of the equations used for the tunneling model, ad- F. Miao, J. J. Yang, D. R. Stewart, R. S. Williams, and C. N. Lau, Appl.
ditional experimental data, and data supporting short and long time ex- Phys. Lett. Appl. Phys. Lett. 95, 113503 共2009兲.
trapolations of the dynamical model. 23
N. Mott and R. W. Gurney, Electronic Processes in Ionic Crystals 共Dover,
18
J. L. Borghetti, D. B. Strukov, M. D. Pickett, J. J. Yang, and R. S. Will- New York, 1964兲.
iams, J. Appl. Phys. 共submitted兲. 24
19 T. Tamura, T. Hasegawa, K. Terabe, T. Nakayama, T. Sakamoto, H. Su-
J. Simmons, J. Appl. Phys. 34, 1793 共1963兲.
20 namura, H. Kawaura, S. Hosaka, and M. Aono, J. Phys.: Conf. Ser. 61,
H. Tang, K. Prasad, R. Sanjinès, P. E. Schmid, and F. Lévy, J. Appl. Phys.
75, 2042 共1994兲. 1157 共2007兲.
25
21
J. J. Yang, M. D. Pickett, X. Li, D. A. A. Ohlberg, D. R. Stewart, and R. P. Boggs, J. Donaldson, R. Byrd, and R. Schnabel, ACM Trans. Math.
S. Williams, Nat. Nanotechnol. 3, 429 共2008兲. Softw. 15, 348 共1989兲.