Journal of

Materials Chemistry C
View Article Online
PAPER View Journal
Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

Substrate dependent resistive switching in

amorphous-HfOx memristors: an experimental
Cite this: DOI: 10.1039/c9tc06736a
and computational investigation†
Pradip Basnet, *a Darshan G Pahinkar,b Matthew P. West,a
Christopher J. Perini, a Samuel Grahamb and Eric M. Vogela

While two-terminal HfOx (x o 2) memristor devices have been studied for ion transport and current evolution,
there have been limited reports on the effect of the long-range thermal environment on their performance. In
this work, amorphous-HfOx based memristor devices on two different substrates, microscopic glass (B1 mm)
and thin SiO2 (280 nm)/Si, with different thermal conductivities in the range from 1.2 to 138 W m1 K1 were
fabricated. Devices on glass substrates exhibit lower reset voltage, wider memory window and, in turn, a
higher performance window. In addition, the devices on glass show better endurance than the devices on the
SiO2/Si substrate. These devices also show non-volatile multi-level resistances at relatively low operating
Received 9th December 2019, voltages which is critical for neuromorphic computing applications. A multiphysics COMSOL computational
Accepted 9th March 2020 model is presented that describes the transport of heat, ions and electrons in these structures. The combined
DOI: 10.1039/c9tc06736a experimental and COMSOL simulation results indicate that the long-range thermal environment can have a
significant impact on the operation of HfOx-based memristors and that substrates with low thermal
rsc.li/materials-c conductivity can enhance switching performance.

Introduction its promising physical and chemical properties, such as good

scalability (o10 nm), fast switching speed (Bns), and fair
Metal oxide (MO) based resistive random access memory switching endurance (Z1010 cycles).2,8,9 Oxide-based VCMs
(ReRAM) devices have been drawing a lot of attention recently for are one type of memristive device, where changes in the
their various potential applications, including neuromorphic resistance or conductance can be induced by redox reactions
computing.1–3 Many researchers have demonstrated that a hybrid in the oxide within a filament or at the oxide–electrode inter-
system, composed of complementary metal oxide semiconductor faces under DC bias voltage or a given voltage pulse.10,11 HfOx is
(CMOS) neurons and memristor synapses, can support important both chemically and physically stable and also compatible with
neuronal functions such as spike timing dependent plasticity CMOS technology.9 Amorphous (a)-HfOx offers a few additional
(STDP).4,5 Wang et al. have reported that a diffusive memristor advantages, such as high uniformity over large area, low
and its dynamics can accomplish a direct emulation of both temperature synthesis and controlled electrical properties.12
short- and long-term plasticity of biological synapses, while other To date, both amorphous and poly-crystalline stoichiometric
researchers have established MO based synapses at the device and sub-stoichiometric (oxygen deficient) HfOx based memristors
level.6,7 have been studied experimentally and computationally,12–18 and
Among the typical binary transition MOs (e.g. TiO2, TaO2, more importantly, stand-alone synaptic devices have been
ZrO2 and HfO2), HfOx (x o 2; sub-stoichiometric) is very demonstrated.19 While there has been significant prior research
popular as a valence change memory (VCM) material due to related to digital switching in filamentary adaptive HfO2, there
has been a very limited fundamental understanding related to the
a
School of Materials Science and Engineering, Georgia Institute of Technology,
primary factors controlling analog resistance change.3,20 The
Atlanta, Georgia 30332, USA. E-mail: [email protected] analog synaptic properties of HfOx such as multilevel set and
b
Electronics Manufacturing and Reliability Laboratory, Department of Mechanical reset states need further investigation. It is worth noting that
Engineering, Georgia Institute of Technology, Atlanta, Georgia 30332, USA biological synapses show analog behavior with multilevel synaptic
† Electronic supplementary information (ESI) available: Substrate dependent
weight changes, and analog conductance (or resistance) change
conductive filament (CF) formation; the CF/s stabilization; the current–voltage
(I–V) curves of stabilized CF/s; substrate dependent switching with the stable CF/s
of two-terminal MOs based memristor devices is quite similar.3
at different reset voltages; and substrate dependent power consumption. See DOI: Recent research efforts along these lines have made a lot of
10.1039/c9tc06736a progress. For example, Long et al.8 have observed the sharp and

This journal is © The Royal Society of Chemistry 2020 J. Mater. Chem. C

 View Article Online

Paper Journal of Materials Chemistry C

gradual reset process in unipolar switching in HfO2, while a migration at the metal–oxide interfaces.14,17,30,31 Syu et al. showed
compliance-free, digital set, and gradual-reset has been reported that the high-speed resistive switching (RS) behaviors of a
in TaOx by Abbas et al.10 memristive device is due to only a few atoms involved in the redox
Forming a conductive filament (CF) in the MOs involves a reactions and mixed ionic–electronic transport at the interface,
creation of oxygen-deficient (or metal rich) region that requires which makes experimental study even more complicated.29
breaking metal–oxygen bonds followed by migration of oxygen Recently, Kim et al. studied TaOx device behavior using three
anions and/or vacancies; both of these processes are controlled different TEs, namely: Pd, Ru, and W, with thermal conductivity
by the local temperature field and the applied electric field.1,3,21 values in the range of 71.8 to 173 W m1 K1, and reported the
Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

As a result, a ‘‘filamentary based’’ memristor operates at effect of the TBMs on overall resistive switching.26 However, the use
extremely harsh conditions of electric field up to 10 MV cm1 of different TEs to study the effect of TBMs can be misleading as
and current density B1 mA cm2 or above.13 It is well accepted other properties of the top electrode, like its work function, oxygen
that bipolar device behavior is controlled by the motion of affinity etc., can heavily affect the RS. Although TiO2 and HfO2
negatively charged oxygen ions (or the positively charged oxygen based memristors have been fabricated on flexible (e.g. plastic)
vacancies) through a conducting filament (CF).22,23 However, substrates,12,32 a systematic study of the impact of the thermal
very little is known about how the thermal boundary materials conductivity of the substrate has not been performed.
(TBMs) affect the oxygen ions’ drift, diffusion, and CF evolution In this work, the performance of 5 nm sub-stoichiometric
process.24–26 The main difficulties in understanding the device a-HfOx (x E 1.8) based memristor devices, fabricated on either
behavior is that the filament has been measured to be on order thin SiO2 (280 nm)/Si or microscopic glass (B1 mm) substrates,
of a nanometer in diameter and re-oxidation of the filament tip are compared and reported. The thin SiO2 (280 nm)/Si substrate
only occurs about a nanometer in distance from the oxygen has an effective thermal conductivity of 138 W m1 K1,
reservoir near the top electrode (TE) as shown in Fig. 1 whereas the glass substrate has a thermal conductivity of
(top).16,17,26–28 Local temperature rises as high as 1000 K have 1.2 W m1 K1. Experimental results for electroforming,
been reported, which depends on the thermal conductivities of digital (abrupt) set, analog (gradual) reset and endurance were
the oxide layer, filament, electrodes, and substrate.14,25 Since compared. A multiphysics COMSOL model that simulates the
both the drift and diffusion of the oxygen ions and/or vacancies simultaneous drift, diffusion and thermophoresis of oxygen
can occur simultaneously, it is very difficult to quantify the vacancies was used to estimate the local and temporal tem-
digital or analog switching behavior that can occur on the perature profile and its effect on the overall device performance.
nano-second timescale.25,29 A variety of approaches, including The comparison of current–voltage (I–V) curves from the COMSOL
computational and experimental methods, have been used to model and those from the experiments provide the validity of the
explain fundamental processes association with vacancy/ion chosen approach in these models. Our combined experimental and

Fig. 1 (a) Schematic of the as-prepared Au (BE)/HfOx/Ti/Au (TE) devices, 2D views: (left) top- and (right) cross-sectional views. (b) A typical electrical
characterization set up and the process of filament formation: (left to right) substrates thermal properties, device layers, and nanoscale filament formation
in HfOx. Arrows showing the heat transfer directions, namely: axial q(z) and radial q(r), from the filament to the surroundings. Note: drawings are not to
scale.

J. Mater. Chem. C This journal is © The Royal Society of Chemistry 2020

 View Article Online

Journal of Materials Chemistry C Paper

modeling findings provide critical insights into the impact of the (B4.0 to 4.6 V in this example), the current fluctuates. This may
long-range thermal environment on memristor operation. indicate that there is a competition of ion–electron migration due
to thermal and electrical fields resulting in an increase or
decrease in the filament number or size. This step plays a critical
Results and discussion role in determining the overall resistance or size/number of the
filament(s). Statistical result shows that the forming voltage (Vf)
Prior to characterizing the resistive switching behavior, it is for both thin SiO2/Si and glass substrates is approximately the
Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

important to first demonstrate that the substrates did not impact same as shown in Fig. 3(b). However, the filament resistance after
the intrinsic structure and behavior of the materials and devices. B3 V is substrate dependent (see ESI†) and it could be attributed
Fig. 2(a) and (b) show the capacitance and initial resistance of the to the local heating near the beginning of filament formation
devices, of different areas in the range: 5  5 to 100  100 mm2, (say, near point D). To make a fair comparison of the RS, a
measured prior to forming the devices. Results show that the stabilization process is performed after the filament formation,
measured capacitance (at 10 kHz) increases linearly as a function as described in detail in the ESI† (S1–S3).
of device area but is independent of substrate for a given area. For Fig. 4(a) shows the on state device resistance (Ron) immediately
example, the capacitance of both the devices on thin SiO2/Si and following filament formation and stabilization. The glass devices
glass substrates of sizes 100 and 10 000 mm2 are measured to be show lower Ron than the Si/SiO2 devices suggesting that the low
B5.0 and B500.0 pF, respectively. These results agree well with the thermal conductivity substrates assist forming either wider or more
thickness of oxide layer (B5 nm) while considering the devices as a filaments. The overall process of growing wider/more filament(s)
parallel plate capacitor, whose capacitance can be estimated from: could be due to a thermally assisted mechanism during the
C = (e0eoxA)/tox; where e0, eox and tox are the permittivity of vacuum filament formation or stabilization process. Fig. 4(b)–(d) show the
(= 8.85  1012 F m1), dielectric constant (HfOx B 20) and set and reset hysteretic I–V characteristics for maximum reset
thickness of the oxide layer, respectively. Similarly, the initial voltages of 1.0 V, 1.5 V and 1.8 V, respectively. The current
device resistance is substrate independent for the given area, following set (reset) for the glass devices is higher (lower) than the
A = 10  10 mm2 (Fig. 2(b)). The capacitance and initial resistance Si/SiO2 devices. Furthermore, the tripping voltage (the voltage at
results, therefore, confirm that the intrinsic electrical device which the device begins to reset) is smaller for the devices on glass
behaviors are substrate independent. These results strongly as compared to SiO2/Si. This earlier tripping voltage and smaller
suggest that any differences in forming and switching behavior LRS (lower resistance state) for the case of glass substrate may be
are not related to differences in the structure of the active attributed to presence of larger filaments as compared to SiO2/Si
materials. substrates, resulting in a higher current and filament temperature.
Fig. 3(a) shows representative results for forming, reset, and This means that the filament re-oxidation begins at a lower voltage
set cycles of a device on the thin SiO2/Si substrate. The resulting in a larger HRS (high resistance state), and larger memory
resistance decreases by approximately nine orders of magnitude window for the glass devices. A higher HRS to LRS ratio (high to
from point A to point E (B1013 O to B104 O). The current from low resistance ratio, HLR) of the glass devices, i.e. in the range:
point A to point E at B3.5 V is small (o1011 and associated with 4 to 100 with |Vr| = 1 to 1.8 V (see Fig. S4(b) in the ESI†)
the pristine device behavior without a filament). From B3.5 V to confirmed the better performance of the devices on glass.
B4.0 V (point B to point C), there is a formation of an initial The digital reset/set of the devices is well explained in the
filament resulting in a significant increase in the current by literature with the help of atomically thin barrier layer that is
approximately two orders of magnitude. From point C point D created/destroyed during the reset/set processes.29 However, there

Fig. 2 Electrical properties of the Au (BE)/a-HfOx/Ti/Au (TE) pristine devices, i.e. before electroforming, deposited on SiO2 (280 nm)/Si and glass
(B1 mm) substrates: (a) capacitance of three different sizes and (b) resistance of 10  10 mm2 size devices as indicated. The error bars are obtained from
18 devices of each specified.

This journal is © The Royal Society of Chemistry 2020 J. Mater. Chem. C

 View Article Online

Paper Journal of Materials Chemistry C

Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

Fig. 3 (a) Electroforming (/or forming; curve (i): (A) to (E)), initial reset (ii) and set (iii) steps of a pristine Au (BE)/a-HfOx/Ti/Au (TE) device, fabricated on a
thin SiO2 (280)/Si substrates. Note that the three steps (i) to (iii) shown in (a) helps getting stable conductive filament/s and stable resistive switching.
(b) Comparison of substrate dependent forming voltage, Vf of 24 devices from each SiO2 (280)/Si and glass (B1 mm) substrate.

Fig. 4 Substrate dependent resistive switching performance comparisons of the Au/a-HfOx/Ti/Au (TE) devices: (a) initial set or ON state device
resistance after forming and stabilizing the filament/s and (b)–(d) digital (abrupt) set and analog (gradual) reset switching starting with the same
compliance current, Icc = 1.0 mA, but different reset voltages, |Vr| = 1.0 to 1.8 V.

is still lack of a detailed mechanism for the smooth analog reset LRS values for the same range of set and reset voltages, and they
response of the HfOx based memristor. A more detailed can function without a visible degradation in the operating voltages
mechanism will be discussed later, in the next section, with as shown in Fig. 5(a). In other words, the devices on glass substrate
the help of a Multiphysics COMSOL simulation results. exhibited clearly distinguishable memory windows at lower reset
Fig. 5 shows the representative multi-level resistance at voltages (starting from 0.8 V, in our case), whereas no such
various set and reset voltages, each of 100 test cycles for the reliable separation of LRS and HRS was observed with the devices
devices on both thin SiO2/Si and glass substrates. Results show that on SiO2/Si substrate (results not shown here). As expected, the
all the devices exhibit multi-level resistance (or conductance) values LRS values of the glass devices are smaller than that of the
at some appropriate applied voltages. Also, it is evident that the devices on SiO2/Si substrate, while the HRS values are greater
devices on glass substrate retain more distinguishable HRS and (see Fig. 5(a) and (b)). The difference in HRS values becomes

J. Mater. Chem. C This journal is © The Royal Society of Chemistry 2020

 View Article Online

Journal of Materials Chemistry C Paper

Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

Fig. 5 Substrate dependent analog resistive switching of Au/a-HfOx/Ti/Au (TE) devices: (a) effect of incremental applied reset voltages (|Vr| = 0.8 to 1.8 V),
showing multi-level resistive switching starting with different set and reset voltages. (b) (Magnified) showing the performance of glass and SiO2/Si devices at
1.0 V (set) and 1.5 V (reset) voltages.

more prominent at higher Vr, resulting the higher HLR values of field, generated via Joule-heating. For this, a COMSOL physical
the glass devices as observed in the digital response. Further- model that is reported in our recently published article37 has
more, it is noteworthy that the cycle-to-cycle variations in the been used to analyze the intermediate variables such as local
multi-level resistances of the SiO2/Si devices are slightly higher temperature and oxygen defect (vacancy) concentration, so that
at the larger reset (|Vr| Z 1.5 V) and set (Z1.2 V) voltages. This their role in the device performance can be understood. The
has previously been attributed to the change in filament size thermal model considers the drift, diffusion and thermophoresis
and/or geometry which is extremely sensitive to the stochastic of oxygen vacancies that are instrumental in creating and
nature of the generation and migration of the oxygen ions and suppressing the CF. The vacancy conservation that includes
vacancies.9,33 Baeumer et al. have demonstrated some of the these phenomena is solved in conjunction with current conser-
possible sources for filaments’ or device variability (i.e. including vation and energy conservation equations as described in eqn (1)
cycle-to-cycle and device-to-device) using photoelectron emission through (3).
microscopy (PEEM), combined with XPEEM measurements.33 @nV
Previous work has proposed system-level adaptations to account þ r  ðvV nV Þ ¼ r  ðDV rnV Þ þ r  ðSV DV nV rT Þ (1)
@t
for this device variability.34 The observed low operating voltages
 
and the multiple HLR values especially of the devices on glass @
r  src þ e rc ¼ 0 (2)
substrate is consistent with the literature of sub-/stoichiometric @t
HfOx based memristors.12,13,24,35,36 The devices on glass sub-
strates exhibited noticeably improved memory window, which @T 1 s
¼ r  ðkth rT Þ þ ðrcÞ2 (3)
agree well with the results from digital switching. The enhanced @t rcp rcp
performance of the devices on glass substrate can be attributed
In these equations, nV is the density of oxygen vacancies, vV is the
to the fact that both the formation of CF/s and reset switching
drift velocity of vacancies, DV is the diffusion coefficient of the
were favorable with the low heat dissipation from the substrate
same, SV is the thermophoresis coefficient, s is the electrical
to the environment. However, it is important to understand that
conductivity, c is the voltage potential, e is the permittivity of the
the different filament sizes (and perhaps geometry) for different
oxide material, T is the temperature, r is the density, cp is the
substrates must be verified with high resolution in situ trans-
specific heat, and kth is thermal conductivity. Eqn (4) is used to
mission electron microscopy (TEM) and/or PEEM studies. To
calculate the electrical conductivity of the filament as a function
understand these differences in a better way, a 2-D axisymmetric
of vacancy density and temperature. The use of this equation
COMSOL model was employed to analyze intermediate variables
results in a metal-like filament behavior for high concentration
such as oxygen vacancy density and temperature, and their
of vacancies and an insulator-type behavior for small number of
effect on the measurable parameters such as current through
vacancies. The dependence on temperature at small number of
the device.
vacancies is amplified using activation energy. For further details
Thermal modeling and experimental validation of the model, we refer to our recently published work by
Pahinkar et al.37
The substrate dependent RS performance of the devices with

EAC 
stable CF of known initial size and geometry are discussed in sHfOx  sHfO2
greater details in this section, followed by the experimental s¼ nV e kB T (4)
nV;Max
results validation. The basic hypothesis of this thermal modeling
is that the RS switching behavior of a filamentary device is Finally, the validity of experimental results is confirmed by
controlled by both the applied electric field (E) and local thermal taking two different radii of CFs, namely: 3 nm for SiO2/Si and

This journal is © The Royal Society of Chemistry 2020 J. Mater. Chem. C

 View Article Online

Paper Journal of Materials Chemistry C

Table 1 Thermophysical properties of substrate materials considered in

COMSOL simulations

Parameters Glass SiO2/Si

3
Density, r (kg m ) 2230 2320
Specific heat, cp (J kg1 K1) 840 710
Thermal conductivity, kth (W m1 K1) 1.2 138
Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

Table 2 Constants and physical parameters used in set/reset simulations

Parameter Value
nV,Max 2  1027 m3
f 2  1010 Hz
a 1010 m
Ea 1.5 eV
EAC,Max 0.125 eV
Fig. 6 Illustration of bigger filament for glass substrate (left) and smaller r 9680 kg m3
filament for SiO2/Si substrate (right) considered in the COMSOL simulations. c 60 J kg1 K1
kHfOx 5 W m1 K1
kHfO2 0.5 W m1 K1

4 nm for glass substrates, which are more realistic than taking a

Table 1 shows the thermophysical parameters considered in
single size filament as discussed above (see Fig. 6). It is crucial
the COMSOL simulations that represent the two different
to note that different shapes and sizes of the filament/s impacts
substrates and Table 2 shows additional parameters used in
heavily on the electrical performance of a device, including the
the simulations. While the specific heat and density of the two
I–V curves. A couple of nanometers to tens or even hundreds
materials are comparable, the thermal conductivity of the thin
of nanometers conducting channels are reported in the literature
SiO2 (280 nm)/Si is more than two orders of magnitude greater
for HfOx based devices, while the additional effects including
than that of glass.
different biasing conditions can also affect these sizes
The simulation results closely follow the substrate depen-
remarkably.2,29 Therefore, we used the experimental I–V curves
dent trend for both reset voltage and the memory window and
to deduce a combination of the electrical conductivity as a
this pattern can be confirmed through the examination of the
function of defect density and filament size, resulting a closely
intermediate variables in the models such as oxygen vacancies
matching I–V profiles. Fig. 7 shows a comparison of I–V curves
and local temperature. The difference in the slope of I–V curves
from the experiments and the model simulations demonstrating
on the HRS side could be attributed to the empirical treatment
a good agreement, especially on the reset side. The earlier
of the current dependence on temperature. The faster decline
tripping voltage of the glass device further confirm the validity
in the current of the glass device after the tripping voltage
of this model: bigger or wider CF of glass device as shown in
clearly represents higher HRS compared to the SiO2/Si device.
Fig. 6, with lower initial resistance, utilizes more power (= V2/R)
The difference in the two slopes of the I–V curves, namely: from
forcing the CF to rapture earlier or vice versa.
0.5 V to 1 V and 1 V to 0 V, may indicate that the device
reset switching is not only controlled by the CF size but also the
substrate’s thermal properties. Thus, the observed higher HRS
and memory window of glass devices can be attributed to the
negligible heat loss, resulting an earlier rapture and re-oxidation
of the CF/s. Similarly, a relatively higher electrical power con-
sumed by the devices on SiO2/Si substrate, that is within the
same operating set/reset voltages, further verifies the higher heat
transfer from the filament/s or devices through the SiO2/Si
substrate of as shown in Fig. S5 in the ESI.† Fig. 8(a and b)
sheds light on how the insulating or dielectric barrier is created
during the reset stage and breakdown of the barrier takes place
during the digital set stage. Because negatively charged oxygen
ions move from the capping layer Ti into the filament region and
a barrier layer, close to the BE, is created at the end of the reset
stage as shown in Fig. 8(a). This barrier layer has a lower
Fig. 7 Substrate dependent comparison of I–V plots for thin SiO2
electrical and thermal conductivity than the region elsewhere
(280 nm)/Si and glass substrates obtained from the experiments and COMSOL
model simulations. Note that different sizes of conducting filaments (CFs),
and becomes the bottleneck for the current flow. This layer also
namely: 3 and 4 nm in diameter, were used in the modeling for SiO2/Si and increases the local temperature gradient, it being the region of
glass substrates respectively. the high potential drop. Thus, an HRS is achieved. When the

J. Mater. Chem. C This journal is © The Royal Society of Chemistry 2020

 View Article Online

Journal of Materials Chemistry C Paper

Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

Fig. 8 Oxygen vacancy distribution (a) at the end of the reset stage (b) at the end of the set stage for glass substrate. Substrate dependent temperature of
(c) filament tip and (d) top electrode (TE) surface, plotted with the applied voltage for thin SiO2/Si and glass substrates. Note that 3 and 4 nm of
conducting filaments (CFs) were used in the COMSOL modeling for SiO2/Si and glass substrates respectively.

polarity is reversed to execute the set stage, the temperature the impact of substrate material on the TE temperature is more
increases initially with the current in the off-state and when the tangible. As seen from Fig. 8(d), the temperature rise in the
temperature reaches a threshold after which the vacancies are glass substrate to be more than 40 1C, while that in the SiO2/Si
mobile, drift and diffusion acting the in the same direction substrate is less than 5 1C. This behavior is obviously due to
nearly instantaneously cause the breakdown of the barrier layer different fractions of thermal power going upward toward the
as seen in Fig. 8(b). It can also be seen that the oxygen ions enter TE and downward toward the substrate for different substrates.
the Ti layer after set, which can be inferred from the fact that the The glass substrate provides more resistance to heat transfer
number of vacancies in the Ti layer have dropped. This is than SiO2/Si, hence more fraction of the total thermal power in
consistent with the observed abrupt set for all devices. In the filament goes toward the TE for glass, thereby showing an
conjunction to this, a continuous metal rich filament reestablishes appreciable temperature rise. While this study has focused on
in the HfOx layer at the end of the set stage. the fundamental understanding of the impact of thermal
The role of substrates dependent temperature field on the boundary materials on analog reset, future work will explore
defect concentration in HfOx or the oxygen vacancy activation the practical impact of these materials on the set behavior of
and movement resulting different RS performance was also the device and neuromorphic computing applications and
studied. Fig. 8(c) and (d) show the comparison of the filament behavior such as long-term potentiation and depression.38,39
tip temperature (at the interface between Ti and HfOx) and the Finally, the models were simulated to confirm that the
exposed surface of TE temperature, as a function of voltage, factors considered in the present model that affect the oxygen
respectively. When the simulations for glass and thin SiO2 anions and/or vacancies migration can indeed reproduce the
(280 nm)/Si substrates begin with the same initial conditions, non-volatile nature of the ‘memory’ or internal resistance of the
the temperature of both the filament tip and the TE in the device. Fig. 9 shows the typical gradual reset I–V curves for a
glass substrate is marginally higher than that in the SiO2/Si series of |Vr| in the range: 0.5 V to 1.0 V for a glass substrate.
substrate. This is clearly attributed to more heat remaining The ion migration and barrier layer creation at every voltage
near the device for the same power (the current and voltage are step is remembered by the device perfectly validating the
the same for both substrates), because of the glass thermo- approach considered, when compared with the experimental
physical properties. This means that the temperature required results (see Fig. S3(a), ESI†). Hur et al. also reported a physical
for onset of the migration of vacancies is achieved earlier in the model for a bipolar resistive switching behavior, with an
glass device than that in the SiO2/Si device. As the migration experimental validation, for multi-level resistance of Pt/Ta2O5x/
begins creating the dielectric barrier, the current begins to drop TaO2x/Pt devices.42 However, to the best of our knowledge, our
lowering the temperature as we observed from the experiments type modeling for the non-volatile reset has not yet been reported
(see Fig. 7). Although, the difference in the temperature profiles in the literature for a-HfOx based memristor. The use of this
of filament tip in the set side are not significant for our devices, modeling approach to further validate nanoscale ion transport

This journal is © The Royal Society of Chemistry 2020 J. Mater. Chem. C

 View Article Online

Paper Journal of Materials Chemistry C

oxidized to create B280 nm thick SiO2, using dry oxidation at

1000 1C. While the microscope glass slides (Gold Seal Catalog
No. 3010) were used as received. The sneak current path free
devices of different sizes were fabricated for different experi-
ments (details are presented in the Result section). Prior to spin
coating the substrates with negative photoresist, NPR (NR-71
3000P), we cleaned the substrates with organic solvents and
Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

piranha solution, then dried with N2 blower. As described in

our recent paper, both the bottom and top electrodes (BE and
TE) were patterned using a maskless aligner (Heidelberg Instru-
ments MLA 150),40 followed by the metal depositions using
e-beam evaporation (20 nm and 5 nm Ti underneath the 70 nm
BE and 150 nm TE respectively, deposited at the fixed deposition
Fig. 9 The I–V curves from the COMSOL model showing gradual reset rates 1.0 Å s1). Both the BE and TE layers were deposited using
with a non-volatile resistance memory. 99.999% Au source at ultra-high vacuum (B3  106 Torr) and
the Ti layers were deposited without breaking vacuum using
99.95% Ti source pellets (Kurt J. Lesker). Note that the 20 nm
phenomena as a function of current flow and temperature field for
adhesion layer of Ti film was deposited to improve the adhesivity
HfOx and other MOs exhibiting similar properties is thus warranted
of Au layer, while the 5 nm Ti was used as a buffer layer
and ongoing.
(/or oxygen getter capping layer) to improve the oxygen diffusion,
as described in the literature.5,31 The HfOx (B5 nm) was
Conclusion synthesized using atomic layer deposition (ALD), Cambridge
Nanotech Fiji F202 system. Thermal ALD process was used at
In summary, Au (BE)/a-HfOx (5 nm)/Ti/Au (TE) memristor devices 250 1C, with the Tetrakis(dimethylamido)hafnium (TDMA-Hf)
were fabricated on two different substrates, thin SiO2 (280 nm)/Si and DI water as oxygen precursors. The pulse times for both the
and glass, and characterized to study substrate dependent device precursors (water = 0.06 s and TDMA-Hf = 0.25 s) were fixed for
behavior. The reported memristor device fabrication technique is whole deposition cycles. Before depositing the HfOx film on the
simple, scalable, and highly reliable (almost 100% device yield in actual device’s substrate, the deposition rate was confirmed with
our case). Our study showed that the bipolar RS, both digital and a test run to ensure the accuracy for the final thickness and used
analog, of the devices is dependent on the thermal diffusivities of the optimized rate to estimate total number of cycles needed
the substrates. Devices on glass substrates showed the best RS (i.e. 45 cycles at 1.12 Å per cycles for 5 nm). Then, the final
switching performance, including a wider memory window and a thickness and the uniformity of the HfOx films were also
lower reset voltage, than the devices on SiO2/Si substrates. This is confirmed with the multi-point thickness measurements using
attributed to the low thermal conductivity of the glass, enabling the an ellipsometer (J. A. Woollam M2000), measured to be 5.0 
device to hold the temperature and hence improve performance 0.1 nm (with uniformity 2%). After the BE and TE deposition, lift-
due to favorable diffusion of oxygen ions at the HfOx–Ti interface. off processes were performed with acetone at room temperature
Comprehensive computational modeling results confirm that the for a period of at least 8 hours. Also, the samples were sonicated
oxygen ions move back and forth between the Ti ‘‘capping layer’’ for about 30 s before the final rinse off to remove the PR
and the HfOx ‘‘active oxide layer’’ during reset and set stages. The completely. Finally, the HfOx film, outside the crossbar region,
movement of ions is reported as a primary function of applied was etched by reactive ion etch (RIE) starting with SF6/Ar mixture
electric field, thermophoresis, and diffusion, while the kinetics are (flow rates: 30/5 sccm) for a minute, setting the RF power at
governed by the material properties of the substrates. Our results 200 W and the pressure at 40.0 mTorr. It is noteworthy that the
indicate that the a-HfOx based memristor devices can be a as-prepared devices were characterized for the electrical proper-
promising candidate for neuromorphic computing. However, ties without any post-fabrication treatments.
for commercial production, besides the requirement of device
size and behavior, other physical parameter and limitations such Materials and device characterization
as compatibility of substrates, device packaging, etc., could affect As-deposited HfOx films were characterized for some of the
the RS switching, and therefore, it may require an extensive important physical and chemical properties that are required,
study to further improve the analog response of these devices. or at least helpful, to investigate the device RS behavior. For
example, amorphous phase was confirmed with glancing angle
Experimental section X-ray diffraction (XRD) measurements (PANalytical Materials
Research Diffractometer, MRD), with Cu Ka (l = 1.540598 Å)
Device fabrication and 2y in the range from 201–801, at a step size of 0.011. And,
All devices were fabricated on about 100  100 square size cleaned X-ray photoelectron spectroscopy (XPS) was performed for chemical
substrates. The Si (Orient: h100i, test grade) wafer were pro- composition using a Thermo Scientific K-alpha XPS system with a
cured from University Wafer, and the surface was thermally monochromated Al Ka X-ray source and a hemispherical analyzer

J. Mater. Chem. C This journal is © The Royal Society of Chemistry 2020

 View Article Online

Journal of Materials Chemistry C Paper

in constant analyzer energy mode (50 eV pass energy, 0.1 eV step 6 M. Hu, C. E. Graves, C. Li, Y. N. Li, N. Ge, E. Montgomery,
size, and a beam size of 400 mm). From the XPS analysis, the Hf : O N. Davila, H. Jiang, R. S. Williams, J. J. S. Yang, Q. F. Xia and
ratio was estimated to be 1 : 1.8, which corresponds to a stoichio- J. P. Strachan, Adv. Mater., 2018, 30, 1705914.
metric amorphous HfO1.8. 7 Z. R. Wang, S. Joshi, S. E. Savel’ev, H. Jiang, R. Midya, P. Lin,
M. Hu, N. Ge, J. P. Strachan, Z. Y. Li, Q. Wu, M. Barne,
Electrical characterization G. L. Li, H. L. Xin, R. S. Williams, Q. F. Xia and J. J. Yang,
All electrical measurements were performed using a Keithley Nat. Mater., 2017, 16, 101–108.
Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

4200-SCS (Semiconductor Characterization System) and a 8 S. B. Long, L. Perniola, C. Cagli, J. Buckley, X. J. Lian,
Cascade Microtech probe station at room temperature and E. Miranda, F. Pan, M. Liu and J. Sune, Sci. Rep., 2013,
ambient conditions. As reported in our previous work, we 3, 2929.
characterized the devices with and without an external transistor 9 S. M. Yu, B. Gao, Z. Fang, H. Y. Yu, J. F. Kang and
to confirm if the transistor is required to avoid the possible H. S. P. Wong, Adv. Mater., 2013, 25, 1774–1779.
damages to the devices: for example, permanent damage of the 10 Y. Abbas, Y. R. Jeon, A. S. Sokolov, S. Kim, B. Ku and C. Choi,
conductive filament or the performance degradation.41 Bias was Sci. Rep., 2018, 8, 1228.
applied to the TE while the BE was grounded. For DC sweeping 11 F. Messerschmitt, M. Kubicek and J. L. M. Rupp, Adv. Funct.
voltage ramping was also fixed (= 0.02 V) for all measurements Mater., 2015, 25, 5117–5125.
in both positive and negative potential windows. Initial device 12 J. Shang, W. H. Xue, Z. H. Ji, G. Liu, X. H. Niu, X. H. Yi,
characteristics were compared by measuring the capacitance– L. Pan, Q. F. Zhan, X. H. Xu and R. W. Li, Nanoscale, 2017, 9,
voltage (C–V) and initial resistance of the pristine devices, of all 7037–7046.
four different sizes as mentioned above. Then, the switching 13 K. L. Lin, T. H. Hou, J. Shieh, J. H. Lin, C. T. Chou and
performance of 10  10 mm2 size devices, fabricated on both thin Y. J. Lee, J. Appl. Phys., 2011, 109, 84104.
SiO2 (280 nm)/Si and glass substrates, were measured and 14 M. Schie, M. P. Muller, M. Salinga, R. Waser and R. A. De
analyzed. Please note that all the test devices, on both glass Souza, J. Chem. Phys., 2017, 146, 94508.
and SiO2/Si substrates, were characterized without using the 15 A. Padovani, L. Larcher, O. Pirrotta, L. Vandelli and
thermal paste; and this can cause a slight change in the device G. Bersuker, IEEE Trans. Electron Devices, 2015, 62,
behavior that might be not easy to quantify. 1998–2006.
16 A. Wedig, M. Luebben, D. Y. Cho, M. Moors, K. Skaja,
V. Rana, T. Hasegawa, K. K. Adepalli, B. Yildiz, R. Waser
Conflicts of interest and I. Valov, Nat. Nanotechnol., 2016, 11, 67–74.
17 B. Traore, P. Blaise, B. Sklenard, E. Vianello, B. Magyari-
The authors declare no competing financial interest.
Kope and Y. Nishi, IEEE Trans. Electron Devices, 2018, 65,
507–513.
Acknowledgements 18 S. Clima, Y. Y. Chen, A. Fantini, L. Goux, R. Degraeve,
B. Govoreanu, G. Pourtois and M. Jurczak, IEEE Electron
The authors acknowledge Professors William A. Doolittle and Device Lett., 2015, 36, 769–771.
Wei-Cheng Lee for the helpful discussion. This work is supported by 19 S. Yu, Y. Wu, R. Jeyasingh, D. Kuzum and H. S. P. Wong,
the Air Force Office of Scientific Research MURI entitled, ‘‘Cross- IEEE Trans. Electron Devices, 2011, 58, 2729–2737.
disciplinary Electronic-ionic Research Enabling Biologically Realistic 20 R. Waser and M. Aono, Nat. Mater., 2007, 6, 833–840.
Autonomous Learning (CEREBRAL)’’ under award number FA9550- 21 Y. C. Yang and W. D. Lu, IEEE Trans. Nanotechnol., 2016, 15,
18-1-0024. This work was performed in part at the Georgia Tech 465–472.
Institute for Electronics and Nanotechnology, a member of the 22 R. Waser, Nanoelectronics and Information Technology, Wiley-
National Nanotechnology Coordinated Infrastructure, which is sup- VCH, 2012.
ported by the National Science Foundation (grant ECCS1542174). 23 G. Bersuker, D. Veksler, D. M. Nminibapiel, P. R. Shrestha,
J. P. Campbell, J. T. Ryan, H. Baumgart, M. S. Mason and
References K. P. Cheung, J. Comput. Electron., 2017, 16, 1085–1094.
24 W. Wu, H. Wu, B. Gao, N. Deng, S. Yu and H. Qian, IEEE
1 H. P. Wong, H. Lee, S. Yu, Y. Chen, Y. Wu, P. Chen, B. Lee, Electron Device Lett., 2017, 38, 1019–1022.
F. T. Chen and M. Tsai, Proc. IEEE, 2012, 100, 1951–1970. 25 G. Bersuker, D. C. Gilmer, D. Veksler, P. Kirsch, L. Vandelli,
2 J. J. Yang, D. B. Strukov and D. R. Stewart, Nat. Nanotechnol., A. Padovani, L. Larcher, K. McKenna, A. Shluger, V. Iglesias,
2013, 8, 13–24. M. Porti and M. Nafria, J. Appl. Phys., 2011, 110, 124518.
3 S. M. Yu, Proc. IEEE, 2018, 106, 260–285. 26 S. Kim, S. Choi and W. Lu, ACS Nano, 2014, 8, 2369–2376.
4 S. H. Jo, T. Chang, I. Ebong, B. B. Bhadviya, P. Mazumder 27 D. Ielmini, IEEE Trans. Electron Devices, 2011, 58,
and W. Lu, Nano Lett., 2010, 10, 1297–1301. 4309–4317.
5 A. Subramaniam, K. D. Cantley, G. Bersuker, D. C. Gilmer 28 S. Kim, S. J. Kim, K. M. Kim, S. R. Lee, M. Chang, E. Cho,
and E. M. Vogel, IEEE Trans. Nanotechnol., 2013, 12, Y. B. Kim, C. J. Kim, U. I. Chung and I. K. Yoo, Sci. Rep.,
450–459. 2013, 3, 1680.

This journal is © The Royal Society of Chemistry 2020 J. Mater. Chem. C

 View Article Online

Paper Journal of Materials Chemistry C

29 Y.-E. Syu, T.-C. Chang, J.-H. Lou, T.-M. Tsai, K.-C. Chang, 36 W. F. He, H. J. Sun, Y. X. Zhou, K. Lu, K. H. Xue and
M.-J. Tsai, Y.-L. Wang, M. Liu and S. M. Sze, Appl. Phys. Lett., X. S. Miao, Sci. Rep., 2017, 7, 10070.
2013, 102, 172903. 37 D. G. Pahinkar, P. Basnet, B. Zivasatienraj, A. Weidenbach,
30 G. Broglia, G. Ori, L. Larcher and M. Montorsi, Modell. M. West, W. A. Doolittle, E. Vogel and S. Graham, Computa-
Simul. Mater. Sci. Eng., 2014, 22, 065006. tional Investigation of Nanoscale Memristor Devices for
31 P. Calka, M. Sowinska, T. Bertaud, D. Walczyk, J. Dabrowski, Neuromorphic Computing, 18th IEEE Intersociety Conference
P. Zaumseil, C. Walczyk, A. Gloskovskii, X. Cartoixà, J. Suñé and on Thermal and Thermomechanical Phenomena in Electronic
Published on 10 March 2020. Downloaded by National University of Singapore on 3/16/2020 2:54:16 PM.

T. Schroeder, ACS Appl. Mater. Interfaces, 2014, 6, 5056–5060. Systems (ITherm), 2019, pp. 219–225.
32 S. Kim, H. Y. Jeong, S. K. Kim, S. Y. Choi and K. J. Lee, Nano 38 S. Yu, Y. Wu and H. S. P. Wong, Appl. Phys. Lett., 2011,
Lett., 2011, 11, 5438–5442. 98, 103514.
33 C. Baeumer, R. Valenta, C. Schmitz, A. Locatelli, T. O. Mentes- , 39 S. Kim, J. Chen, Y. C. Chen, M. H. Kim, H. Kim, M. W.
S. P. Rogers, A. Sala, N. Raab, S. Nemsak, M. Shim, C. M. Kwon, S. Hwang, M. Ismail, Y. Li, X. S. Miao, Y. F. Chang
Schneider, S. Menzel, R. Waser and R. Dittmann, ACS Nano, and B. G. Park, Nanoscale, 2019, 11, 237–245.
2017, 11, 6921–6929. 40 C. J. Perini, P. Basnet, M. P. West and E. M. Vogel, ACS Appl.
34 P. Y. Chen, B. B. Lin, I. T. Wang, T. H. Hou, J. P. Ye, Mater. Interfaces, 2018, 10, 39860–39871.
S. Vrudhula, J. S. Seo, Y. Cao, S. M. Yu and IEEE, 2015 IEEE/ 41 B. Chakrabarti and E. M. Vogel, Microelectron. Eng., 2013,
ACM International Conference on Computer-Aided Design, 109, 193–196.
2015, pp. 194–199. 42 J. H. Hur, K. M. Kim, M. Chang, S. R. Lee, D. Lee, C. B. Lee,
35 G. Niu, H.-D. Kim, R. Roelofs, E. Perez, M. A. Schubert, M. J. Lee, Y. B. Kim, C. J. Kim and U. I. Chung, Nanotechnology,
P. Zaumseil, I. Costina and C. Wenger, Sci. Rep., 2016, 6, 28155. 2012, 23, 225702.

J. Mater. Chem. C This journal is © The Royal Society of Chemistry 2020