UNIVERSITY OF PENNSYLVANIA

Conducting Nanofilaments in Metal Oxide Resistive

Switching Memory Qualifying Exam Manuscript

Yang Lu

Department of Material Science and Engineering

6 May, 2013

Contact Email: yanglu1@seas.upenn.edu

CONTENT

RESEARCH PROGESS ............................................................................................................................... 1

ABSTRACT .................................................................................................................................................. 2

1. INTRODUCTION ................................................................................................................................ 3

1.1. Hafnium Oxide ............................................................................................................................ 3

1.2. Resistive Random Access Memory ............................................................................................ 3

1.3. Conductive Filaments in RRAM and Related Switching Mechanisms .................................. 5

2. EXPERIMENT AND RESULTS ......................................................................................................... 7

2.1. Experiment Procedures .............................................................................................................. 7

2.2. I-V characteristics of TiN/HfO2 system ..................................................................................... 7

2.3. Identification of the conductive region ...................................................................................... 8

2.4. Chemical Properties of the Conductive Region...................................................................... 10

2.5. Electronic Properties of the Conductive Region .................................................................... 10

2.6. Structural Properties of the Conductive Region .................................................................... 11

3. DISCUSSION AND CRITIQUE ........................................................................................................ 12

3.1. Switching Polarity Analysis...................................................................................................... 12

3.2. Oxygen Vacancy Generation Mechanism ............................................................................... 13

3.3. Oxygen Vacancy Conduction Mechanism .............................................................................. 17

3.4. Amorphization in Resistive Switching .................................................................................... 18

4. SUGGESTIONS FOR FUTURE WORK ........................................................................................... 19

4.1. Verification of Filamentary Effect ........................................................................................... 19

4.2. Identification of Conducting Nanofilaments .......................................................................... 20

5. CONCLUSION ................................................................................................................................... 21

REFRENCE ................................................................................................................................................ 22

1

RESEARCH PROGESS

The topic of my PhD thesis is about metal-insulator transition in random materials for nanometallic

resistance switching random access memory (RRAM) application. My research is mainly focused on the

purely electronic switching behaviors in new material system. I have gained solid foundations related to

this area and accomplished several preparations for further in-depth research:

1. Comprehensive readings about experimental methods, typical electrical behaviors, physical

models and main challenges in this area, such as PhD thesis of previous students, published

papers of current group and group meeting slides.

2. Training of instruments for material characterization: Lasker probe station in controlled

environments, Asylum Atomic Force Microscope (AFM) at NBIC and Quanta 600 Scanning

Electron Microscope (SEM) at PRNF.

3. Training of tools for thin film fabrication: Atomic Layer Deposition (ALD), Plasma Enhanced

Chemical Vapor Deposition (PECVD) and Magnetron sputtering (all in Penn nanofab).

4. Conducting basic electrical measurement for different structures, including nano-cross bar arrays

and complementary devices and divices made of various amorphous mixtures of oxides and

metals.

5. Testing device fatigue for AC reading and writing regime in different conditions, such as in

vacuum and with ALD capping layers. Device Endurance of 105 was achieved, already meeting

the requirement for flash memories.

6. Conducting AFM tests for electrical properties of nano scale cells and force induced resistive

switching.

In the next year, I want to get myself more involved in research work and plan to investigate these

issues including: carrier properties in different resistance states, effects of electrode materials on

resistance switching and integration of selectors into devices for memory arrays application.

2

ABSTRACT

Resistive random access memory (RRAM) has been extensively studied as one of the most promising

candidates for the next generation of memory technology due to excellent memory performances. RRAM

is commonly found in oxide thin films, especially those of transition metal. Although switching

mechanism is not fully understood, many switching behaviors can be explained by assuming there’re

conductive filaments connecting top and bottom electrodes. Therefore, experimental identification and

detailed investigation of conductive filaments in nanometer scale become one of the most challenging and

important issues in this area.

This manuscript involves two papers about conducting nanofilaments in metal oxide RRAM. In the

main paper to be critiqued [1], P Calka et al. dealt with resistive switching devices of the binary metal

oxide system: HfO2 with TiN as bottom electrode. They investigated structural, chemical and electronic

properties of oxygen-deficient regions (HfO2-x regions) in the oxides by C-AFM, STEM-EELS and

HRTEM. They conclude that those oxygen-deficient regions can be identified as conducting

nanofilaments and resistive switching can be explained as breaking of Hf-O bonds and migration of

oxygen ions / vacancies. In the other paper [2], Deok-Hwang Kwon and coworkers demonstrated the

presence of TinO2n-1 (known as Magnéli phases) in TiO2-based RRAM and discussed their roles of

conducting current in TiOx-based resistive switching memory.

In this manuscript, I will first give a brief introduction of various resistive switching material systems,

especially binary metal oxide, and different switching mechanisms. Secondly, experiment methods and

important results of the main paper will be presented. Then I will discuss physical explanations for

observed phenomena and critique possible flaws of this paper, including switching polarity, physics of

oxygen vacancies generation, and roles of HfO2-x regions in electron transport. Lastly, I will propose

some possible future work that may further identify conducting nanofilaments in metal oxide RRAM and

elucidate its switching mechanism.

3

1. INTRODUCTION

1.1. Hafnium Oxide

Hafnium is a transition metal of group IV in the periodic table with atom atomic number 72. Electron

orbital structure for Hf is [Xe]6s24f

145d

2. And its oxidation states are +4, +3 or +2. Hafnium oxide with

the formula HfO2, also known as hafnia, is one of the most common and stable compounds of hafnium.

Under normal pressure, single-crystal HfO2 is available in three polymorphic modifications: a low-

temperature monoclinic phase (space group P21/c), a tetragonal phase space group (P42/n mc) obtained at

temperatures above 2000 K, and the cubic phase (Fm3m) that forms at temperatures above 2870 K [3].

Crystal structures are shown in Fig.1.

Fig.1 (adapted from Ref.3) Crystal structures for the crystalline phase of HfO2

HfO2 is an electrical insulator with a band gap of 5.7eV. Due to its high dielectric constant (25 in

comparison with 3.9 of silicon dioxide), HfO2 can be used to achieve the same value of oxide capacitance

with a larger thickness, which decreases the leakage current thus the power consumption. Therefore,

HfOx-based materials have been employed as a high-k dielectric for the gate insulator for high-

performance CMOS MOSFETs [4]. Deposition of HfO2 is usually done by atomic layer deposition (ALD)

with metal iodide, chloride or amide precursors [5-6].

In addition, HfOx is one of the most mature RRAM materials explored due to its rich defects. The

TiN/HfOx/Pt structure is usually employed to achieve the bipolar switching characteristics [7]. 10nm x

10nm HfOx-based crossbar RRAM device with excellent performance has already been realized [8].

1.2. Resistive Random Access Memory

4

Resistive random access memory (RRAM) has been extensively studied as one of the most promising

candidates for the next generation of memory technology due to the excellent memory performances such

as fast switching speed (<10ns), lower operating voltage (<2V), great potential of scalability and high

density 3D integration [9]. It stores the information “0” and “1” through two different resistance levels:

low resistance state (LRS) and high resistance state (HRS). The device structure is simply an oxide

material sandwiched between two metal electrodes, called the metal–insulator–metal (MIM) structure.

Metals can be either the same (usually inert metals) or different (one is inert and the other is active) for

the two sides. A huge variety of insulators in the MIM configuration have been reported to show

hysteretic resistance switching, such as transition metal oxides (NiO[10], TiO2[11], HfO2[12]), perovskite

oxides (SrTiO3[13]), chalcogenides (Ag2S[14]) and even organic compounds [15]. Among these material

systems, binary metal oxides, as shown in Fig.2 [16], have drawn the most attentions.

Fig.2 (adapted from Ref.16) binary oxide that exhibits resistance switching

Typically, an initial electroforming step such as a current-limited electric breakdown is induced in the

virgin sample. This step preconditions the system which can subsequently be switched between LRS and

HRS. Usually, switching process from HRS to LRS is called SET process; corresponding trigger voltage

is called SET voltage. And process from LRS to HRS is called RESET process; corresponding trigger

voltage is called RESET voltage. There are two different schemes with respect to the electrical polarity to

clarify a RRAM device: unipolar and bipolar, which is schematically shown in Fig.3 [17]. For unipolar

device, SET and RESET voltage have the same bias polarity and only differ in magnitude (RESET

voltage is usually less than SET voltage). For bipolar device, SET and RESET voltage are in opposite

5

direction (This anti-symmetry is often regarded to be introduced by different electrodes). There’s always

a current compliance in SET or forming process to prevent damage from large current to the stack.

Fig.3 (adapted from Ref.17) Typical I-V curves of RRAM. a) Unipolar switching whose SET and RESET voltage

have the same polarity. b) Bipolar switching whose SET and RESET voltage have opposite polarity.

1.3. Conductive Filaments in RRAM and Related Switching Mechanisms

SET and RESET process are typically reported as a confined, filamentary effect rather than a

homogeneously distributed one. Therefore, many aspects of the switching behavior can be understood by

assuming that the current flows through conductive filaments (CF) connecting top and bottom electrodes

[18]. Schematic view of conductive filaments in MIM structure is shown in Fig.4 [17]. Experimental

identification and detailed investigation of conductive filaments in nanometer scale are essential to

elucidate switching mechanisms for RRAM. Besides these two papers [1-2] discussed in this manuscript,

there are other works about observation and verification of nanofilaments [19-21].

Fig.4 (adapted from Ref.17) Filamentary conduction in RRAM a) vertical configuration b) lateral configuration

6

The switching mechanisms for RRAM differ in various material systems. They can be roughly

classified into three different types: valence-change mechanism related to ionic effects (typically for

bipolar switching), charge-trap mechanism related to electronic effects and fuse-antifuse mechanism

related to thermal effects (typically for unipolar switching).

Valence-change mechanism is related to an ion-migration-induced redox process. Ions can either be

cations or anions. For the first case, corresponding MIM system consists of an electrode made from an

electrochemicall active metal (such as Ag), a solid electrolyte as an ion-conducting ‘I’ layer (such as

Ag2S), and a counter electrode made from an inert metal (such as Pt). Metal cations migrate towards the

cathode (inert metal), where they are reduced and form the conductive filaments of metal atoms. After

changing the polarity of voltage bias, metal filaments are dissolved near inert electrode. In this way,

RRAM is RESET to HRS. For anions migration, it is particularly the case in transition metal oxides

RRAM, where oxygen ions or vacancies are much more mobile. Electroforming or SET process is

regarded as dielectric soft breakdown associated with the generation of oxygen-deficient region and

migration of O2−

toward the anode [22]. Transition metal cations accommodate this deficiency by

trapping electrons emitted from the cathode. The reduced valence states of transition metal cations

typically turn the oxide into a metallically conducting phase. At the anode, the oxidation reaction of O2-

leads to the accumulation of oxygen atoms or release of oxygen gas. RESET process with opposite

voltage bias can be regarded as the anode interface localized switching: accumulated oxygen is ionized

and drifts to recombine oxygen vacancies. In this way, conducting filaments are ruptured near the anode

and RRAM is back to HRS.

Charge-trap model can be considered as another origin of resistive switching. After injection of

charges by Fowler–Nordheim tunneling at high electric fields, they are trapped at sites such as defects or

nanoclusters. Consequently, electrostatic barrier changes and resistance switching occurs.

As for thermally related mechanisms, electroforming or SET process is similar to the dielectric soft

breakdown mentioned in valence-change mechanism. However, RESET process results from thermally

activated diffusion of oxygen ions either from oxygen accumulation region near anode [22] or defects

nearby oxygen-rich phases [23].

It’s easy to see that valence-change mechanism, which is polarized ionic effect, is mainly responsible

for bipolar switching; while fuse-antifuse mechanism, without polarity, is mainly responsible for unipolar

switching. There’re also unified models of both unipolar and bipolar switching for metal oxides RRAM

[23].

7

2. EXPERIMENT AND RESULTS

2.1. Experiment Procedures

A 20 nm thick TiN layer is deposited on SiO2/Si (100) substrate by sputtering. To prevent the

possible influence of oxidation of TiN, chemical etching with 1% HF was done prior to HfO2 deposition.

Then HfO2 thin film layer with 10 nm thick is deposited by atomic layer deposition (ALD) at 350°C,

which uses HfCl4 as precursor and H2O as oxidant. Oxide layer is characterized by Rutherford back-

scattering spectroscopy (RBS), which shows x=2.3, demonstrating the initially oxygen rich film. A partial

cross-section is prepared by focus ion beam (FIB) using a gallium beam. Two parallel trenches (1 x 19

um2) with 2 nm depth are made on both side of the as fabricated HfO2 area. Resistive switching is then

performed in these trenches for further localization and characterization.

I-V characteristics are measured using atomic force microscopy in the conducting mode (C-AFM).

Sample structure and measurement configuration are shown in Fig.5. In the measurement, TiN bottom

electrode serves as cathode while AFM tip (highly doped diamond coated Si) serves as anode (a positive

bias is applied). After resistive switching (SET process), conductive region in the oxide is localized by

scanning spreading resistance microscopy (SSRM). After protective coatings of SiOCH and Pt are

deposited and cross section is lifted out, this conductive region is characterized by scanning transmission

electron microscopy (STEM). In this process, chemical composition is investigated by electron energy

loss spectroscopy (EELS) and structure information is obtained by high-angle annular dark field detector

(HAADF) and high-resolution TEM (HRTEM).

Fig.5 Experimental setup for resistive switching of the TiN/HfO2 stack.

2.2. I-V characteristics of TiN/HfO2 system

Fig. 6 shows electrical properties of the TiN/HfO2 system before (c) and after (d) considering tip

resistance and voltage distribution. As mentioned in the introduction, electroforming is needed before

8

SET/RESET switching, which is shown as the blue line. A forming voltage of 4.2V is observed, where

the current starts to increase dramatically (resistance decreases to LRS) until it reaches compliance

current (500uA). In the second voltage sweep, initial current remains at high level, indicating the

persistent state of LRS and drops down near RESET voltage around 4V, where HRS is achieved. During

the third sweep, current starts in HRS as expected and changes again to LRS after SET voltage around

1.3V with the same polarity. Therefore, this TiN/HfOx stack exhibits resistance hysteresis behavior and

can be regarded as unipolar RRAM. Considering tip resistance and the real voltage dropped on the device,

we find that RESET voltage is actually smaller than SET voltage (It is typically the case for unipolar

RRAM). This is related to the thermally activated switching mechanism: current in HRS is smaller than

that in LRS, which requires larger voltage to provide enough Joule heating and electric field for switching.

Issues about voltage polarity and switching mechanism will be fully discussed in Section.3.

Fig.6 I-V curves of HfOx-based unipolar RRAM. C) Va represents total voltage applied, d) Ve represents the voltage

dropped on device (taking account of tip resistance).

2.3. Identification of the conductive region

Using the same measurement, DC voltage sweep is done for as fabricated HfO2 sample in the partial

cross section after FIB. After this electroforming, resistance mapping with SSRM is performed in contact

mode while bottom electrode is biased at 1V. Corresponding results are shown in Fig.7. As shown in

Fig.7 a), forming voltage is around 6V, larger than 4V in Fig.6. Reading operation is performed with

voltage less than 1V, resistance of 10MΩ in LRS is observed. LRS resistance reaches such a large value

because conducting filaments are suppressed by the low current compliance and small effective contact

area of AFM tip. A conductive region accompany with a protrusion with area of 100nm x 100nm is

observed in resistance map and topography as shown in Fig.7 b). This conductive region is corresponding

to the contact area of AFM tip, which has a radius of 20-50 nm. Resistance of this region is 10MΩ,

9

consistent with the value obtained in I-V curve of reading, indicating no modification of SSRM. The

protrusion area may be related to oxygen accumulation region resulting from the ion migration during

electroforming. Further discussion of oxygen accumulation will be addressed in Section.3.

Fig.7 a) I-V characteristics of the electroforming and reading process. Inset: SEM image of HfO2 surface after

electrical stresses. b) Topography and resistance maps of HfO2 in LRS.

Fig.8 a) STEM-EELS chemical oxygen map of the conductive region. b) Oxygen concentration and relative

thickness along the dashed line in a).

10

2.4. Chemical Properties of the Conductive Region

STEM-EELS measurements of O-K, N-K and Ti-L2,3 edges are performed for the conductive region

in Fig.7 b). Ti and N maps show no diffusion of metallic atoms into the oxide. While in oxygen map, a

localized oxygen-deficient region appears near the top of HfO2 layer with an extension of 20 nm as shown

in Fig.8 a). It’s worth noting that this oxygen-deficient region does not connect top and bottom electrode

as a complete filament. And no obvious oxygen accumulated interfaces are found, even near the

protrusion. In Fig.8 b), oxygen concentration and thickness profile are plotted along the oxygen-deficient

and oxygen-rich regions, indicating a clear drop of oxygen. However, its shape and location are not

supported by any proposed physical mechanisms so far. Relationship between electroforming and

oxygen-deficient region and its origin will be discussed in Section.3. Identification of this oxygen-

deficient region as nanofilaments actually remains in doubt.

2.5. Electronic Properties of the Conductive Region

Fig.9 EELS oxygen K-edges spectra in oxygen rich and deficient regions.

To investigate the electronic properties of HfO2-x conductive region, oxygen K-edge EELS spectra are

measured in both oxygen-deficient (HfO2-x) and oxygen-rich (HfO2) regions. As compared in Fig.9,

differences between these two spectra can be summarized as: 1. Decrease of total intensity in oxygen poor

region; 2. Additional shoulder III on the left and IV on the right side of peak I; 3. Shapes and separation

of peak I and II. Based on several literatures, these differences can be attributed to the creation of oxygen

vacancies, which introduce localized states within the band gap and local disorder, making the excitation

of O 1s electron more easily. However, due to the incomplete connection of oxygen vacancies “path”

with electrodes, their roles in electron conduction need to be carefully verified which will be presented in

Section.3.

11

2.6. Structural Properties of the Conductive Region

HRTEM image of the same conductive region is shown in Fig.10 a). Main central grain labeled as

HfO2-x has similar position and shape to the oxygen deficient region illustrated in Section 2.4. HRTEM

image shows that oxygen-deficient region does not fully extended along the electric field with some parts

missing near cathode, which is explained by hot electron injection model. They claim that electrons are

injected from cathode to the oxide through grain boundaries. It’s the intersection between three grain

boundaries where defects start to be generated. Amorphousness appears in the regions near two electrodes

and in oxygen-deficient regions due to different effects. Near the cathode (bottom electrode) where

current density is high, amorphization is probably caused by local heating. While near the anode,

accumulation of oxygen may attribute to the amorphousness. Partial amorphization of HfO2-x region (e.g.

A2) is verified by Fig.11 b), which shows additional rings compared to that of HfO2 crystalline region.

However, this generation model of oxygen vacancies and explanations about amorphization cannot be

fully supported by current data in this paper. We’ll come back to the verification of this model in

Section.3.

Fig.10 a) HRTEM image of conductive region, dashed line represents grain boundaries. b) FFT of selected areas

together with simulated diffraction patterns.

12

3. DISCUSSION AND CRITIQUE

3.1. Switching Polarity Analysis

Unipolar switching properties of the TiN/HfO2 system are demosntrated in Section 2.2. In literature,

polarity of HfOx-based RRAM varies. With Pt as both of the top and bottom electrodes, this system is

reported to be unipolar [25]. With Pt as one electrode and TiN as another one, bipolar switching is often

observed [26]. For metal oxide RRAM, polarity of switching is believed to be related to electrode/oxide

interface [22]. Switching types in different material systems are summarized in Table.1 [16, 22].

Unipolar Bipolar

Pt/NiO/Pt Pt/NiO/SrRuO3

Pt/TiO2/Pt TiN/TiO2/Pt

Pt/ZnO/Pt TiN/ZnO/Pt

Pt/ZrO2/Pt Ti/ZrO2/Pt

Pt/HfO2/Pt TiN/HfO2/Pt

Pt/Al2O3/Ru Ti/Al2O3/Pt

Table.1 Switching polarity for different material systems in metal oxide RRAM

As mentioned in Section 1.3, unipolar switching is caused by a thermally activated process while

bipolar switching is related to polarized effects like ion drift. In metal oxide RRAM, electroforming/SET

process of both unipolar and bipolar is similar to dielectric soft breakdown. This process has to be

controlled by electric field instead of thermal heating because current in HRS is too low to generate

sufficient heat. Conductive filaments of oxygen vacancies are formed together with accumulation of

oxygen ions near the anode. If the electrode is oxidizable, for example, TiN/HfO2/Pt system, an oxidized

interface layer of TiON will be generated as shown in Fig.11 [27]. For RESET process, oxygen ions need

to migrate towards oxygen vacancies. If there’re no interfacial layers, diffusion of oxygen ions can be

thermally activated with the existing concentration gradient, resulting in unipolar switching. If there’re

interfacial barriers of the oxidized layer, diffusion is not easy to happen. Voltage with opposite polarity

must be applied to make oxygen ions drift to the filament, making it a bipolar switching.

It’s worth noting that co-exists of unipolar and bipolar switching is also reported in TiN/HfO2/Pt

system [28], Pt/TiO2/Pt system [29] and Ni/NiO/Ni system [30]. By increasing the current compliance in

SET/electroforming, bipolar switching can be converted into unipolar switching. This can be explained by

the thermally activated diffusion model. When a large current compliance is applied, there’ll be more

oxygen vacancies and perhaps grain boundaries generated and more oxygen ions accumulated (they may

13

locate not only near the anode but also near the filament). Because diffusion of oxygen is mainly along

grain boundaries and highly related to concentration gradient, thermal diffusion is easier with larger

current compliance. Oxygen ions can migrate toward oxygen vacancies without the help of electric fields

and thus unipolar switching happens.

Fig.11 (adapted from Ref.27) Cross-section TEM of (a) the single-layer sample and (b) the double-layer sample;

HR-TEM of (c) the cell in the single layer sample, (d) the bottom cell, and (e) the top cell in the double-layer sample.

Just based on one direction sweep of DC voltage in Fig.6, it’s inappropriate to claim that TiN/HfO2

RRAM system in this paper is purely unipolar. Based on the discussion about switching polarity above,

polarity of switching is not an intrinsic property of the metal oxides but an interfacial effect between

electrode and oxide. Switching behaviors will be different if a bias with opposite direction is applied

(negative voltage is applied on the tip). When TiN serves as anode in this case, oxygen ions will migrate

toward TiN and an interfacial layer of TiON may be generated as shown in Fig.11 and traps oxygen ions.

In this way, TiN/HfO2 with AFM tip as cathode may behave like a bipolar RRAM. In addition, if bipolar

switching is achieved in the opposite voltage polarity, location and shape of oxygen-deficient regions may

be different. Further characterization could be done to verify if this region is near bottom electrode (based

on their electron injection model) and if TiON interfacial layer acatually exists. Therefore, DC sweep in

both directions and further characterization should be done to fully illustrate electrical properties and

make possible comparison of oxygen-deficient regions.

3.2. Oxygen Vacancy Generation Mechanism

14

As shown in Fig.12(b), an oxygen-deficient region near top electrode is observed with extension of

20nm in the conductive HfOx area. In my opinion, this oxygen-deficient region cannot be attributed to the

electroforming process based on current data for the following two reasons.

Fig.12 a) STEM-HAADF image of the conductive region; STEM-EELS b) oxygen chemical map (c) nitrogen

chemical map (d) titanium chemical

Firstly, there’s no comparison of the observed oxygen-deficient region with the identical position in

as fabricated HfO2 sample and in HRS sample (after RESET). Theoretically, in pristine sample, there

should be very few oxygen vacancies; and in HRS sample, oxygen-deficient region should be somehow

ruptured or disappear. Without information about development of oxygen vacancies during switching,

there’re many possibilities to explain the observed phenomena. For example, oxygen-deficient region

may be introduced by fabrication before electroforming, and does not even change during switching.

Secondly, shape and location of this oxygen-deficient region are not supported by any proposed

physical mechanisms so far. For metal oxide RRAM, unipolar switching can be explained by fuse-

antifuse mechanism mentioned in Section 1.3 and 3.1. Growth of conducting nanofilaments can be

anticipated based on above mechanism:

1. Growth or accumulation of oxygen vacancies has to start from the cathode along the direction of

applied electric field for two reasons. Firstly, electrons, needed by metal to accommodate oxygen

deficiency, are injected from cathode. Secondly, positively charged oxygen vacancies are

15

attracted by cathode along electric field. This means that even though some vacancies are

generated elsewhere, they will eventually migrate to the cathode and pile up to the anode.

2. As filament is closer to anode, it becomes weaker, because generation of defects is highly field

related. Initially, oxygen vacancies will accumulate near cathode with a large lateral dispersion

because the electric field is almost uniform so is the generation rate. After several layers of

defects are piled up, resistance of the incomplete filament is negligible compared to the dielectric

oxide. All the applied voltage will drop on this shorter ruptured region, so the electric field is

increased, confining generation of vacancies in this narrow area. Finally, a complete filament is

formed and current through the oxides increases dramatically, further expansion of this filament

near anode or growth of a new filament will be suppressed by the applied current compliance.

3. Rupture of filaments happens near the anode for two reasons. One is that due to the shape of

filament with thinner end near anode, current density reaches maximum at this end and thus the

temperature. Migration of oxygen ions is thermally activated so recombination is strongest in this

end. Another reason is that oxygen ions are stored near anode as non-lattice atoms (perhaps

trapped in an interfacial layer) after SET or forming process [31]. Interface of anode/oxides

serves as an “oxygen reservoir” to supply oxygen ions to the defects. Therefore, oxygen ions will

be first captured by oxygen vacancies near anode, leading to the rupture of filament.

These predictions are verified by HRTEM image of nanofilaments in TiOx-based RRAM shown in

Fig.13 [2] and many other papers [23, 32-33].

Fig.13 (adapted from Ref.13) nanofilaments of Magnéli phase in TiOx-based RRAM (negative bias is applied to TE

and BE serves as anode) when the film is a) after SET f) after RESET.

16

Now let’s take a look at the observed oxygen-deficient region in this paper. At first, it grows laterally

instead of vertically along the electric field. Secondly, there’s no obvious conical shape of the so called

filament. Thirdly, no obvious oxygen accumulated interfaces are found, even near the protrusion. On the

contrary, areas near cathode (bottom electrode) seem to contain richer oxygen than the area near anode.

Last but not least, oxygen vacancies only appear in the top part of the oxide layer, they don’t even

connect the two electrodes.

Possible conduction mechanism of this incomplete “filament” will be discussed in Section 3.3, here

we just verify its physical origins. In this paper, the missing part of “filament” is explained by hot electron

injection model of Bersuker et al [34]. It states that electron injected from cathode moves through grain

boundaries. The intersection of three grain boundaries will be critical point where generation of vacancies

starts. I think this explanation is not convincing. Firstly, generation rate of oxygen and its relation with

grain boundaries are not fully illustrated in the paper. Generation of oxygen vacancy is schematically

shown in Fig.14.

Fig.14 Schematic view of oxygen vacancy generation in electric filed

In lattice, oxygen atom can be considered as trapped in a barrier with height E0 and lattice constant a

as shown in left. At the presence of an electric field, barrier height E0 is decreased by qEa in one direction.

Therefore, the generation possibility p can be expressed as:

This equation shows that generation of oxygen vacancies will start at the region that has the lowest E0

and strongest E, which is satisfied by the grain boundary.

E0

a

E0

a E

0 - qEa

Oxygen in lattice point Oxygen vacancy Oxygen ion

Electric Field E

17

Secondly, supported by the statement in [35], once a positively charged oxygen vacancy is generated

in the intersection with the highest electric field, it will directly drift to cathode due to ion migration

thermally accelerated by electrons. There’s one grain boundary vertically towards cathode shown in

Fig.10 a). Oxygen vacancies generated near grain boundary will drift along this direction and accumulate

near cathode. Therefore, the proposed electron injection model in the paper is not consistent with the

observed absence of oxygen vacancy at bottom.

3.3. Oxygen Vacancy Conduction Mechanism

To prove conducting nanofilaments are composed of oxygen vacancies and directly related to the

observed oxygen-deficient region, besides physical origins of defects generation, their roles in electron

conduction must be investigated. Based on the oxygen-deficient area shown in Fig.8 a) and Fig.12 b),

oxygen vacancies in HfO2-x don’t constitute a complete path connecting top and bottom electrode, making

its ability to transport electrons highly doubtful. Therefore, their conduction mechanism in this special

condition must be verified.

Fig.15 (adapted from Ref.36) Possible electron conduction mechanism in oxide-based RRAM

As shown in Fig.15 [36], there’re many possible mechanisms of electron conduction in oxide-based

RRAM (typically n-type conducting): Schottky emission (1), Fowler–Nordheim (F–N) tunneling (2),

direct tunneling (3), and trap-assisted tunneling. Trap-assisted tunneling can be further categorized as

tunneling from cathode to traps (4); emission from trap to conduction band (5), which is essentially the

18

Poole–Frenkel emission; F–N-like tunneling from trap to conduction band (6); trap to trap hopping or

tunneling (7), and tunneling from traps to anode (8). Among these possibilities, electrons would seek the

least resistive paths among all the possibilities. For LRS conduction, many reports show linear or Ohmic

I-V. While for HRS, as illustrated in [36], trap-assisted tunneling is used to model electron conduction.

In this paper, although authors claim that oxygen vacancies are known to be related to localized states

within the HfO2 band gap, there’s no clear data for LRS or HRS conduction I-V characteristics. In Fig.7

a), the reading curve of LRS seems to be non-linear with a large LRS resistance. The distance between

cathode and oxygen-deficient region is found to be around 10nm in Fig.8 a). It’s possible that there’re

isolated vacancies in the conducting path. So trap-assisted tunneling is likely to happen. However,

without further electrical characterization including temperature dependence measurement, conductivity

of oxygen-deficient region cannot be verified. Another possible explanation might be that under electrical,

chemical and structural characterization (especially when the bottom electrode is biased at 1V for SSRM),

oxygen vacancies near bottom electrode might be recombined. After characterization, another DC sweep

should be done to make sure this sample still remains in LRS.

3.4. Amorphization in Resistive Switching

As shown in Fig. 10a), amorphousness appears in areas near two electrodes and partial amorphization

can be observed in oxygen-deficient region. In the paper, amorphization near bottom electrode is

attributed to local heating and oxygen loss. However, in Fig.12 b), we can’t observe any oxygen

deficiency near bottom electrode. As for local heating, it is believed to cause amorphousness if fast

cooling is accompanied. In DC sweep, fast cooling process is impossible. In addition, the author claim

amorphous character on the top side is due to oxygen accumulation. As already mentioned in Section 3.2,

no oxygen-rich region is observed near top electrode in Fig.12 b).

It's worth noting that in HRTEM image of SET sample for TiOx-based RRAM does not show

obvious amorphousness near electrodes as shown in Fig.16 [2]. Admittedly, fabrication and structural

properties of TiO2 and HfO2 are different in these two papers. At least, one thing for sure is that this

observed amorphization is not responsible for resistive switching and caused by other reasons.

Amorphous character in HfO2 system may be introduced by fabrication process (350°C ALD) or caused

by intrinsic properties of HfO2.

19

Fig.16 (adapted from Ref.2) HRTEM image of the sample after SET in TiOx system.

4. SUGGESTIONS FOR FUTURE WORK

4.1. Verification of Filamentary Effect

Filamentary effect of HfOx-based RRAM is not fully verified in this paper. The conductive region

shown in SSRM result is around 100nm x 100nm, which is similar to the contact of the tip (with radius

20-50nm). To prove switching behavior is related to a confined, filamentary effect rather than a

homogeneously distributed one, following methods could be used.

RRAM devices with top electrode, such as TiN/HfO2/Pt stack, should be fabricated using similar

procedure. After electroforming, STEM-EELS is conducted for a large area of the film to check whether

oxygen vacancies are localized in a small area or dispersed in the whole oxide. Compliance current

dependence could also be tested, because larger compliance is believed to result in more filaments with

larger width.

LRS without area dependence is another demonstration of filamentary effect (electrons conduction is

through the filament not the whole film). HfOx-based devices with different areas of top electrode could

20

be fabricated with the same thickness of oxides. Under the same DC sweep voltage, resistance in LRS

could be compared to check the area dependence.

According to switching mechanisms in Section.2, filamentary type RRAM is always related to ionic

diffusion or drift effect while other RRAM always has electron based switching. In this sense,

filamentary effect could be verified by following two tests.

1. Temperature dependence tests to check the change of switching parameters, such as

switching voltage or current, switching speed and corresponding resistance because ionic

migration and defects generation are supposed to be thermally activated.

2. Fast switching test to investigate switching time dependence on voltage. Ionic migration is

much slower than electron transport, so electron based switching is supposed to be voltage

independent.

4.2. Identification of Conducting Nanofilaments

As what is pointed out in Section.3, this paper doesn't provide enough evidence to identify oxygen-

deficient region as the conducting nanofilaments. To verify this, following methods could be used in

addition to the current procedure.

1. STEM-EELS should be firstly performed for pristine HfO2 sample in order to demonstrate

the absence of oxygen vacancies before switching.

2. To confirm the existence of oxygen-deficient region and oxygen-rich region anticipated by

mechanism, oxygen profile in vertical direction should be done besides lateral one. After

mapping oxygen concentration of the conductive region, I-V characteristics should be tested

to check if this sample is still in LRS.

3. Switching the sample to HRS and do STEM-EELS again in the same region (this region is

easy to be located by the protrusion mentioned in Fig.7). After this, check the resistance state

again by DC sweep.

4. More precise characterization could be done by in-situ switching as in [2]. Instead of STEM-

EELS, HRTEM and corresponding diffraction should be used.

5. In this way, development of oxygen vacancies during switching is demonstrated. To further

verify the their conductivity, models fitting for conduction mechanisms mentioned in Section

3.3 should be done with I-V curves in LRS or HRS and corresponding temperature

dependence data.

21

5. CONCLUSION

The main reference deals with HfOx-based RRAM and presents chemical, electronic and structural

properties of the conductive region by C-AFM, STEM-EELS and HRTEM. Identification of conducting

nanofilaments as the observed oxygen-deficient region is mainly critiqued in this manuscript. Several

related issues for metal oxide RRAM are addressed:

1. Polarity of switching voltage is related to electrode/oxide interface. Unipolar and bipolar

switching characteristics of TiN/HfO2 system are not fully investigated in this work.

2. Resistive switching of metal oxide based RRAM is due to ionic migration and local heating

induced construction and rupture of filaments. The shape and location of observed oxygen-

deficient region are not consistent with the prediction of these mechanisms.

3. Conduction of oxygen vacancies in observed locations are highly doubtful without further

conduction data.

4. Physical origins of amorphization are not consistent with the observations in this paper.

Amorphous character of the oxides is explained by comparing similar material system in

another paper.

Based on these critiques, this work lacks supportive evidence to demonstrate its statement. Further

possible work to verify and identify conducting filament in metal oxide RRAM is suggested at last.

22

REFRENCE

1. Calka, P., et al. "Chemical and structural properties of conducting nanofilaments in TiN/HfO2-based resistive

switching structures." Nanotechnology 24.8 (2013): 085706.

2. Kwon, Deok-Hwang, et al. "Atomic structure of conducting nanofilaments in TiO2 resistive switching

memory." Nature nanotechnology 5.2 (2010): 148-153.

3. Perevalov, T. V., et al. "Atomic and electronic structure of amorphous and crystalline hafnium oxide: X-ray

photoelectron spectroscopy and density functional calculations." Journal of Applied Physics 101.5 (2007):

053704-053704.

4. Wilk, Glen D., Robert M. Wallace, and J. M. Anthony. "High-κ gate dielectrics: Current status and materials

properties considerations." Journal of applied physics 89.10 (2001): 5243-5275.

5. Kukli, Kaupo, et al. "Comparison of hafnium oxide films grown by atomic layer deposition from iodide and

chloride precursors." Thin Solid Films 416.1 (2002): 72-79.

6. Hausmann, Dennis M., et al. "Atomic layer deposition of hafnium and zirconium oxides using metal amide

precursors." Chemistry of materials14.10 (2002): 4350-4358.

7. Goux, Ludovic, et al. "Evidences of oxygen-mediated resistive-switching mechanism in TiN/HfO2/Pt

cells." Applied Physics Letters 97.24 (2010): 243509-243509.

8. Govoreanu, B., et al. "10× 10nm2 Hf/HfOx crossbar resistive RAM with excellent performance, reliability and

low-energy operation." Electron Devices Meeting (IEDM), 2011 IEEE International. IEEE, 2011.

9. Sawa, Akihito. "Resistive switching in transition metal oxides." Materials Today 11.6 (2008): 28-36.

10. Lee, Byoungil, and H-SP Wong. "NiO resistance change memory with a novel structure for 3D integration and

improved confinement of conduction path." VLSI Technology, 2009 Symposium on. IEEE, 2009.

11. Kim, Kyung Min, Byung Joon Choi, and Cheol Seong Hwang. "Localized switching mechanism in resistive

switching of atomic-layer-deposited TiO2 thin films." Applied physics letters 90.24 (2007): 242906-242906.

12. Chen, Y. S., et al. "Highly scalable hafnium oxide memory with improvements of resistive distribution and read

disturb immunity." Electron Devices Meeting (IEDM), 2009 IEEE International. IEEE, 2009.

13. Cho, Suyeon, et al. "Effect of Cr concentration on resistance switching in Cr-doped SrZrO films and surface

accumulation of Cr ions." Journal of Applied Physics 108 (2010): 103716.

14. Terabe, K., et al. "Quantized conductance atomic switch." Nature 433.7021 (2005): 47-50.

15. Bandyopadhyay, Anirban, and Amlan J. Pal. "Large conductance switching and memory effects in organic

molecules for data-storage applications."Applied physics letters 82.8 (2003): 1215-1217.

16. Wong, H-SP, et al. "Metal–oxide RRAM." Proceedings of the IEEE 100.6 (2012): 1951-1970.

17. Waser, Rainer, and Masakazu Aono. "Nanoionics-based resistive switching memories." Nature materials 6.11

(2007): 833-840.

18. Kim, Kyung Min, et al. "Anode-interface localized filamentary mechanism in resistive switching of TiO2 thin

films." Applied physics letters 91.1 (2007): 012907-012907.

23

19. Yang, Yuchao, et al. "Observation of conducting filament growth in nanoscale resistive memories." Nature

Communications 3 (2012): 732.

20. Yao, Jun, et al. "In situ imaging of the conducting filament in a silicon oxide resistive switch." Scientific reports

2 (2012).

21. Wei, Z., et al. "Demonstration of high-density ReRAM ensuring 10-year retention at 85 C based on a newly

developed reliability model." Electron Devices Meeting (IEDM), 2011 IEEE International. IEEE, 2011.

22. Yu, Shimeng, and H-SP Wong. "A phenomenological model for the reset mechanism of metal oxide

RRAM." Electron Device Letters, IEEE 31.12 (2010): 1455-1457.

23. Chen, Y. S., et al. "Microscopic mechanism for unipolar resistive switching behaviour of nickel

oxides." Journal of Physics D: Applied Physics 45.6 (2012): 065303.

24. Gao, B., et al. "Oxide-based RRAM: Unified microscopic principle for both unipolar and bipolar

switching." Electron Devices Meeting (IEDM), 2011 IEEE International. IEEE, 2011.

25. Kim, Yong-Mu, and Jang-Sik Lee. "Reproducible resistance switching characteristics of hafnium oxide-based

nonvolatile memory devices." Journal of Applied Physics 104.11 (2008): 114115-114115.

26. Cagli, C., et al. "Experimental and theoretical study of electrode effects in HfO2 based RRAM." IEDM Tech.

Dig 28.1 (2011).

27. Yu, Shimeng, et al. "A HfOx Based Vertical Resistive Switching Random Access Memory Suitable for Bit-

Cost-Effective Three-Dimensional Cross-Point Architecture." ACS nano (2013).

28. Goux, Ludovic, et al. "On the gradual unipolar and bipolar resistive switching of TiN\ HfO2\ Pt memory

systems." Electrochemical and Solid-State Letters13.6 (2010): G54-G56.

29. Jeong, Doo Seok, Herbert Schroeder, and Rainer Waser. "Coexistence of Bipolar and Unipolar Resistive

Switching Behaviors in a Pt∕ TiO2∕ Pt Stack."Electrochemical and solid-state letters 10.8 (2007): G51-G53.

30. Goux, L., et al. "Coexistence of the bipolar and unipolar resistive-switching modes in NiO cells made by

thermal oxidation of Ni layers." Journal of Applied Physics 107.2 (2010): 024512-024512.

31. Zhou, P., et al. "Role of TaON interface for CuxO resistive switching memory based on a combined

model." Applied Physics Letters 94.5 (2009): 053510-053510.

32. Kim, Kyung Min, and Cheol Seong Hwang. "The conical shape filament growth model in unipolar resistance

switching of TiO2 thin film." Applied Physics Letters 94.12 (2009): 122109-122109.

33. Russo, Ugo, et al. "Self-accelerated thermal dissolution model for reset programming in unipolar resistive-

switching memory (RRAM) devices."Electron Devices, IEEE Transactions on 56.2 (2009): 193-200.

34. Bersuker, G., et al. "Metal oxide resistive memory switching mechanism based on conductive filament

properties." Journal of Applied Physics 110.12 (2011): 124518-124518.

35. Yang, J. Joshua, et al. "Metal oxide memories based on thermochemical and valence change mechanisms." MRS

bulletin 37.02 (2012): 131-137.

36. Yu, Shimeng, Ximeng Guan, and H-S. Philip Wong. "Conduction mechanism of TiN/HfOx/Pt resistive

switching memory: A trap-assisted-tunneling model."Applied Physics Letters 99 (2011): 063507.

Chemical and structural properties of conducting nanofilaments in TiN/HfO2-based resistive

switching structures

This article has been downloaded from IOPscience. Please scroll down to see the full text article.

2013 Nanotechnology 24 085706

(http://iopscience.iop.org/0957-4484/24/8/085706)

Download details:

IP Address: 165.123.214.77

The article was downloaded on 12/02/2013 at 04:05

Please note that terms and conditions apply.

View the table of contents for this issue, or go to the journal homepage for more

Home Search Collections Journals About Contact us My IOPscience

IOP PUBLISHING NANOTECHNOLOGY

Nanotechnology 24 (2013) 085706 (9pp) doi:10.1088/0957-4484/24/8/085706

Chemical and structural properties ofconducting nanofilaments inTiN/HfO2-based resistive switchingstructuresP Calka, E Martinez, V Delaye, D Lafond, G Audoit, D Mariolle,N Chevalier, H Grampeix, C Cagli, V Jousseaume and C Guedj

CEA, LETI, MINATEC Campus, 17 rue des Martyrs, 38054 Grenoble Cedex 9, France

E-mail: eugenie.martinez@cea.fr

Received 10 September 2012, in final form 21 December 2012Published 5 February 2013Online at stacks.iop.org/Nano/24/085706

AbstractStructural, chemical and electronic properties of electroforming in the TiN/HfO2 system areinvestigated at the nanometre scale. Reversible resistive switching is achieved by biasing themetal oxide using conductive atomic force microscopy. An original method is implemented tolocalize and investigate the conductive region by combining focused ion beam, scanningspreading resistance microscopy and scanning transmission electron microscopy. Resultsclearly show the presence of a conductive filament extending over 20 nm. Its size and shape ismainly tuned by the corresponding HfO2 crystalline grain. Oxygen vacancies together withlocalized states in the HfO2 band gap are highlighted by electron energy loss spectroscopy.Oxygen depletion is seen mainly in the central part of the conductive filament along grainboundaries. This is associated with partial amorphization, in particular at both electrode/oxideinterfaces. Our results are a direct confirmation of the filamentary conduction mechanism,showing that oxygen content modulation at the nanometre scale plays a major role in resistiveswitching.

(Some figures may appear in colour only in the online journal)

1. Introduction

Resistive random access memories (RRAM) are promisingcandidates for the next generation of non-volatile memories.They are of potential interest to achieve high integrationdensity, high speed of operation and low power consump-tion [1, 2]. The RRAM consists of a metal–insulator–metal(MIM) structure that exhibits a reversible change of resistanceunder biasing. At a specific threshold voltage, known as theforming voltage, the resistance evolves from the pristine highresistive state (HRS) to a low resistive state (LRS). Duringthis so-called electroforming step, a soft dielectric breakdownis assumed to occur and resistive switching is activated. Acurrent limitation is enforced to reduce damage in the MIMstructure and avoid permanent dielectric breakdown. Succes-sive alternating switching cycles (>106) between both LRS

and HRS states can then be performed, these may be eitherunipolar and bipolar depending on whether the switchingdirection depends on the polarity of the applied voltage.

The resistive switching phenomenon has been observedin a wide variety of materials, such as the transition metaloxides (NiO [3], TiO2 [4], TaOx [5], WOx [6], ZrO2 [7]).Among these, HfO2 has attracted much attention [8–11]because it offers low fabrication costs and good compatibilitywith the conventional metal oxide semiconductor field effecttransistors (MOSFETs) fabrication process. However, theswitching mechanism is still not fully understood. It isbelieved that an active region is formed in a confinedarea between the top and bottom electrodes, the so-calledconductive filament (CF) [1, 2]. To improve our knowledgeof the physical mechanism involved in resistive switching,a direct observation of this conductive filament is required,

10957-4484/13/085706+09$33.00 c© 2013 IOP Publishing Ltd Printed in the UK & the USA

Nanotechnology 24 (2013) 085706 P Calka et al

but this is challenging. Indeed, the conductive region isthought to be unique and randomly located between theelectrodes, thus making its localization and observationvery difficult. Resistive switching has been achieved usingconductive atomic force microscopy (C-AFM) [12–16],offering the possibility to analyse individual conductivefilaments. Recently, direct observations of conductive regionshave been made on NiO [17, 18], TiO2 [19] and Ta2O5 [20],but regarding HfO2, nothing has been reported yet in theliterature to our knowledge.

In this study, a specific protocol has been implementedto address this issue. Conductive regions are created in HfO2and then localized to investigate the structural and chemicalproperties after the electroforming process. We focus here onthis first critical step, which defines both the CF properties andthe switching characteristics of resistive memories. Workingwithout a permanent top electrode allows easier localizationof the conductive path. The oxide is locally switched at thenanometre scale with an AFM nanoprobe (top electrode). Theconductive region is then localized by scanning spreadingresistance microscopy (SSRM). A direct observation is madeby scanning transmission electron microscopy (STEM) toinvestigate the chemical and electronic changes involved inthe switching mechanism. Significant results are obtainedregarding the properties of the conductive filaments. Answersregarding the chemical composition, crystallinity, electronicstructure and morphology are obtained at the nanometre scale.Our results are in agreement with the filamentary conductionmechanism based on the segregation of oxygen vacancies.This is, to our knowledge, the first experimental evidence ofsuch conducting nanofilaments in HfO2-based structures.

2. Experiment

A 20 nm thick TiN layer is sputtered on top of a Si(100)wafer covered by a thin (2 nm) native SiO2 film. De-oxidationof the TiN surface is done by chemical etching (HF1%) before deposition of a 10 nm thick HfO2 layer byatomic layer deposition (ALD) at 350 ◦C using HfCl4 andH2O. Both the TiN electrode and HfO2 oxide have apolycrystalline structure, as already reported elsewhere [21].The stoichiometry of the pristine HfO2 oxide is measuredby Rutherford back-scattering spectroscopy (RBS). The O/Hfratio is found to be 2.3, with an uncertainty of 5%, showingthat the initial oxide is presumably oxygen rich, probablybecause of residual –OH groups due to the H2O-based ALDgrowth.

A specific protocol in five steps has been developedin order to produce and localize a conductive region inthe TiN/HfO2 system. A partial cross-section (CS) is firstprepared by focused ion beam (FIB) using a gallium beamin a dual beam Helios system from FEI. Two parallel trenches(1 × 10 µm2) of approximately 2 nm depth are designed onboth sides of a pristine HfO2 area (1× 10 µm2). The depth iskept very low to preserve the electrical contact of the bottomTiN electrode. Resistive switching is then performed in thisdelimited area to facilitate further localization and observationof the conductive area.

Resistive switching is performed using AFM D3100equipment from Bruker in the conducting mode (C-AFM).The sample is previously heated at 120 ◦C for 12 hunder atmospheric pressure to remove moisture and surfacecontamination. Measurements are then performed in a glovebox with a constant N2 flux, i.e. with reduced humidity(<2% of H2O). These two precautions are taken to minimizewater dissociation and subsequent detrimental effects such asoxide regrowth [22], leading to an increase of the formingvoltage, or injection of charged species (OH−,H+), leading toparasitic leakage currents [23]. Highly doped diamond coatedSi tips from Nanosensors are used (Rtip = 5–8 k�). The tipradius is in the range 20–50 nm, depending of the shape ofthe diamond grain at the tip apex. The tip force is set in the0.1–0.5 µN range to ensure efficient electrical contact andlow mechanical pressure (<0.5 GPa). The resulting possibledamage is thus assumed to be rather weak. The TiN bottomelectrode is grounded whereas the tip is biased using an HP4155A parameter analyser. Resistive switching is performedwith the tip held at a fixed location while the voltage isswept. The current flowing through the oxide lies in the range10−14–10−3 A.

In order to localize the conductive region afterresistive switching, topography and conductivity maps aresimultaneously acquired in contact mode by SSRM on thecross-section. A positive voltage (1 V), which is below the set(Vset) and forming (Vforming) voltages, is applied to the TiNlayer while the tip is grounded. The resistance is measuredwithin the detection limits, from 104 to 1010 �. To enable adirect observation of the conductive region by STEM, the finalcross-section is prepared by FIB. Protective coatings (SiOCHand Pt) are deposited on top of the HfO2 partial CS beforethe lift-out. The lamella is prepared by thinning this section athigh energy (30 keV), with a final cleaning step performedat low energy (5 keV) to minimize surface amorphizationeffects.

Finally, the chemical composition of the conductiveregion is studied by electron energy loss spectroscopy(EELS) using the Cs probe and image corrected TitanUltimate Microscope from FEI, equipped with a QuantumGatan energy filter and a scanning module. STEM–EELSmeasurements are carried out with an acceleration voltageof 200 keV. Elemental maps are estimated by measuringthe absorption edge area, after a power law backgroundsubtraction. Edge area densities (at× cts nm−2) are estimatedusing the formula NITotal = Ik/σk, where N is the atomicconcentration, and Ik and σk are respectively the intensity andionization cross-section of the absorption edge. A chemicalsensitivity of 0.1 at.% is achieved. STEM–EELS chemicalmaps are recorded with an energy resolution of 1.6 eV anda spatial resolution of 0.5 nm. The relative lamella thicknesst/λ is calculated using the formula t/λ = ln(ITotal/IElastic),where ITotal and IElastic are respectively the intensity of thewhole spectrum and zero-loss peak, λ is the electron inelasticmean free path and t is the lamella thickness [24]. Additionalinformation about the structure and composition of the activeregion is obtained with Z-contrast imaging using a high-angleannular dark-field detector (HAADF) and high-resolutiontransmission electron microscopy (HR-TEM).

2

Nanotechnology 24 (2013) 085706 P Calka et al

Figure 1. (a) Experimental setup used for resistive switching of the TiN/HfO2 stack with a polarized conductive AFM diamond coated Sitip as a top electrode. (b) Corresponding electrical schematics. (c) I(Va) characteristics measured by biasing the AFM tip (Va = 4 V).(d) I(Ve) characteristics obtained by taking into account the tip resistance and effective voltage Ve seen by the TiN/HfO2 stack.

3. Results and discussion

3.1. Electrical properties of the TiN/HfO2 system

First, the resistive switching properties of the TiN/HfO2system are tested using the blanket Si/SiO2/TiN/HfO2sample. C-AFM measurements are performed by applying theanalyser voltage (Va ∼ 4 V) directly on the AFM conductivetip. The experimental setup and the corresponding equivalentelectrical circuit are displayed in figures 1(a) and (b). The tipresistance (Rtip) affects the electroforming process, especiallywhen high current densities flow through the oxide. Apotential drop is expected over the AFM tip, thus decreasingthe effective voltage (Ve) applied to the TiN/HfO2 stack. Notethat Rtip is in the 5–8 k� range, i.e. 2–3 orders of magnitudehigher than for a conventional metallic electrode.

Several successive voltage sweeps are performed in theunipolar mode by biasing the tip with positive voltages(figure 1(c)). During the first voltage sweep, the initial currentintensity flowing through the oxide is around 10 fA. At thethreshold voltage of 4.2 V (Vforming), the current intensityincreases dramatically. During this electroforming process,the insulating oxide evolves into a conductive state (LRS).The compliance current Icc is set at 500 µA to reduce damageinduced by high current densities flowing through the stack.During the second voltage sweep, the initial current (∼2 µA)indicates that the conductive state is persistent. The oxideis thus stabilized in the LRS state. When increasing thevoltage, the current intensity reaches the mA range, sinceno compliance current is applied. At a threshold voltage of3.9 V (Vreset), a drop of the current is observed (3 orders ofmagnitude). This is the so-called reset process, bringing the

oxide into a high resistive state (HRS). During the third sweep,the oxide resistance changes again from HRS to LRS. This isthe set process, occurring at a voltage of 1.3 V (Vset). For thesame reason as for the electroforming process, the compliancecurrent is also activated.

After taking into account the tip resistance, assumed tobe the same for the electroforming, reset and set processesand equal to 5 k�, the resulting I(Ve) curves are plotted infigure 1(d). The shape of the curves is modified only forhigh current densities, where the tip resistance correctioninduces a decrease of the effective voltage (Ve). The newthreshold voltages are measured with the methodologydescribed in a previous work [11]. Vforming and Vset are notaffected by the tip resistance, whereas Vreset decreases, thusleading to the expected behaviour regarding the unipolarswitching mode [25] (Vset higher than Vreset). Note thatfor quantitative measurements of the film conductivity, forinstance the resistance ratio between LRS and HRS states, aSchottky barrier effect should be considered at the tip/oxideinterface [26], which might introduce a significant resistanceand modify the Vreset voltage as well.

These results show that the TiN/HfO2/diamond (AFMtip) stack exhibits unipolar resistive switching. The resistancechange is reversible and successive set/reset processes can beactivated at the same bias polarity. The current intensity andforming voltage measured for the reset process are similar tothat of our integrated devices [27].

3.2. Electroforming and identification of the conductivefilament

Following the same C-AFM protocol, resistive switching isnow performed on the pristine HfO2 area contained in the

3

Nanotechnology 24 (2013) 085706 P Calka et al

Figure 2. (a) I(Ve) characteristics of the electroforming (bluesquares) and reading (red circles) steps performed on the HfO2partial cross-section with the conductive AFM tip before lift-out ofthe TEM lamella. Inset: SEM image of the HfO2 surface after theelectroforming/reading electrical stresses. (b) Topography andresistance maps measured by SSRM on the switched HfO2.

partial CS prepared by FIB, as described in section 2. Thecorresponding I(Ve) characteristic is plotted in figure 2(a).Once again, a compliance current of 500 µA is appliedduring the electroforming process. A second bias voltagesweep (up to 1 V) is performed subsequently, confirming thestability of the LRS (red rings). In the inset of figure 2(a),the SEM (scanning electron microscopy) image shows a darkcontrast corresponding to the electrically stressed region.The SEM contrast is usually attributed to a change of thesurface topography or to charge trapping. Here, this contrastis presumably related to the conductive region.

To confirm this assumption and clearly identify thisarea as being highly conductive, a precise localization ofthe conductive region is done at the nanometre scale byresistance mapping with SSRM before thinning and lift-outof the cross-section. The surface is scanned in contact modewhen biasing the TiN bottom electrode (1 V). Topographyand resistance maps are simultaneously acquired. Both imagesare plotted in figure 2(b). A conductive area together witha small protrusion is observed and perfectly correlated onboth maps. This protrusion has a resistance of 107 � at 1 V,therefore the current in this conductive region is 10−7 A.This value is in good agreement with the current intensitymeasured for the LRS state (see figure 2(a)), showing that theSSRM measurement does not modify the conductive region.This was also checked on a pristine HfO2 area (not shown),since no modifications were observed after doing SSRMin the same conditions. The conductive area is round andextends over 100 nm at the HfO2 surface. However, this sizemight be overestimated because of the relatively high contactarea (∼100 nm) between the tip and the HfO2 surface. Toget a better insight of the size, shape and composition ofthis conductive region, and thus of the physical phenomena

Figure 3. STEM–HAADF image measured for the conductiveregion identified in figure 2(b).

involved in resistive switching, the partial cross-section islifted-out and thinned to allow direct observation by STEM.

Figure 3 shows the STEM–HAADF image measured inthe region of interest. All the layers of the Si/SiO2/TiN/HfO2stack are identified by Z-contrast imaging. The conductiveregion is clearly visible, corresponding to a protrusion 3 nmin height. As seen on this image, the morphology of the HfO2layer has changed, whereas the underlying layers, includingthe TiN bottom electrode, are not modified. Dark contrastis observed at the left side of the conductive region. This isa thickness-related contrast, due to FIB milling, which wasmore important in the left region, yielding a thinner lamella inthe left part. This assumption is confirmed by the protectiveSiOCH and Pt being missing at the top of the lamella.

3.3. Chemical properties of the conductive filament

STEM–EELS measurements are performed in the protrudingregion mentioned in section 3.2 and displayed in the HAADFimage of figure 4(a). Oxygen, nitrogen and titanium maps,acquired by measuring the O–K (532 eV), N–K (401 eV)and Ti–L2,3 (456 eV) edges, are displayed in figures 4(b)–(d).These elemental maps are derived from the integration of thefull edges (532–560 eV for O, 396–421 eV for N, 456–481 eVfor Ti). In the oxygen map (see figure 4(b)), only the SiOCHand HfO2 layers exhibit high oxygen content, as expected.As discussed before, we again observe a thickness-relatedcontrast along the cross-section, corresponding to dark andbright blue contrasts on the left and right sides of the HfO2layer, respectively. Nevertheless, a localized oxygen-deficientregion (dark contrast inside the red circle) is highlighted inthe protruding HfO2 area. This region seems mainly localizedat the top of the HfO2 layer, with an extension of 20 nm inthe lateral direction. Ti and N maps show that no diffusion ofmetallic atoms from the bottom electrode is observed insidethe oxide. Furthermore, there is no migration from the topelectrode material (diamond coated silicon tip) or from carbonsurface contamination in the HfO2 layer, as confirmed by Cand Si maps (not shown).

Further information is obtained by a detailed analysisof the conductive region. Oxygen and thickness profiles areplotted in the horizontal direction, along the oxygen-deficient

4

Nanotechnology 24 (2013) 085706 P Calka et al

Figure 4. (a) STEM–HAADF image of the conductive region:zoom on the HfO2 layer. (b)–(d) STEM–EELS chemical maps ofthe conductive region: (b) oxygen map, (c) nitrogen map,(d) titanium map.

and oxygen-rich regions (see figure 5). A drop of the oxygenconcentration is clearly seen and related to a modification ofthe oxide after electroforming. The lamella thickness, alsoplotted in figure 5, is constant along the same profile (see redline in figure 5(b)). The oxygen drop is thus not correlated toa decrease of the lamella thickness, but really representativeof the electroforming. The oxygen loss is estimated to be ashigh as 55%± 15% in some points. We thus demonstrate herethat the electroforming step creates an oxygen-poor regionin the Hf oxide, labelled as HfO2−x conductive region inthe rest of the paper. Our results clearly show that oxygencontent modulation at the nanometre scale plays a major rolein resistive switching of HfO2-based structures.

3.4. Electronic properties of the conductive filament

Additional information about the electronic properties of theHfO2−x conductive region is obtained by a detailed analysisof the oxygen K-edge (532 eV) EELS spectra measuredat each pixel of the oxygen map. Spectra extracted in theoxygen-deficient (HfO2−x) and oxygen-rich (HfO2) regionsare plotted in figure 6(a).

The reference spectrum, extracted from the oxygen-richregion, is characterized by the typical double peak structure(contributions I and II at 536 and 539 eV), already observedfor crystalline HfO2 [28]. These near-edge fine structuresare attributed to the excitation of a O 1s core level electroninto an unoccupied state in the conduction band of HfO2,mainly composed by cation Hf 5d states [29]. The oxygenK-edge spectrum is explained by an hybridization of Hf 5dand O 2p states, since the s→ d transition is forbidden by

Figure 5. (a) Zoom on the oxygen-deficient region in HfO2corresponding to the red circle of figure 4. (b) Oxygen concentrationand relative thickness (t/λ) of the lamella, plotted along the dashedline of figure 5(a).

the dipole selection rule [30]. The two main peaks are relatedto the so-called crystal-field splitting via this Hf 5d/O 2phybridization. Indeed, energy splitting of Hf 5d orbitals isknown to form the eg and t2g sub-bands [31, 32] because of theelectrostatic field created by the oxygen atoms around each Hfatom.

Significant differences are observed for the oxygenK-edge measured in the HfO2−x conductive region. First, adecrease of the intensity of the global spectrum is observedand presumably related to the reduced number of oxygenatoms in this region. Second, the shape is modified, withthe occurrence of a shoulder (III) at the low-energy sideof component I, which is worth noting although veryweak, together with an additional contribution (IV) betweencomponents I and II. The shoulder on the low-energyside of component I could be related to the creation ofoxygen vacancies which induce localized states [25, 33]within the band gap of HfO2. This feature has already beenobserved by Baik et al [33] at HfO2 grain boundaries, knownto be oxygen deficient and confirmed by first-principlescalculations including O vacancies in HfO2 supercells [33].This additional structure is explained by the excitation ofa O 1s core level electron into localized states within theband gap. This transition requires less energy than a transitioninto the conduction band. Thus, this structure appears shiftedtowards lower energies. The additional contribution (IV) isalso probably related to oxygen vacancies. Similarities can bemade with Wilk’s results [28], showing that the gap between

5

Nanotechnology 24 (2013) 085706 P Calka et al

Figure 6. (a) EELS oxygen K-edges spectra of the oxygen-rich(HfO2) and oxygen-poor (HfO2−x) regions acquired in STEMmode. (b) EELS oxygen K-edges spectra measured in the TiNbottom electrode and SiOCH upper layer. All the measurements aredone in regions where the lamella thickness is the same (see blacksquares on figure 7).

peaks I and II is filled for as-deposited HfO2 compared toannealed samples, i. e. with increasing oxygen vacancies andthus differing oxygen coordination around metal Hf ions. Thistrend has also been reported by Ostanin et al [34] when dopingzirconia with trivalent yttrium (YSZ for yttria-stabilizedzirconia), which introduces oxygen vacancies for chargeneutrality reasons. Additional structures are seen to appear inbetween components I and II for calculated ELNES spectraof such materials [35]. Note that, in our case, this additionalstructure is not attributed to electron beam damage, as shownby the spectrum measured for the oxygen-rich region, whichis not influenced by such a process. Finally, the initial doublepeak structure also evolves for the oxygen-deficient region.Indeed, the width and separation of peaks I and II aredifferent and might also be related to oxygen vacancies. Thebroadening of these structures has already been observed byStemmer et al [29] for amorphous Hf1−xAlxOy oxides andassigned to local atomic disorder. In our case, the expectedmovement of oxygen atoms probably induces such localdisorder as well as changes in the point defect chemistry. Theenergy difference between components I and II is also knownto vary with the material composition [36] and in particularwith oxygen loss in our case. This is confirmed by Ostanin’sresults [34, 35], reporting a shift of the second peak (II)

Figure 7. Images of the a–d coefficients used to model the oxygenK-edges EELS spectra measured at each pixel of the STEM–EELSchemical maps. Reference spectra for each compound (HfO2,HfO2−x, TiN and SiOCH) were extracted from regions marked bysmall black squares.

towards higher energies when increasing the concentration ofO vacancies in YSZ.

The near-edge fine structures of the oxygen K-edge areused to obtain information about the chemical and electronicproperties of the conductive region at the nanometre scale.EELS spectra measured at each pixel of the oxygen map (seefigure 4) are modelled using the minimum linear least squares(MLLS) routine of the Gatan digital micrograph software.Reference oxygen K-edge spectra are extracted from theoxygen-rich (HfO2) and oxygen-poor (HfO2−x) regions (seefigure 6(a)) as well as from the neighbouring TiN and SiOCHlayers (see figure 6(b)). No oxygen edge is measured for thebottom TiN electrode, confirming the absence of an interfacialTiON layer. The oxygen K-edge measured for SiOCH has asingle peak at higher energies (540 eV). A linear combinationof these spectra is used to reproduce the O K-edge (SO–K)

measured at each pixel of the O map, following the equation:

SO–K = aSO–K(HfO2)+ bSO–K(HfO2−x)

+ cSO–K(SiOCH)+ dSO–K(TiN) (1)

where SO–K(HfO2), SO–K(HfOx), SO–K(SiOCH), andSO–K(TiN) are the reference oxygen K-edges spectra. Theresulting images are plotted in figure 7 for the regionof interest. These coefficients highlight the location ofunoccupied electronic states into the conduction band for eachcompound (HfO2, HfO2−x, TiN and SiOCH).

The first two images (see figures 7(a) and (b)) arecomplementary and exhibit respectively low and highcontrasts in the conductive region. This result is consistentwith the existence of an oxygen-poor region, as discussed in

6

Nanotechnology 24 (2013) 085706 P Calka et al

Figure 8. (a) High-resolution TEM image of the HfO2−x conductive region. The grain boundaries are identified by dashed lines and theamorphous regions are labelled (A). (b) Fast Fourier transformed micrographs of selected areas (left HfO2 crystalline grain and amorphousarea labelled A2 of HfO2−x) of the high-resolution TEM image together with the corresponding simulated diffraction patterns formonoclinic HfO2.

section 3.3. The conductive region is characterized by a highdensity of defects such as oxygen vacancies, thus modifyingthe HfO2 electronic properties after the electroforming step.Oxygen vacancies are known to be related to localized stateswithin the HfO2 band gap. These intermediate band gapstates might favour electron conduction along the oxide,thus explaining the resistive switching. Figures 7(c) and(d) show the location of unoccupied states relative to theTiN and SiOCH oxygen K-edges. They are located ineach corresponding layer, thus confirming the absence ofinter-diffusion phenomena.

3.5. Structural properties of the conductive filament

When looking at the STEM–EELS oxygen map, the oxygenvacancies and thus the conductive region seem to be mainlylocated at the top of the HfO2 layer. More precise informationabout the CF location, shape and crystalline structure isobtained by HR-TEM. The image measured when zoomingon the conductive region is presented in figure 8(a). Theposition, dimension and shape of the main central grain aresimilar to the conductive region identified in sections 3.3 and3.4. This region is thus identified as an individual conductivefilament, which is not limited to the grain boundaries (seedashed lines) but extended over one specific grain 20 nmin size. This observation is in agreement with the resultsreported by Bersuker et al [37], indicating a spatial extensionof the leakage path at the deca-nanometre scale. This maincentral grain is characterized by a different crystallographicorientation compared to the left and right HfO2 grains. Thisis also the case for the TiN grain, which is perfectly alignedwith the CF. This particular configuration might favour thegrowth of the CF at this specific location. The properties ofthe CF are therefore probably strongly related to the structureof both materials and thus to the corresponding depositiontechniques. Fast Fourier transformed (FFT) micrographs

of selected areas of this HR-TEM image are presentedin figure 8(b) together with their corresponding simulateddiffractions patterns obtained using the JEMS software andassuming a HfO2 monoclinic structure. The FFT micrographextracted from the central part of the CF, labelled A2,highlights additional rings compared to the one extracted froma neighbouring HfO2 grain. This indicates partial localizedamorphization of the CF. Similar results are obtained for theregions labelled A1 and A3.

The CF is known to grow parallel to the electric fieldapplied between both electrodes and to be preferentiallycreated at the grain boundaries (GB), as stated by Lanzaet al [38]. Our results show that the CF is not fully extendedalong the vertical direction but some part is missing at thebottom side. This CF was probably initiated along the smallvertical grain boundary at the bottom side where the electricfield and conductivity is much higher. Our results support thehot electron injection model of Bersuker et al [37], whereelectrons are injected from the cathode into the oxide throughgrain boundaries. The intersection point between the threegrain boundaries seen on the HR-TEM image must have beena critical point to favour the spatial extension of the CFwhile growing along the oxide. At this point, the injectedelectrons start to create defects such as oxygen vacancies.This is confirmed when looking at figures 7(a) and (b),where we clearly see that most of the oxygen vacancies arecreated at this point and inside the main central grain. Oxygenions then probably migrate towards the anode under thehigh electric field. This conductive region remains crystallinedespite some local atomic disorder related to oxygen loss.This is confirmed by the diffraction patterns extracted fromselected areas of the CF, showing partial amorphization (seefigure 8(b)). On the central and bottom side (A2 and A3),amorphization is probably due to local heating and subsequentoxygen loss. The temperature must have increased drasticallybecause of the high current density flowing along the smallvertical grain boundary. On the top side (A1), the amorphous

7

Nanotechnology 24 (2013) 085706 P Calka et al

character as well as the deformation of the oxide layer isprobably related to the accumulation of oxygen. This isconsistent with the protrusion observed both by SSRM andSTEM–HAADF. This oxygen is assumed to come mainlyfrom the HfO2−x region, as oxygen migration towards theanode (AFM tip) is expected. However, we cannot rule outhere possible oxygen incorporation after resistive switchingduring air exposure between the SSRM and FIB preparationsteps and surface amorphization due to the AFM-inducedwear under mechanical and electrostatic forces.

3.6. Discussion

STEM–HAADF measurements show that a localized(∼20 nm) conductive region is created during the electroform-ing step induced by C-AFM in the TiN/HfO2 system. Analysisof the structural, chemical and electronic properties is doneat the nanometre scale by direct STEM–EELS observation.An oxygen-deficient region is identified, corresponding toa protrusion 3 nm in height with clear morphologicalchanges. Analysis of the near-edge fine structures observedon the EELS oxygen K-edge spectra highlights the presenceof oxygen vacancies. HR-TEM observations confirm thatelectroforming is a local phenomenon based on oxygenmigration, starting and propagating through grain boundariesand finally extending over a larger area.

Regarding the HfO2-based systems, the conductionmechanisms mentioned in the literature are mainly based onelectrical measurements [39–42]. Our work brings furtherinsights about the physico-chemical phenomena involved inresistive switching. Experimental proof of the major roleof oxygen is done thanks to direct observation of theconductive region. STEM–EELS results show the existenceof an oxygen-deficient region containing a high density ofoxygen vacancies and associated band gap states. This is adirect confirmation of the filamentary conduction mechanism,based on the breaking of Hf–O bonds and migration ofoxygen under high electric fields (>10 MV cm−1) andelevated temperatures. Doubly charged oxygen vacancies(V2+

o ) and oxygen interstitials ions (O2−i ) appear to be the

most energetically favourable point defects that can be createdin HfO2, as shown by ab initio calculations [43, 44]. Apossible physical interpretation of the switching mechanismis thus based on the creation of such defects, as described bythe following relationship:

HfO2 → HfO2−x + xV2+o + xO2−

i . (2)

Negatively charged interstitial oxygen ions are created and,near the vacancy, the resulting positive charge (2+) is locallytransferred to the metallic d states of the surrounding Hfatoms. Oxygen ions (O2−

i ) are driven towards the anode(positively biased AFM tip) under electrical stress. Theyprobably accumulate near the HfO2 surface yielding the3 nm high protrusion observed by SSRM and HR-TEM (seefigure 8). They might also oxidize near the anode (AFMtip) with charge transfer and oxygen desorption (2O2−

i →

O2 + 4e). However, the CF does not seem to protrude fromthe cathode, connecting both electrodes during electroforming

as a result of oxygen vacancies segregation. On the contrary,our results show that the conductive path is first generated byelectron injection along grain boundaries. Then, at a smalldistance from the cathode, a high concentration of oxygenvacancies is created (see figures 7(b) and 8). Subsequentgrowth of the CF is achieved in one particular grain bysegregation of oxygen vacancies when increasing the currentdensity and local temperature during the electroformingevent. The conductive region is thus characterized by itslocal character (∼20 nm) and high density of defects,such as oxygen vacancies acting as possible traps forelectron conduction through the oxide. Note that the CFproperties mentioned here are strongly correlated to thespecific conditions of the electroforming process in termsof effective voltage, current overshoot, and C-AFM protocol(tip composition and size, N2 atmosphere). In particular, theoxygen-poor environment used for resistive switching mightinduce field-assisted oxygen extraction, thus increasing theoxygen deficiency in the CF.

4. Conclusion

To conclude, reversible resistive switching of the TiN/HfO2

stack is performed using C-AFM. A precise identificationof the conductive region created during electroforming isachieved at the nanometre scale by SSRM. Direct Z-contrastimaging, STEM–EELS analysis and HR-TEM measurementsare made on this individual CF. Results clearly show thatthe conductive region extends over 20 nm and is still mainlycrystalline despite morphological changes and local atomicdisorder. This area appears to be oxygen deficient. EELSoxygen K-edges highlight the presence of oxygen vacanciesassociated with localized states in the HfO2 band gap.HR-TEM shows that generation and propagation of the CFis done along grain boundaries, where the electric field andconductivity is much higher. Partial amorphization is observedat both electrode/oxide interfaces due heating effects oroxygen accumulation. These results are a direct confirmationof the filamentary conduction mechanism, showing thatelectroforming is a local phenomenon related to a strongmodulation of the oxygen content. The process is mainlybased on the injection of electrons from the cathode, creationof oxygen vacancies in the oxide and movement of oxygenions towards the anode. These results contribute to a betterunderstanding of the switching mechanism in HfO2-basedOxRRAM structures for future device optimization.

Acknowledgments

All the measurements were performed at the CEA-LETINanoCharacterisation Centre (NCC) of Minatec. The authorsare grateful to the French ‘Recherche Technologique de Base’program for the measurements performed with the FEI TitanUltimate and to Amal Chabli and Francois Bertin for fruitfuldiscussions. This work was partially funded by the Frenchpublic authorities through the Nano2012 research program.

8

Nanotechnology 24 (2013) 085706 P Calka et al

References

[1] Sawa A 2008 Mater. Today 11 28[2] Waser R, Dittmann R, Staikov G and Szot K 2009 Adv. Mater.

21 2632[3] Baek I G et al 2004 Tech. Dig.—Int. Electron Devices Meet.

p 587[4] Kim S, Biju K P, Jo M, Jung S, Park J, Lee J, Lee W, Shin J,

Park S and Hwang H 2011 IEEE Electron Device Lett.32 671

[5] Lijie Z, Ru H, Minghao Z, Shiqiang Q, Yongbian K, Dejin G,Congyin S and Yangyuan W 2010 IEEE Electron DeviceLett. 31 966

[6] Chien W C et al 2010 IEEE Electron Device Lett. 31 126[7] Chih-Yang L, Chen-Yu W, Chung-Yi W, Tzyh-Cheang L,

Fu-Liang Y, Chenming H and Tseung-Yuen T 2007 IEEEElectron Device Lett. 28 366

[8] Gonon P, Mougenot M, Vallee C, Jorel C, Jousseaume V,Grampeix H and El Kamel F 2010 J. Appl. Phys.107 074507

[9] Walczyk D, Walczyk C, Schroeder T, Bertaud T, Sowinska M,Lukosius M, Fraschke M, Tillack B and Wenger C 2011Microelectron. Eng. 88 1133

[10] Yu-Sheng C, Heng-Yuan L, Pang-Shiu C, Tai-Yuan W,Ching-Chiun W, Pei-Jer T, Chen F, Ming-Jinn T andChenhsin L 2010 IEEE Electron Device Lett. 31 1473

[11] Jousseaume V et al 2011 Solid-State Electron. 58 62[12] Son J Y and Shin Y H 2008 Appl. Phys. Lett. 92 222106[13] Szot K, Speier W, Bihlmayer G and Waser R 2006 Nature

Mater. 5 312[14] Lee D, Seong D J, Jo I, Xiang F, Dong R, Oh S and

Hwang H 2007 Appl. Phys. Lett. 90 122104[15] Choi B J et al 2005 J. Appl. Phys. 98 033715[16] Son J Y, Kim D Y, Kim H, Maeng W J, Shin Y S and

Shin Y H 2011 Electrochem. Solid State Lett. 14 H311[17] Fujii T et al 2011 J. Appl. Phys. 109 053702[18] Uenuma M, Zheng B, Kawano K, Horita M, Ishikawa Y,

Yamashita I and Uraoka Y 2012 Appl. Phys. Lett.100 083105

[19] Kwon D H et al 2010 Nature Nanotechnol. 5 148[20] Wei Z et al 2011 Electron Devices Mtg (IEDM), 2011 IEEE

Int. 31.4.1[21] Calka P, Martinez E, Lafond D, Dansas H, Tirano S,

Jousseaume V, Bertin F and Guedj C 2011 Microelectron.Eng. 88 1142

[22] Arinero R, Hourani W, Touboul A D, Gautier B, Ramonda M,Albertini D, Militaru L, Gonzalez-Velo Y, Guasch C andSaigne F 2011 J. Appl. Phys. 110 014304

[23] Hourani W, Gautier B, Militaru L, Albertini D,Descamps-Mandine A and Arinero R 2011 Microelectron.Eng. 51 2097

[24] Egerton R F 2009 Rep. Prog. Phys. 72 25[25] Waser R and Aono M 2007 Nature Mater. 6 833[26] Lee M-H and Hwang C-S 2011 Nanoscale 3 490[27] Cagli C et al 2011 Electron Devices Mtg (IEDM), 2011 IEEE

Int. 28.7.1[28] Wilk G D and Muller D A 2003 Appl. Phys. Lett. 83 3984[29] Stemmer S, Chen Z Q, Zhu W J and Ma T P 2003 J. Microsc.

210 74[30] Egerton R F 1996 Electron Energy-Loss Spectroscopy in the

Electron Microscope (New York: Plenum)[31] Jang J H, Jung H S, Kim J H, Lee S Y, Hwang C S and

Kim M 2011 J. Appl. Phys. 109 023718[32] McComb D W 1996 Phys. Rev. B 54 7094[33] Baik H S, Kim M, Park G S, Song S A, Varela M,

Franceschetti A, Pantelides S T and Pennycook S J 2004Appl. Phys. Lett. 85 672

[34] Ostanin S, Craven A J, McComb D W, Vlachos D, Alavi A,Finnis M W and Paxton A T 2000 Phys. Rev. B 62 14728

[35] Ostanin S, Craven A J, McComb D W, Vlachos D, Alavi A,Paxton A T and Finnis M W 2002 Phys. Rev. B 65 224109

[36] Vlachos D, Craven A J and McComb D W 2001 J. Phys.:Condens. Matter 13 10799

[37] Bersuker G et al 2011 J. Appl. Phys. 110 124518[38] Lanza M, Zhang K, Porti M, Nafria M, Shen Z Y, Liu L F,

Kang J F, Gilmer D and Bersuker G 2012 Appl. Phys. Lett.100 123508

[39] Yu S, Jeyasingh R, Wu Y and Wong H S P 2012 Phys. Rev. B85 045324

[40] Yu S, Jeyasingh R, Wu Y and Wong H S P 2011 Appl. Phys.Lett. 99 232105

[41] Wong H S P, Lee H Y, Yu S, Chen Y S, Wu Y, Chen P S,Lee B, Chen F T and Tsai M J 2012 Proc. IEEE 100 1951

[42] Goux L, Czarnecki P, Chen Y Y, Pantisano L, Wang X P,Degraeve R, Govoreanu B, Jurczak M, Wouters D J andAltimime L 2010 Appl. Phys. Lett. 97 243509

[43] Zheng J X, Ceder G, Maxisch T, Chim W K andChoi W K 2007 Phys. Rev B 75 104112

[44] Foster A S, Lopez Gejo F, Shluger A L andNieminen R M 2002 Phys. Rev. B 65 174117

9

Atomic structure of conducting nanofilamentsin TiO2 resistive switching memoryDeok-Hwang Kwon1, Kyung Min Kim1,2, Jae Hyuck Jang1, Jong Myeong Jeon1, Min Hwan Lee1,2,

Gun Hwan Kim1,2, Xiang-Shu Li3, Gyeong-Su Park3, Bora Lee4, Seungwu Han1, Miyoung Kim1* and

Cheol Seong Hwang1,2*

Resistance switching in metal oxides could form the basis for next-generation non-volatile memory. It has been arguedthat the current in the high-conductivity state of several technologically relevant oxide materials flows through localizedfilaments, but these filaments have been characterized only indirectly, limiting our understanding of the switchingmechanism. Here, we use high-resolution transmission electron microscopy to probe directly the nanofilaments in aPt/TiO2/Pt system during resistive switching. In situ current–voltage and low-temperature (�130 K) conductivitymeasurements confirm that switching occurs by the formation and disruption of TinO2n21 (or so-called Magneli phase)filaments. Knowledge of the composition, structure and dimensions of these filaments will provide a foundation forunravelling the full mechanism of resistance switching in oxide thin films, and help guide research into the stability andscalability of such films for applications.

Innovations in modern information technology are criticallydependent on the development of denser, faster and less energy-consuming non-volatile memory (NVM)1. Charge-based mem-

ories, such as dynamic random access memory (DRAM) and flashmemory, will suffer from performance degradation as the scalinglimit is approached2. The development of non-charge-basedmemory is therefore essential for extending Moore’s law over thefew next decades. Among the many contenders for next-generationNVM based on a non-charge mechanism, resistance-switchingrandom access memory (RRAM) has attracted increasing attentionbecause of the advantages in its fabrication process as well as its out-standing device performance3–6. In addition, RRAM is also suitablefor the three-dimensional stacking of memory layers, which can leadto the ultimate high-density memory7.

RRAM is based on the reversible dielectric (soft) breakdown ofan insulator, particularly metal oxides. From a microscopic pointof view, resistance switching in various materials can be classifiedbroadly into two different mechanisms3,8. In the valence-changemechanism, the creation and electromigration of oxygen vacanciesinduces the distribution of the carrier density and the valencestates of cations. For example, it is convincingly demonstrated inref. 9 that the oxygen vacancies in SrTiO3 migrate through the dis-location network and affect the conductivity. The device driven bythis mechanism usually shows bipolar behaviour, in which the con-ducting and insulating states are switched with opposite biaspolarity. In the fuse–antifuse mechanism, in which the interplaybetween the thermal effect and the redox reaction in the filamentand its vicinity has a crucial role, metallic filaments are createdthrough the insulator matrix during the electroforming process,and are fused as a result of Joule heating and the electric field10–14.In the antifuse process, the Joule heating-assisted reduction recon-nects the filament. In this case, resistance switching can be achievedwith only one bias polarity, and is thus termed unipolar switchingbehaviour. Among the various oxide materials demonstrating

unipolar switching behaviour5,10–12,15–19, TiO2 appears to be oneof the most promising switching materials to use this switchingmechanism5,10–12,15,16. Several materials also show both types ofswitching behaviours20–22.

In both valence-change and fuse–antifuse mechanisms, manyaspects of the switching behaviour can be understood by assumingthat the current flows through localized filaments in the conductingstate11,13,16. However, basic information about the conducting fila-ments, such as their composition, size and density, has been inferredonly indirectly23,24. Consequently, it is very difficult to understand theresistance-switching phenomena in terms of detailed chemical pro-cesses. The low density of metallic filaments poses a significanthurdle to characterizing their physical properties10. In this study,extensive and careful high-resolution transmission electronmicroscopy (HRTEM) and electron diffraction analyses have shownthat the conducting filaments in TiO2 are composed of TinO2n21(mostly n¼ 4 or 5), known as Magneli phases. In situ localcurrent–voltage (I–V) measurements in TEM indicate that this trans-formed structure can indeed constitute an electrical conduction path.The conductivity measurements at low temperature (�130 K) and insitu switching experiments confirm that the overall resistance switch-ing was induced by the Magneli phase filaments.

Electrical switching behaviourThree types of TiO2 films were examined: pristine, SET and RESETsamples. The pristine sample was prepared using plasma-enhancedatomic layer deposition of 40-nm-thick TiO2 thin films followed bythe deposition of a platinum electrode, as shown in Fig. 1a (seeMethods). To switch the pristine metal–insulator–metal (MIM)sample to the conducting state (forming process), a negative biasis applied to the top electrode with an appropriate compliancecurrent (Fig. 1b). The TiO2 film in this low-resistance state iscalled the ‘SET’ sample. The scanning electron microscopy (SEM)image in Fig. 1c shows that a part of the platinum top electrode is

1Department of Materials Science and Engineering, Seoul National University, Seoul 151-744, Korea, 2Inter-university Semiconductor Research Center, SeoulNational University, Seoul 151-744, Korea, 3Analytical Research Laboratory, Samsung Advanced Institute of Technology, PO Box 111, Suwon 440-600, Korea,4Department of Physics, Ewha Womans University, Seoul 120-750, Korea. *e-mail: mkim@snu.ac.kr; cheolsh@snu.ac.kr

ARTICLESPUBLISHED ONLINE: 17 JANUARY 2010 | DOI: 10.1038/NNANO.2009.456

NATURE NANOTECHNOLOGY | VOL 5 | FEBRUARY 2010 | www.nature.com/naturenanotechnology148

© 2010 Macmillan Publishers Limited. All rights reserved.

blown off after the forming process, as indicated by the black arrow.The small explosion is probably caused by the sudden evolution ofcompressed excess oxygen gas16. The blown-off region of thesample may correspond to a location where the strongest filamentshave developed, and was useful for identifying the filaments at theinitial stage of the present work (Supplementary Fig. S1). However,this part does not act as a current path in the following switchingprocess, because the top electrode is missing. Two of the capacitorstructures were electroformed in this manner. A negative bias wasfurther applied to one of the two electroformed structures to switchthe sample into a high-resistance state. The TiO2 film in this high-resistance state is called the ‘RESET’ sample. Details regarding resist-ance switching by means of the I–V sweeps are reported in ref. 11.

HRTEM observationThe electron diffraction (ED) patterns and fast Fourier transformed(FFT) micrographs of the HRTEM images from each sample wereexamined extensively to determine the crystallographic structureof the filaments and the remaining part of the TiO2 films. In thecase of the pristine TiO2 sample, the majority phase was identifiedas a meta-stable brookite structure (Supplementary Fig. S2), ratherthan a rutile or anatase structure. This is understandable, becausethe deposition conditions of the pristine sample were far from thethermodynamic stability condition, even with plasma applicationat a growth temperature of 250 8C. It is also noted that nanoscalenon-stoichiometric TiOx (x , 2) phases were pervasive throughoutthe thin films, determined by the FFT of local regions of a few nano-metres, although their volume ratio is relatively small.

The ED patterns from the SET and RESET samples, however, havefeatures clearly distinguishable from pristine TiO2. First, a substantialamount of the dielectric films was changed, in both cases, into therutile or anatase phases. Thermal heating during the I–V sweepsmight have triggered the transition to more stable phases. Second,ED spots with smaller scattering angles than those of the anatase

[101] spots appear. Because the anatase [101] spots, correspondingto a d-spacing of 0.351 nm, have the smallest observable diffractionangle in the ED patterns from stoichiometric TiO2 with the brookite,anatase and rutile structures, those extra diffraction spots are indica-tive of the presence of non-stoichiometric TinO2n21 (mostly n¼ 4 or5). (Possible diffraction spots that can be excited by multiple scatter-ing, both from the same grain and from separate grains, were carefullytraced and were not considered for further examination.) This isknown as a Magneli phase25. The Magneli phase is a defective struc-ture derived from the ideal rutile phase, and can be described by thetwo-dimensional rutile (121) slabs made from octahedral TiO6 withan n-layer thickness in the direction normal to the slab plane. Theadjacent two slabs were displaced by 1=2½0�11� to accommodate theoxygen deficiencies25. Therefore, the nth (121) plane constitutes anantiphase boundary, which is known as the crystallographic shearplane. It is also well known that most of these Magneli phases are met-allic conductors near room temperature26. Therefore, the presence ofa Magneli phase in the SET and RESET samples could be responsiblefor the observed resistance-switching behaviours.

Most Magneli structures, whether connected or disconnected inthe HRTEM images, were fairly straight in both SET and RESETsamples. This implies that most nanofilaments are normal to theelectrode; if the filaments deviate significantly from the vertical direc-tion, one should be able to observe an image of slant nanofilaments inHRTEM. This is reasonable, because the nanofilaments are likely tobe formed along the direction of applied electric fields, which is thenormal direction to the film surface in the planar MIM geometry.

In the following, a more detailed analysis of ED patterns for SETsamples is presented. Figure 2a, for the SET state, shows a clear andconical pillar with diameters of �15 and 3 nm at the cathode (TE)and anode (BE) interfaces, respectively, comprising a Magneli phase.The microscopic structure was confirmed from an ED pattern inFig. 2b, which shows the diffraction spot with a d-spacing of 0.62 nm(marked with a circle). This was identified as the (002) spot of aMagneli phase with n¼ 4. The dark-field image from this spot is dis-played in Fig. 2c. The high-resolution image of the bright area in thedark-field image is shown in Fig. 2a. The FFT image in Fig. 2d also con-firms that the structure is Ti4O7. The diffraction image in Fig. 2e is thesimulated [140] diffraction pattern of Ti4O7, which coincides well withthe FFT image. Another image for the connected filament in the SETsample can be found in Supplementary Fig. S3.

An extensive examination of the SET samples revealed only sixconnected filament images in five 10-mm-wide focused ion beam(FIB) samples. (This includes two strong filaments found withinthe blown-off region.) Considering the low probability of the thinTEM specimen containing nanofilaments and the informationlimit of TEM, this does not necessarily mean that there are onlysix conducting filaments in such a large area of the sample. Inaddition, the observation of the nanofilaments in TEM requiresthat the filaments must be in specific crystallographic orientationsfor the given electron-beam directions.

Besides the nanofilaments in the connected shape, severaldisconnected nanofilaments in Magneli phases were also found.(In fact, most of the nanofilaments were disconnected.) This canbe understood based on the nanofilament growth process. Duringthe electroforming step, many nanofilaments may start to grow. Asthe nanofilaments connect the top and bottom electrodes, large cur-rents would flow through these metallic paths. The bias voltage islargely reduced when the current level reaches the compliancelimit, preventing further growth of other nanofilaments. Figure 2fshows a typical partial filament. (The images in Fig. 2f–j correspondto the images in Fig. 2a–e, respectively.) Most of the incompleteTi4O7 or Ti5O9 pillars were present near the top electrode. Thismeans that the nanofilaments usually grow from the cathode. Inaddition, these pillars typically have conical shapes, with a widerdiameter at the cathode side. (In the connected nanofilaments,

c

SET

Blown offregion

10 µm

TiO2

Pt

SiO2 Si

Pta b

0 1 20.00

0.01

0.02

0.03

Cur

rent

(A)

Voltage (V)

ICC

SETstate

RESETstate

Figure 1 | Schematic of the device structures and SEM image after the

forming process. a, Schematic of the Pt/TiO2/Pt stack. b, Typical I–V curves

of the MIM sample showing three different conduction states. c, SEM image

of the top electrode in the SET state. The image shows the blown-off region

marked as the black box in the low-magnification SEM image in the inset.

NATURE NANOTECHNOLOGY DOI: 10.1038/NNANO.2009.456 ARTICLES

NATURE NANOTECHNOLOGY | VOL 5 | FEBRUARY 2010 | www.nature.com/naturenanotechnology 149

© 2010 Macmillan Publishers Limited. All rights reserved.

three of the observed filaments have conical shapes; see Fig. 2a, Fig. 5cand Supplementary Fig. S3.) These findings are consistent with a fila-ment-growth model for TiO2 as already proposed11,15.

TEM analysis of the RESET samples indicates that they aresimilar to SET samples, except that no connected nanofilamentcould be found. In the RESET operation, the connected nanofila-ment should be ruptured, possibly by thermally assisted electromi-gration of oxygen ions10,11,15. However, disconnected nanofilamentswould not be affected by the RESET operation. (One of the discon-nected nanofilaments found in the RESET sample is shown inSupplementary Fig. S4.)

The observed filament diameters (measured at the middle pointalong the length) in the SET and RESET states ranged from 5 to10 nm, with the distance between them being between 0.1 and 5 mm(considering both connected and incomplete filaments). These areessential parameters for estimating the ultimate packing density ofthe RRAM device. The relatively large distances between filamentsare an unfavourable feature of the device, because this can comprise ascaling limit. However, it should be noted that once a filament is estab-lished, there would be no further filament formation in the adjacentregion, because the current flows mostly through the connected fila-ment. Therefore, even if the device size becomes very small, namely

� 100� 100 nm2, it is anticipated that there would be at least onefilament with which resistance switching could occur. In fact, a lowerdensity of filaments could be advantageous for device applications,because once a filament is formed in a nanoscale memory cell, it is unli-kely that another will form in the same cell. (Multiple filaments maydeteriorate the reproducibility of the resistance switching behaviour.)In this sense, a smaller cell size is conducive to better uniformity andrepeatability. Therefore, the size of the filament is a more importantparameter. The small size of the filaments suggests that the memorycell size can in principle be scaled down to tens of nanometres.

Even though the Magneli phase is a metal in the bulk phase, it isunclear whether the nano-sized Magneli pillar is also conducting.To establish this, the electrical conductivity of the Magneli nanofila-ments in the SET sample was measured using a local in situ I–Vscan in TEM (Fig. 3a) using the scanning tunnelling microscopy(STM) tip operating in conductive atomic force microscopy(CAFM) mode. The sample was prepared from the region where thetop electrode was blown off. Figure 3b shows the I–V curves obtainedfrom the Magneli structure (red circles) and from the area �50 nmaway from the filament where the Magneli structure was not identified(blue squares). The local I–V curve showed an electrical conductivityratio of �1,000 between the two locations. It was noted that the

TE

BE

(002)

a f

b c

d e

g h

i j

TE

BE

(002)

10 nm 10 nm

20 nm 20 nm

0.62 nm

2 nm−1 2 nm−1

0.62 nm

002

Ti4O7Ti4O7 [140]

413

002

Figure 2 | Magneli structures in the SET sample. a–e, High-resolution TEM image of a Ti4O7 nanofilament (a), selected-area diffraction pattern of the film

(b), dark-field image obtained from the diffraction spot marked as a circle in the diffraction pattern (c), a fast Fourier transformed micrograph of the high-

resolution image of Ti4O7 (d), and the corresponding simulated diffraction pattern (e). The Bloch-wave method was used to simulate the diffraction patterns.

f–j, Disconnected Ti4O7 structure in the conical shape. The images are presented in the same manner as for the connected filament in a–e.

ARTICLES NATURE NANOTECHNOLOGY DOI: 10.1038/NNANO.2009.456

NATURE NANOTECHNOLOGY | VOL 5 | FEBRUARY 2010 | www.nature.com/naturenanotechnology150

© 2010 Macmillan Publishers Limited. All rights reserved.

current variation due to the difference in probe pressure on the samplewas �15%, and there was a �10% difference with and without inci-dent high-energy electrons (for the TEM observation). The contactresistance might also complicate measurement of the absolute resist-ance values. However, the large conductivity ratio suggests that theobserved Magneli phases are indeed localized conducting paths. Inpassing, it is worth mentioning that there are three orders of differencein the current level between Figs 1b and 3b. This is because a largenumber of filaments contribute to the total current in the pad-typedevice in Fig. 1. In contrast, Fig. 3b was obtained for TEM samplesin which only one filament was probed.

Low-temperature measurementsAlthough TEM analysis in the above provided a clear observation of theMagneli phase filaments, it needs to be further confirmed that the nano-filaments are responsible for the conductivity in the SET state. This isbecause other parts of the TiO2 film, which may contain a largeportion of oxygen-deficient TiOx phase (or even metal atoms), mayalso contribute to electrical conduction27,28. To this end, the tempera-ture-dependent I–V curves of the MIM sample in the SET state weremeasured. It is known that the Magneli-structured Ti4O7 phase exhibitsa metal (high-temperature) to semiconductor (low-temperature) tran-sition at �150 K, with an abrupt drop in electrical conductivity by afactor of �1,000, caused by charge ordering of Ti3þ and Ti4þ ions29.Figure 4a shows the current values of the MIM sample measured at0.1 V with the temperature varying between 129.5 and 200 K. It isfound that the current drops sharply at 130 K. In Fig. 4b, the

corresponding I–V curves indicate that the conduction behaviourchanges from metallic to semiconducting (or insulating) near 130 K.(The I–V curve at 305 K was almost identical to that at 200 K.) Thisis similar to results reported for single-crystal Ti4O7 in ref. 29, exceptfor the transition temperature (130 K versus 150 K), and confirmsthat the electrical conduction in the SET state is governed by theMagneli nanofilaments. The discrepancy in the transition temperaturecould be ascribed to different correlation effects between the nanostruc-ture of the Magneli phase in the matrix of TiO2 and the single crystal.The I–V measurements at temperatures ranging from 313 to 363 K(data not shown) showed that the on-state resistance increases withthe temperature, which confirms metallic conduction in the SET state.

In situ switching in TEMTo further confirm that the observed structure and the accompanyingelectrical properties of Magneli phase nanofilaments are manipulatedby the applied voltage, an in situ RESET operation on a connectedMagneli structure was performed in TEM using the experimentalset-up shown in Fig. 5a. The STM tip approached the TE with theBE grounded. The Magneli structure of the connected nanofilamentwas confirmed (Fig. 5c) and then the I–V curve of the SET state wasobtained (red circles in Fig. 5b). The current density calculated fromFig. 5b reaches �1� 106 A cm22. Although this current density isnot high enough to induce elecromigration of atoms in the filament,the accompanying Joule heating effect may thermally anneal the fila-ment. An I–V sweep to a lower voltage (21.0 V) was then performedto reset the TEM sample. The I–V curve after the RESET operation

1 10 10010−1

100

101

102

103

baTi4O7

TiO2

Cur

rent

(nA

)

TiO2 Ti 4O 7

Pt

Voltage (mV)

Figure 3 | In situ I–V scan on nanofilaments. a, Schematic to show the experimental set-up. b, Local I–V curves measured on the Ti4O7 structure (red circles)

and on the anatase structure that is �50 nm away from the Ti4O7 (blue squares). The conductivity ratio between the two locations is �1,000.

0.1

1

10

a b

Cur

rent

(mA

at 0

.1 V

)

Temperature (K)

−10−0.4200 180 160 140 120 −0.2 0 0.2 0.4

−5

0

5

10 C

urre

nt (m

A)

Voltage (V)

Temp. (K) 200 180 160 150 145 132 130 129.5

Figure 4 | Temperature-dependent conduction behaviours of the MIM sample in the SET state. a, Current values measured at 0.1 V with the temperature

varying between 129.5 and 200 K. b, Corresponding I–V curves. The I–V curve at 305 K was almost identical to that of 200 K (not shown). It is found that

the sample changes its conduction behaviour from metallic (down to 132 K) to semiconducting (or insulating) at 130 and 129.5 K.

NATURE NANOTECHNOLOGY DOI: 10.1038/NNANO.2009.456 ARTICLES

NATURE NANOTECHNOLOGY | VOL 5 | FEBRUARY 2010 | www.nature.com/naturenanotechnology 151

© 2010 Macmillan Publishers Limited. All rights reserved.

represented in Fig. 5b (blue squares) indicates that the conductivity waslowered by a factor of 10–20. Concurrently, the structure of the formerMagneli nanofilament was converted into other structures (probablyanatase), as shown in Fig. 5d. The diffraction spots from theMagneli structure also disappeared as expected. This confirms thatthe RESET operation corresponds to the phase transformation ofnanofilaments from Magneli to other stoichiometric phases. Becausethe experiment was performed in a high vacuum condition, it islikely that the oxygen atoms were supplied from the neighbouringoxide phase. In Fig. 5b, the conductivity change in the in situRESET was less pronounced when compared to the pad-type switch-ing in Fig. 1c. From a comparison with Fig. 3b, it is found that thecurrent in the SET state is lower than the corresponding value inFig. 3b, meaning that this particular filament is less conducting. Thetwo orders of difference in the SET-state currents in Figs 3b and 5bcould be attributed to the disparate filament shapes. In Fig. 3b, the fila-ment is straight with a diameter of �10 nm, and the measurement tipis positioned in the middle of the filament. The filament in Fig. 5b iscone-shaped with diameters of 10 (top) and 5 nm (bottom) and the tipis positioned on the electrode. Consideration of these geometric differ-ences can account for one order of difference. In addition to this, thesharp and narrow shape of the filament near the bottom electrode maycomplicate the transport behaviour at the interface. The non-ohmic be-haviour in Fig. 5b might also be caused by this.

The in situ SET experiment was also tried and the conductivity sig-nificantly increased to a value much higher than that of the initial SETstate with the reappearance of the diffraction spots of the Magneli

phase. However, the connected Magneli phase was not identified inthe TEM observation area. This suggests that the newly formed nano-filament was located in a region outside the area observed by HRTEM.(The observable area by HRTEM was only �50 nm wide, and thediameter of the area contributing to the diffraction patterns wasover �1 mm across.) Even though the filament in the present in situexperiment was completely eliminated and was not recovered, thisdoes not exclude the possibility of partial rupture of the filament. Infact, from a CAFM study on the TiO2 film (data not shown), it wasoften observed that some filaments were recreated at the same spot,suggesting that those filaments were partially broken.

ConclusionsThe identification of the Magneli phase as conducting filaments hasimportant implications for the switching mechanism in RRAM. Atthe initial stage, a random oxygen vacancy will be created as theoxygen atoms are displaced from the bulk position by the externalfield and thermal effects. Above a critical density, oxygen vacancieswill spontaneously rearrange to form an ordered structure. Indeed,first-principles calculations with the same computational set-up asin ref. 28 found that the Magneli phase is more stable than therutile phase with a uniform distribution of oxygen vacancy, by�2 eV per Ti4O7 formula unit. Therefore, with increasing concen-tration of oxygen vacancies, there should be a strong thermodynamicdriving force to form the Magneli phase. The presence of a stablemetallic phase explains the outstanding endurance of many RRAMdevices, which is difficult to rationalize if the conducting paths

−1 −10

a b

c d

−100 −1,000−10−1

−100

−101

−102

−103

In situ SETRESET

Cur

rent

(nA

)

Voltage (mV)

TiO2

Ti 4O 7

Pt

Pt

111

0.43 nm

102

Ti4O7 [231]10 nm 10 nm

2 nm−1 2 nm−1

Figure 5 | Structural transformation after an in situ RESET experiment. a, Schematic to depict the experimental set-up. b, Local I–V curves in a log scale

before and after RESET. The STM probe approached the top electrode, and the I–V curves represent the electrical conduction between the top and bottom

electrodes. c, High-resolution image, diffraction pattern and fast Fourier transformed micrograph of the Magneli structure before RESET. d, The corresponding

images after RESET. The diffraction spot (marked as a circle in c) from the Magneli structure disappeared after RESET.

ARTICLES NATURE NANOTECHNOLOGY DOI: 10.1038/NNANO.2009.456

NATURE NANOTECHNOLOGY | VOL 5 | FEBRUARY 2010 | www.nature.com/naturenanotechnology152

© 2010 Macmillan Publishers Limited. All rights reserved.

comprise a random distribution of point defects. In this respect, itwould be worth reexamining the conducting path in other oxidesin terms of vacancy ordering or oxygen-deficient phases. A recentreport on the valence reduction of a switching filament in CuOcells is, thus, very interesting30.

MethodsA 40-nm-thick TiO2 thin film was deposited on a 100-nm-thick sputteredplatinum/SiO2/silicon substrate by plasma-enhanced atomic layer deposition at250 8C using titanium tetraisopropoxide and plasma-activated O2 as the precursor andoxygen source, respectively. Structural and chemical characterization using grazing-angle incidence X-ray diffraction and X-ray photoelectron spectroscopy revealed theas-grown TiO2 film to have an amorphous to polycrystalline mixed anatase andbrookite structure with an oxygen/titanium ratio of �2.1. The grain shape wasgranular, and no specific preferred crystallographic orientation was observed.A 50-nm-thick circular platinum top electrode with a diameter of 300 mm was thenfabricated by electron-beam evaporation followed by a lift-off photolithographicprocess. The resistive switching behaviour of the MIM structure was measured at roomtemperature using an HP4145B semiconductor parameter analyser in the I–V sweepmode. The switching behaviour of the sample was measured by applying a negativebias voltage to the top electrode with the bottom electrode grounded. Two sampleswere electroformed using the I–V sweep with a compliance current of 20 mA. One ofthe two samples was set again to the high-resistance state by applying another reset I–V sweep. Therefore, SET samples experienced only one I–V sweep and the RESETsamples experienced two I–V sweeps. To confirm that the observed Magneli phasefilaments governs the overall resistive switching behaviour of the MIM sample, theconductance of the MIM sample in the SET state was measured at low temperature(down to 129.5 K) using a low-temperature stage of the CAFM (JEOL JSPM 5400) in ahigh-vacuum (1� 10–6 Torr) condition. A platinum-coated STM-type tip was used tominimize the contact resistance. The presence of intervening SiO2 below the MIMstructure adversely interferes with the fluent heat transfer between the MIM structureand the stage. To mitigate this problem, the TiO2 film surface was thermally connectedto the cold stage surface using silver paste.

The TEM specimens were prepared using a FIB etching technique to include thedielectrics directly under the top platinum electrode. A 300 kV field emission TEM(Jeol 3000F) and a 200 kV field emission TEM (Tecnai F20) were used for electrondiffraction, dark-field imaging and HRTEM. In situ localized I–V measurementsin Fig. 3 were performed on the Magneli structures, which were confirmed by ED andHRTEM before in situ experiments. A two-terminal I–V curve was measuredusing a double-tilt STM (Nanofactory Instruments AB, ST-1000) installed on a TEMholder of Tecnai F20G serving as a manipulator. I–V was measured in CAFMmode. The contact between the tungsten probe and the TiO2 sample was made usingthe STM unit, which could control the probe movement at sub-nanometreresolution with a piezotube. The integrity of the contact was constantly monitoredduring the I–V measurement. The TEM specimen for the in situ RESET experimentwas prepared by FIB from the SET sample, retaining both TE and BE, and negativebias was applied to the TE to keep the electrical bias direction the same as that inpad-type switching. The STM tip approached the TE instead of the TiO2 film to avoidinevitable specimen damage due to excessive heating related to the high contactresistance between the tip and TiO2. The connecting area between the top electrodeand the TEM copper grid was removed carefully by FIB in order to cut off the currentpath between the top and bottom electrodes through the copper grid.

Received 6 October 2009; accepted 25 November 2009;published online 17 January 2010

References1. Burr, G. W. et al. Overview of candidate device technologies for storage-class

memory. IBM J. Res. Dev. 52, 449–464 (2008).2. Waser, R. (ed.) Nanoelectronics and Information Technology Ch. 4

(Wiley-VCH, 2003).3. Waser, R. & Aono, M. Nanoionics-based resistive switching memories. Nature

Mater. 6, 833–840 (2007).4. Strukov, D. B., Snider, G. S., Stewart, D. R. & Williams, R. S. The missing

memristor found. Nature 453, 80–83 (2008).5. Seo, S. et al. Reproducible resistance switching in polycrystalline NiO films.

Appl. Phys. Lett. 85, 5655–5657 (2004).6. Meijer, G. I. Who wins the nonvolatile memory race? Science 319,

1625–1626 (2008).7. Likharev, K. K. & Strukov, D. B. Prospects for the development of digital circuits.

Proc. IEEE Int. Symp. Nanoscale Architectures 109–116, 2007.8. Sawa, A. Resistive switching in transition metal oxides. Mater. Today 11,

28–36 (2008).9. Szot, K., Speier, W., Bihlmayer, G. & Waser, R. Switching the electrical resistance

of individual dislocation in single-crystalline SrTiO3. Nature Mater. 5,312–320 (2006).

10. Choi, B. J. et al. Resistive switching mechanism of TiO2 thin films grown byatomic-layer-deposition. J. Appl. Phys. 98, 033715 (2005).

11. Kim, K. M., Choi, B. J., Shin, Y. C., Choi, S. & Hwang, C. S. Anode-interfacelocalized filamentary mechanism in resistive switching of TiO2 thin films.Appl. Phys. Lett. 91, 012907 (2007).

12. Rohde, C. et al. Identification of a determining parameter for resistive switchingof TiO2 thin films. Appl. Phys. Lett. 86, 262907 (2005).

13. Sato, Y., Kinoshita, K., Aoki, M. & Sugiyama, Y. Consideration of switchingmechanism of binary metal oxide resistive junctions using a thermal reactionmodel. Appl. Phys. Lett. 90, 033503 (2007).

14. Russo, U., Ielmini, D., Cagli, C. & Lacaita, A. L. Self-accelerated thermaldissolution model for reset programming in unipolar resistive-switchingmemory (RRAM) devices. IEEE Trans. Electron. Dev. 56, 193–200 (2009).

15. Kim, K. M. & Hwang, C. S. The conical shape filament growth model in unipolarresistance switching of TiO2 thin film. Appl. Phys. Lett. 94, 122109 (2009).

16. Jeong, D. S., Schroeder, H., Breuer, U. & Waser, R. Characteristic electroformingbehaviour in Pt/TiO2/Pt resistive switching cells depending on atmosphere.J. Appl. Phys. 104, 123716 (2008).

17. Shim, H. et al. Resistance-switching characteristics of polycrystalline Nb2O5 fornonvolatile memory application. IEEE Electron. Device Lett. 26, 292–294 (2005).

18. Lee, D. et al. Resistance switching of the nonstoichiometric zirconium oxide fornonvolatile memory applications. IEEE Electron. Device Lett. 26,719–721 (2005).

19. Kim, K. M. et al. Resistive switching in Pt/Al2O3/TiO2/Ru stacked structures.Electrochem. Solid State Lett. 9, G343–G346 (2006).

20. Jeong, D. S., Schroeder, H. & Waser, R. Coexistence of bipolar and unipolarresistive switching behaviors in a Pt/TiO2/Pt stack. Electrochem. Solid State Lett.10, G51–G53 (2007).

21. Lee, S., Kim, W.-G., Rhee, S.-W. & Yong, K. Resistance switching behaviors ofhafnium oxide films grown by MOCVD for nonvolatile memory applications.J. Electrochem. Soc. 155, H92–H96 (2008).

22. Yang, Y. C., Pan, F., Liu, Q., Liu, M. & Zeng, F. Fully room-temperature-fabricated nonvolatile resistive memory for ultrafast and high-density memoryapplication. Nano Lett. 9, 1636–1643 (2009).

23. Fujiwara, K. et al. Resistance switching and formation of a conductive bridge inmetal/binary oxide/metal structure for memory devices. Jpn J. Appl. Phys. 47,6266–6271 (2008).

24. Lee, M.-J. et al. Electrical manipulation of nanofilaments in transition-metaloxides for resistance-based memory. Nano Lett. 9, 1476–1481 (2009).

25. Bursill, L. A. & Hyde, B. G. Crystallographic shear in the higher titanium oxides:structure, texture, mechanisms and thermodynamics. Prog. Solid State Chem. 7,177–253 (1972).

26. Inglis, A. D., Page, Y. L., Strobel, P. & Hurd, C. M. Electrical conductance ofcrystalline TinO2n21 for n¼ 4–9. J. Phys. C 16, 317–333 (1983).

27. Yang, J. J. et al. Memristive switching mechanism for metal/oxide/metal nanodevices. Nature Nanotech. 3, 429–433 (2008).

28. Cho, E. et al. First-principles study of point defects in rutile TiO22x. Phys. Rev. B73, 193202 (2006).

29. Lakkis, S., Schlenker, C., Chakraverty, B. K., Buder, R. & Marezio, M. Metal–insulator transitions in Ti4O7 single crystals: crystal characterization, specificheat and electron paramagnetic resonance. Phys. Rev. B 14, 1429–1440 (1976).

30. Yasuhara, R. et al. Inhomogeneous chemical states in resistance-switchingdevices with a planar-type Pt/CuO/Pt structure. Appl. Phys. Lett. 95,012110 (2009).

AcknowledgementsThis work was supported by National Research Foundation of Korea grant funded by theMinistry of Education, Science and Technology (2009-0083038) and MEST-AFOSR NBITProgram. C.S.H., K.M.K., M.H.L. and K.H.K. acknowledge support by the NationalProgram for 0.1 Terabit NVM Devices of the Korean Government, the National ResearchFoundation of Korea (NRF) funded by the Ministry of Education, Science and Technology(grant no. 2009-0081961), and World Class University program through the Korea Scienceand Engineering Foundation funded by the Ministry of Education, Science and Technology(grant no. R31-2008-000-10075-0). B.L. and S.H. were supported by the QuantumMetamaterials Research Center (grant no. R11-2008-053-03001-0).

Author contributionsD.-H.K., J.H.J. and J.M.J. performed the TEM experiments and analysed the diffraction data.X.-S.L., G.-S.P. and D.-H.K. performed the in situ switching experiments in STM–TEM.K.M.K. and G.H.K. fabricated the samples and performed electrical switching experiments.M.H.L. performed the low temperature experiment. B.L. and S.H. performed the first-principles calculation. M.K. and C.S.H. conceived and designed the experiments. M.K., S.H.and C.S.H. co-wrote the paper. All authors discussed the results and commented onthe manuscript.

Additional informationThe authors declare no competing financial interests. Supplementary informationaccompanies this paper at www.nature.com/naturenanotechnology. Reprints andpermission information is available online at http://npg.nature.com/reprintsandpermissions/.Correspondence and requests for materials should be addressed to M.K. and C.S.H.

NATURE NANOTECHNOLOGY DOI: 10.1038/NNANO.2009.456 ARTICLES

NATURE NANOTECHNOLOGY | VOL 5 | FEBRUARY 2010 | www.nature.com/naturenanotechnology 153

© 2010 Macmillan Publishers Limited. All rights reserved.