Graphene resistive random memory — the promising memory device in next generation
Wang Xue-Feng1, 2, Zhao Hai-Ming1, 2, Yang Yi1, 2, †, Ren Tian-Ling1, 2, ‡
Institute of Microelectronics, Tsinghua University, Beijing 100084, China
Tsinghua National Laboratory for Information Science and Technology (TNList), Tsinghua University, Beijing 100084, China

 

† Corresponding author. E-mail: yiyang@tsinghua.edu.cn RenTL@tsinghua.edu.cn

Abstract

Graphene-based resistive random access memory (GRRAM) has grasped researchers’ attention due to its merits compared with ordinary RRAM. In this paper, we briefly review different types of GRRAMs. These GRRAMs can be divided into two categories: graphene RRAM and graphene oxide (GO)/reduced graphene oxide (rGO) RRAM. Using graphene as the electrode, GRRAM can own many good characteristics, such as low power consumption, higher density, transparency, SET voltage modulation, high uniformity, and so on. Graphene flakes sandwiched between two dielectric layers can lower the SET voltage and achieve multilevel switching. Moreover, the GRRAM with rGO and GO as the dielectric or electrode can be simply fabricated. Flexible and high performance RRAM and GO film can be modified by adding other materials layer or making a composite with polymer, nanoparticle, and 2D materials to further improve the performance. Above all, GRRAM shows huge potential to become the next generation memory.

1. Introduction

Resistive random access memory (RRAM) is a kind of memory based on a resistive switching mechanism controlled by external voltage.[16] Its simple metal-insulating layermetal structure can break through technological and physical limitations existing in ordinary memory devices such as dynamic random access memory (DRAM) and flash as the device scales down and has higher storage capacity.[7] Apart from it, facile processing, fast switching, good endurance, compatibility with conventional semiconductor processing, and high stability[811] make RRAM a promising next generation nonvolatile memory (NVM). RRAM has two stable states, a high resistance state (HRS) and a low resistance state (LRS). The resistive mechanism, such as the conductive filament growth and rupture, has been proved both theoretically[1214] and experimentally.[15] There are many kinds of materials used as the resistive switching (active) layer, such as transition metal oxides (ZnO,[16] ZrO2,[17] TiO2,[18] NiO,[19] HfOx,[20] ), organic materials,[22] chalcogenides,[23] and so on. However, ordinary RRAM has some problems such as the limitation of electrode thickness, which affects RRAM further scaling down, reading current decreasing in LRS, and flexible fabrication. Recently, 2D material graphene[24] with carbon atoms packed into a plane honeycomb crystal structure becomes a research hotspot due to its extremely high mobility,[25] ease of being scaled down,[26] high light transmittance,[27] strong intrinsic breaking strength,[28] and good electrical and mechanical properties. Many graphene-based devices[29,30] are fabricated, such as transistors,[31] solar cells,[32] sound emitting devices,[33] gas detectors,[34] and photodetectors.[35] Many RRAM problems such as scalability and power consumption can be solved by inserting a graphene layer. Moreover, graphene RRAM owns new exciting properties, for example, high transparency and SET voltage modulation, which will be interpreted in the following. Graphene can be large-area grown both on Cu foil[36] and Ni substrate,[37] which is good news for the mass fabrication of graphene RRAM in the future. Other graphene-based materials are also widely used in RRAM fabrication, such as graphene oxide (GO) and reduced graphene oxide (rGO). GO is a layered stack of the planar, graphene-like sheet decorated by epoxy, diol, hydroxyl, ether, and carboxy groups;[38] and rGO is the reduction of GO by laser scribed,[39,40] chemicals,[41,42] or other methods. Due to easy fabrication, flexible and good electrical, thermal, mechanical, and optical properties, GO and rGO are applied to fabricate all kinds of devices, such as pressure sensors,[43] thermal rectifiers,[44] flexible earphones,[45,46] strain sensors,[47] and light emitters.[48] For RRAM applications, rGO can be used as the electrode and rGO or GO as the active layer, which can achieve RRAM simple fabrication, flexibility, relative good endurance, and high stability. Combining with other materials such as π-conjugated polymer will improve the RRAM performance and obtain novel properties such as multilevel switching.

In this paper, we will briefly review the different types of graphene-based RRAMs and explore the roles of graphene materials in RRAM application. In Section 2, we give a simple introduction of graphene RRAMs according to the merits and functions of graphene as the electrode, including low power consumption, higher density, monitoring and detecting oxygen ion drift function, transparency, good uniformity, SET voltage modulation, and graphene as the middle layer. In Section 3, GO/rGO RRAM, including rGO as the electrode and GO/rGO as the dielectric layer is reviewed, where the model, metal-insulator-metal (MIM) structure, and GO modification of GO RRAM are included. Finally, we give concluding remarks in Section 4.

2. Graphene RRAM
2.1. Graphene as electrode
2.1.1. Low power consumption

With the integration scale becoming larger, low power consumption for single RRAM is essential. Large out-of-plane resistance caused by the weak interaction of graphene with other materials[49] can meet the need. References [50] and [51] used single layer grapheme (SLG) to engineer the interface between the oxide layer and the metal electrode and reduce the RESET current, providing a potential way towards further lowering the programming energy of RRAM.

The structure is shown in Fig. 1(a), composed by Pt/HfO2/SLG/Ti/TiN. The thickness of the insulating layer (HfO2) is ∼ 5 nm and TiN/Ti is ∼ 65 nm/8 nm. The specific contact resistivity (Fig. 1(c)) between SLG and the top electrode is ∼1.4×10−5 Ω·cm−2, showing a good contact between them. Compared with the sample without SLG, this RRAM reveals a high-resistance interface barrier (> 900 kΩ) and shows 22 times smaller average RESET current as shown in Fig. 1(d). The programming power consumption (calculated as the RESET voltage multiplied by the peak of the RESET current) shows ∼ 47 times reduction for SLG RRAM. Over 105 s retention time can be realized (Fig. 1(e)).

Fig. 1. (color online) (a) The structure of Pt/HfOx/graphene/Ti/TiN RRAM. (b) The I–V curves of RRAMs with and without graphene, showing that the GRRAM owns lower SET voltage and RESET current. (c) The relation between the contact resistance between graphene and the top electrode and the graphene length, indicating low contact resistance with the top electrode. (d) The relation between the cumulative probability and the RESET current of RRAMs with and without graphene respectively. The RESET current of RRAM with graphene has a 22 reduction compared to that without graphene. (e) Retention time of GRRAM under HRS and LRS, repectively. The figure is reproduced from Ref. [51].
2.1.2. Monitoring and detecting oxygen ion drift

Theoretical researches suggest that the switching mechanism for oxide-based RRAM is oxygen ion drift caused by external electrical field. However the redox reaction caused by oxygen ion drift occurs beneath the top electrode, which is hard to analyze and monitor with a nondestructive method for ordinary RRAM. The graphene RRAM coupled with Raman spectroscopy or XPS will be an effective way to solve the problem since the different state of oxygen ion drift in RRAM will cause the change of graphene Raman and XPS spectra.

For graphene Raman spectroscopy, the G peak shift and 2D peak intensity change can reflect the doping effect in the graphene.[52] Therefore observing these two parameters changing during the graphene RRAM SET/RESET process can monitor and detect oxygen ion drift.[51] As shown in Fig. 2, the G peak position after RESET is left shifted compared to that after SET, indicating the p-type doping becomes lighter. The intensity of the 2D peak is a function of the switching cycle and in each cycle, the 2D peak intensity of SET is always lower than that of RESET. These changes are mainly from the doping changes caused by the migration of oxygen ions. During the SET process, oxygen ions move and diffuse laterally toward graphene under the applied electrical field until they form covalent bonds with DB defects existing on the SLG surface. This process makes graphene p-type doping, which will increase the wave number of the G peak and decrease the intensity of the 2D peak and the resistance of SLG will be reduced.[53] During the RESET process, under the reversed external electrical field, oxygen ions move back to the oxide layer. This process would change the graphene towards intrinsic doping, leading to a decrease of the wave number of the G peak and an increase of the intensity of the 2D peak.

Fig. 2. (color online) The change of graphene G peak and 2D peak Raman shift after repeating the SET and RESET process. The figure is reproduced from Ref. [51].

XPS is a more convincing method to detect the oxygen ion drift as it can directly observe the covalent bond type change during the graphene RRAM SET/RESET process. Reference [54] fabricated SLG/Pt/Al2O3(∼ 1 nm)/TiO2/Pt RRAM and tested its XPS spectra of C 1s, O 1s, Ti 2p, and Al 2s binding states at HRS and LRS, respectively. The LRS is achieved by the SET process under a negative bias, which causes the formation of a TiOx suboxide at the TiO2 surface. After the RESET process, the bond fraction of C–O/C=O increases while that of the Ti3+ BE state decreases. O ions drift toward the SLG/Al2O3 interface and are assumed to convert to the surface Al–OH to Al–O bonds partially. Therefore, Al3+ in Al–O is decreased in LRS compared to that in HRS. From the above XPS analysis, we can figure out that the oxygen ions from the bottom side of the TiO2 layer are activated to move to the upper side by a positive bias applied, cross over the ultrathin Al2O3 layer, and bond with the SLG finally.

2.1.3. Higher density

In order to achieve high density RRAM, 3D RRAM using the edge of the 3D stacked metal plane as one of the RRAM electrodes is becoming prevalent in this field. But in order to make higher density RRAM, the total stack height (metal and dielectric) must be scaled down due to limitations of the etch aspect ratio.[55] Owing to graphene ultrathin thickness and excellent electrical conductivity, 3D RRAM using graphene as the electrode plane as shown in Fig. 3 will dramatically lower the height for single RRAM and make the density much higher than that using ordinary metal (such as Pt) without sacrificing RRAM performance.[56] It is assumed that if the top electrode trench has an 89° slope, the 3D RRAM can achieve up to 200 stacks. This memory window of G-RRAM is significantly (> 10×) larger than that of Pt-RRAM, which is attributed to ultrathin CF formed from the edge of the ultrathin graphene electrode. If the same methodology for the worst-case selected cell of 3D RRAM as in Ref. [57] is used to test with the criteria of a 75% write access voltage margin and an over 100 nA read margin, an array size of > 100 G for a 200-stacked graphene RRAM can meet both criteria, while an array less than 100k Pt-RRAM can only reach the read margin criterion. It is because lower height for 3D graphene RRAM results in overall low pillar resistance in a 3D stack architecture with cross point structure, which makes higher density 3D GRRAM more competitive in practical applications.

Fig. 3. (color online) The structure of graphene 3D stacked RRAM.
2.1.4. Well performed transparent electrode

Only ∼ 2% light is absorbed by graphene,[58] therefore graphene has the huge potential of being a transparent circuit used in electrode application. Compared to ITO-based transparent RRAM (TRRAM), using graphene as the RRAM transparent electrode not only gets high light transmittance in the large wavelength, but also improves the RRAM performance. For example, the graphene sheet resistance (∼ 30 Ω/□) is much lower than ITO[59] and the graphene/ZnO TRRAM[60] SET voltage (< 1 V) is much lower than that without graphene (∼ 3 V), and the sudden drop of leakage current in both SET and RESET processes will not happen, showing graphene-based TRRAM reversible and steady bipolar switching characteristics. Moreover, graphene can eliminate an undesired surface effect better and as a result a high yield of graphene RRAM can be obtained. The ordinary ITO/ZnO/ITO device exhibits a relatively low switching yield in a vacuum and O2 ambience (41.7% and 50.0%, respectively) as compared to the device in the ambience of N2 and air (66.7% and 58.3%, respectively). However for graphene-based devices, the switching yield is greatly increased in all four cases (66.7%, 66.7%, 75.0%, and 75.0% in a vacuum, O2, air, and N2, respectively) and insensitive to the environmental atmosphere, indicating that the graphene can be not only a transparent electrode material but also a passivation layer due to the weak chemisorption of O2 molecules.[61]

For graphene ZnO TRRAM, a forming process is needed, but if the dielectric layer is changed into a rare-earth-oxide-based thin film, the forming process is not necessary owing to its forming-free characteristics.[6264] Reference [65] fabricated multilayer grapheme(MLG)/Dy2O3/ITO TRRAM with 80% transmittance at 550 nm. Dy2O3 has a direct and wide bandgap of ∼ 5.5 eV at room temperature, which meets the intrinsic transparency requirement in resistive switching materials without sacrificing the memory density or capacity.[66] Due to excellent properties of the transparent electrode, the device exhibits unipolar resistance switching with low operation voltage (< 1 V), low operation current (< 100 μA), low power consumption (< 100 μW), high resistance ratio (> 104), reliable data retention, fast switching speed (< 60 ns), and promising cycle endurance (> 200 cycles), exhibiting the strong competitiveness with graphene TRRAM for which the forming process is needed.

2.1.5. Good uniformity

Uniformity measures the uniform level of parameters such as the resistance distribution in HRS or LRS. Long operational RRAM lifetime is dependent on it. Inserting graphene between the metal and dielectric will improve the RRAM uniformity, for example, Al/graphene/VOx/Cu RRAM[67] exhibits a good resistive switching at low current compliance (∼ 2 mA) while VOx RRAM without graphene needs a large one (∼ 20 mA). Although a little larger resistance distribution in LRS exists for the graphene-based RRAM compared to that without graphene, the uniformity of HRS is greatly improved, which indicates that the overall uniformity of VOx RRAM has been improved by using the embedded graphene.

2.1.6. Lowering SET voltage

With the integrated circuit scaling down, the supply voltage must be reduced correspondingly. Low SET voltage for RRAM is requested if it is integrated with other circuits. For some dielectrics such as NiO, the graphene will form a Schottky contact with it and enhance the electrical field intensity in the resistive switching layer, which will help the CF grow. As a result, the dielectric/graphene-based RRAM will own the lower SET and RESET voltages compared with that without graphene.[68]

Reference [69] used an anodic aluminum oxide (AAO) nanotemplate for the formation of the NiO RRAM nanocapacitor (Pt/NiO/graphene capacitor) array as shown in Fig. 4(a). The pore diameter in the nanotemplate is about 30 nm and the interpore distance is about 100 nm. The device is a vertical structure and exhibits the typical unipolar RRAM switching behavior with ON current below 1 nA, and as Fig. 4(c) shows, it has lower SET (∼ 1.2 V) and RESET (∼ 0.4 V) voltages than those on the surface of Nb-doped SrTiO3 no matter for the NiO RRAM nanocapacitors (1.5 V for SET and 0.5 V for RESET ) and the NiO thin film RRAM (2.5 V for SET and 1.5 V for RESET). Despite the work function of graphene (4.66 eV) being higher than that of Nb-doped SrTiO3 (4.2 eV),[70] the NiO/graphene Schottky contact is supposed to contribute to the lower SET and RESET voltages of the graphene NiO RRAM.

Fig. 4. (color online) NiO/graphene RRAMs with (a) vertical and (b) plane structure. (c) I–V curve comparison among RRAMs fabricated with NiO film, NiO dot, and NiO dot/graphene in vertical structure. (d) The repeating test results in vertical structure. Panels (a), (c), and (d) are reproduced from Ref. [69].

Reference [71] provided a plane structure with graphene nanoribbon (GNR)[72] as the electrode and the NiO nanodot covering the ∼ 25 nm width GNR nanogap as the switching layer, as shown in Fig. 4(b). The fabrication of monolayer graphene nanoribbon and deposition of the NiO nanodots onto the GNR nanogap are achieved by polystyrene dip-pen nanolithography (DPN). For this device, the SET and RESET voltages are ∼ 1.2 V and 0.4 V, respectively, which are similar to those with the vertical structure. Both vertical and plane structures own good reproducible switching behavior. After 200 switching cycles, the resistances of HRS and LRS change little for both devices.

2.1.7. Modulating SET voltage

A-B stacked bilayer grapheme (BLG) can open the bandgap by an external electrical field,[7376] which is different from single layer graphene with zero bandgap. For RRAM, this bandgap can be used to control the intensity of the electric field in the resistive switching layer and in turn, adjust the SET voltage.

Due to that the bandgap is determined by the external electrical field, extra bottom dielectric and gate electrode are needed. The structure in Ref. [77] is shown in Fig. 5(a). The main body of RRAM is Al/AlOx/BLG/Au and the voltage applied on the high doping Si substrate is used to provide an electrical field to open the BLG bandgap. When the gate voltage changes from −35 V to 35 V, the SET voltage for RRAM will increase from 0 V to 4.1 V (Figs. 5(b)5(d)), indicating that the SET voltage has a positive correlation with the gate voltage. It is because that for positive gate voltage, the bandgap will open and allow the electrical field from the gate to cross over the BLG into the resistance change material. The device enters the oxygen ions depleted state and the CF forming is prevented by the electrical field. In order for oxygen ions to screen the electrical field from the gate, the SET voltage is required to be higher to excite more oxygen ions and then form the CF. On the other side, the device enters the oxygen ions accumulated state for negative gate voltage. The bandgap is also opened and renders the electrical field to pass though BLG into the resistance switching layer, which in turn attracts more oxygen vacancies. As a result, the CFs could be formed under a lower SET voltage. By this mechanism, the gate voltage achieves the modulation of the SET voltage.

Fig. 5. (color online) (a) The structure of gated Al/AlOx/BLG/Au RRAM, where Al is used as top electrode, Au as bottom electrode, and Si as gate electrode. (b)–(d) I–V curves at different gate voltages, indicating the SET voltage and HRS/LRS ratio modulated by the gate voltage. (e), (f) The SET voltage changing trend varying with the gate. The figure is reproduced from Ref. [77].

The HRS/LRS ratio is also changed with the gate voltage, from 103 at −35 V to 105 at 35 V. It can be interpreted as that under a negative gate voltage, the oxygen vacancies will be attracted, resulting in the resistance decreasing of HRS and the HRS/LRS ratio. Conversely, under a positive gate voltage, some oxygen ions will be depleted, leading to the resistance increasing of the HRS and the HRS/LRS ratio.

Compared with the conventional one-diode-one-RRAM (1D1R) structure where the diode is used as a selector, this BLG RRAM can use the gate as the selector and owns the low OFF current, easy to achieve high density 3D stacking, and more compact merits.

2.1.8. Multilevel resistive switching

Multilevel switching has a great potential for RRAM on industrial memory applications.[7881] Using graphene as an electrode can also achieve stable and repeatable multilevel resistive switching behavior. Reference [82] used nanographene (NG) and nanogap to achieve RRAM multilevel switching. NG owns many appealing properties, such as tunable conductivity, scaled-up capability of integration, facile device fabrication, and fabrication compatible with modern CMOS technology, which make it very suitable to use as an electrode.

The device structure is similar to a back-gate nanographene transistor in its original state. By applying a certain voltage, a forming process between the two electrodes activates its resistive switching behavior in the new device. After the forming process, the electrical breakdown occurs and the nanogap appears in the middle of nanographene due to Joule heating.[83] This nanogap is served as the switching layer of RRAM.

The device shows unipolar resistive switching behavior. The SET voltage is ∼ 2.8 V and the RESET voltage is ∼ 5.5 V, an over 103 ON/OFF current ratio can be obtained at 1 V. The reversible reduction and oxidation of these silicon nanocrystallites in the SiO2 layer and the FN tunneling current[84] controlled by the external applied voltage are the key to the resistance switching effect during the SET and RESET procedures.

By changing the RESET voltage sweeping scope, the device shows multilevel switching properties. The resistance in HRS is relevant to the maximum sweeping voltage. When applying 3 V, 5 V, 7 V, 9 V, and 11 V bias pulses, the resistance changes to different values and varies widely. Besides the retention time can be up to 104 s, showing a very good stability.

2.2. Graphene as middle layer

For some RRAM devices, graphene is sandwiched between two dielectrics and served as the trapping site and is able to generate a local internal field to help filament grow, especially for an organic layer which is usually very thick and makes the external electrical field too weak to form a filament in it. For instance, the structure ITO/PMMA/GFs/PMMA/Al[85] shows great electrical bistable resistive switching property even the test temperature is up to 150 °C, 103 ON/OFF state current ratio and 105 s switching cycles can be obtained without obvious ON and OFF state degradation, which exhibits relatively large switching windows and good stability.

Apart from that, graphene as the middle layer is usually utilized for achieving multilevel switching, which is similar to Subsection 2.8, however the graphene is no longer used as an electrode but as a trapping site. Reference [86] fabricated organic RRAM with graphene nano flakes (GF) which are a good charge trapping[87] and storage[88] medium. This device uses polyvinylidene fluoride (PVDF) as the switching layer because of its better heat resistance, non-reactive nature, low weight, and flexibility. ITO on glass substrate is used as the bottom electrode. Two PVDF polymer layers are fabricated through the spin coating process and GR is obtained from GO by using reducing agent hydrazine monohydrate. The I–V curve shows double set and reset property. As the voltage increases, the current is growing but still in the OFF state. The index for the I–V curve is ∼ 2, which indicates that the space charge limited conduction (SCLC) mechanism[89] dominates the process. When the voltage reaches ∼ 1.8 V, the current suddenly rises and the resistance state jumps to an intermediate low resistance state from HRS. With voltage further going up, the second current jump occurs at ∼ 2.2 V and the device enters the total low resistance state. On the contrary, if a negative voltage applies on the device, the small current jump-down happens at −2.3 V and the large one happens at −2.9 V as the negative voltage becomes larger. The device finally enters the high resistance state.

The explanation of this phenomenon can be interpreted as follows. When a positive voltage is applied on the top gate, the electrons from the ITO electrode will be injected into the lowest unoccupied molecular orbital (LUMO) level of PVDF. As the voltage increases, these electrons sourced from the ITO electrode are captured by the electron trap nodes existing in GR. Once the graphene trapping sites are almost occupied by electrons, they can generate an electrical field strong enough to form a conducting path in the film and make the device transfer from HRS to LRS. The double SET process may attribute to the formation of multichannels with different trapping threshold potentials. On the contrary, when the applied voltage is reversed, it will de-trap the electrons from the graphene trapping sites and eject them back to the ITO electrode. Decreasing the voltage down to some degree, the conducting channel will be disconnected and the current will suddenly jump down. After all channels are broken down, the device will enter the total OFF state. Owing to it, the formation of extra conducting channels can be manipulated by limiting the current flow. Therefore controlling the compliance current can be used to achieve multi-bit storage. For this device, different LRS can be observed with different compliance current and at a certain reading voltage, multi-LRS is able to read out. In addition, multilevel switching can also realized by applying different SET and RESET pulses.

3. GO/rGO RRAM
3.1. rGO as electrode

rGO is a special kind of graphene which is reduced from GO, the thin rGO film behaves as graphene, therefore rGO can be used as the electrode as graphene is. Reference [90] used the laser-scribed technology to reduce GO to rGO (or laser-scribed grapheme (LSG)) and fabricated RRAM with rGO as the electrode, as shown in Fig. 6(a). The device is fabricated on a polythylene terephthalate (PET) flexible substrate, thus it has potential to be integrated into a flexible circuit. The forming voltage (∼ 1 V) is comparable to the devices’ SET voltage (∼ 1.2 V for the first SET), indicating that the device is forming free.[91] The ON/OFF current ratio is ∼ 10 due to the large HRS resistance. Under different current compliance, the device shows the potential of multi-level switching application. The retention time is over 104 and little degradation can be observed after 100 cycles, showing good reliability.

Fig. 6. (color online) (a) The structure of Ag/HfO2/rGO RRAM on PET substrate. (b) I–V curve of the device in the first cycle, 50th cycle, and 100th cycle. (c) I–V curve of the device under different current compliance, showing the potential for multilevel switching application. (d) The switching cycle test of the device in HRS and LRS. (e) The endurance test of the device. The figure is reproduced from Ref. [90].

By changing the material of the top gate from Ag to Pt (Figs. 7(c) and 7(d)) and conducting Ag-based RRAM LRS resistance temperature test as Fig. 7(a) shows, the Ag ions movement is confirmed as playing a dominant role in the growth of the conductive filament. The phenomenon of LRS resistance changing with temperature variation means that the filament growth is unlikely to be relevant to the oxygen vacancy.[92] High LRS resistance between 42 °C and 50 °C is attributed to the composite outcome between the electron scattering increasing and diameter of metal filament reduction as the temperature goes up. The high SET voltage of Pt-based RRAM further confirms that the Ag ions form filaments in Ag-based RRAM since that Pt cannot diffuse into HfO2 and it needs a high voltage for the oxygen vacancy to form the filament.

Fig. 7. (color online) (a) The relation between the resistance in LRS and temperature in Ag/HfO2/rGO RRAM. (b) The forming process for the Ag-based RRAM, indicating low forming voltage needed. (c), (d) I–V curves of Pt/HfO2/rGO under Pt and rGO as top electrode, respectively. The figure is reproduced from Ref. [90].
3.2. GO/rGO as dielectric

As Section 1 describes, GO film is cost-effective, easy to fabricate, flexible, etc. which is suitable to make low lost, flexible graphene circuits. For RRAM, GO is usually used as the switching layer and the simplest structure is the typical MIM structure: metal/GO/metal,[93] or MIS structure: metal/GO/semiconductor.[94] In GO MIM structure, a large range of metallic materials can be used as the electrode, such as metal elements, indium-tin-oxide (ITO),[95,96] and even grapheme.[97]

3.2.1. Model

When GO/rGO is used as dielectric, it becomes the place where the resistive switching happens. However, the mechanism of resistive switching in GO and rGO is different from that of conductive filament growth in most ordinary RRAMs. Therefore the accurate modeling of RRAM with GO/rGO as dielectric is very essential to have a deep understanding of its resistive switching behavior. For most of GO RRAMs published and introduced in the following section, the following reviewed models (such as space–charge-limited current model and Ohmic current model) are used widely to simulate GO/rGO RRAM.

Several models of GO and rGO have been proposed and ascertained.[98100] For RRAM, the potential difference between two electrodes will drive the sp3 and sp2 mutual conversion, which will increase or decrease the concentration of sp2 clusters and thus decide the conductivity of the GO film. The sp2 clusters growth rates can be express based on phenomenological Butler–Volmer kinetics[101] as

(1)
and the mechanism of electron transport between GO layers can be described by the multiphonon trap-assisted tunneling (MTAT) model[102] as
(2)
where νred, νOX, and νC represent the net reaction rates of the GO reduction process (sp3 to sp2 conversion), GO oxidation process (sp2 to sp3 conversion), and total sp3 to sp2 conversion, respectively; k0 is the kinetic constant of the electrochemical reaction, α(= 0.5) is the asymmetry factor, ΔrG0 is the free energy of the reaction at equilibrium, and F is the Faraday constant. Z is the charge-transfer number and η is the redox overpotential and can be estimated as the negative absolute of the applied voltage.[103] R and T are the ideal gas constant and temperature, respectively. For the MTAT model, Nt and ft are the spatial density and the electron occupation probability of the trap, respectively. and are the time constants for the electron capture and emission. Rc and Re are the capture and emission rates.

Equation (1) can be simplified as an Arrhenius-like[104] equation

(3)
which is similar to the filament growth rate in ordinary RRAM, where Eact is the amorphous carbon activation energy (∼ 3.5 eV), ν0 is the fitting parameter, and Vcell is the applied voltage.

The following space–charge-limited current (SCLC) model[105] is used to describe the current characteristics in HRS:

(4)
which can be simplified as JαV + βV2, where J and V are the current density and the applied voltage, respectively, A is a positive constant, μ is the charge carrier mobility, ε and d are the permittivity and the thickness of the GO film, and k and q are the Boltzmann constant and the unit electron charge, respectively. The Ohmic current model is used to fit the current in LRS due to the fact that the conductivity of the device in LRS is very large, the device in LRS can be simply treated as a simple linear resistance. These two models can fit the experimental data very well[106] and prevail in the GO RRAM theoretical analysis.

3.2.2. MIM structure GO RRAM

The first reported GO RRAM is the structure Cu/GO/Pt,[107] the thickness of GO by the vacuum filtration method is ∼ 30 nm. Apart from easy fabrication, the major merit of the device is the low switching voltage. The SET voltage distributes in the range from 0.3 V to 1 V and the RESET one is from −0.3 V to −0.9 V. Low SET and RESET voltages contribute to the diffusion of the top electrode under an external voltage and the adsorption/release of oxygen-related groups during the formation and rupture of conductive filament in the GO film. The current ON/OFF ratio is only 20, which is not very good, indicating that the GO resistance in HRS is not high enough. The retention time is more than 104 s and after 100 switching cycles, no obvious degradation of the ON/OFF ratio is observed, showing a relatively good stability.

Due to the flexibility of the GO film, RRAM can also be fabricated flexibly. Reference [108] proposed Al/GO/Al flexible RRAM as shown in Fig. 8(a), which is on the flexible polyethersulfone (PES) substrate. The GO film is fabricated by a simple and scalable spin-casting process[109] with GO solution and by this method, the GO film can be made large-area uniform and transferred in any kind of substrates. The device shows bipolar switching behavior and the ON/OFF ratio can reach up to 100 (Fig. 8(b)), smaller than that of the ordinary RRAM. But the retention time is over 105 s, competitive with the conventional one. After 100 switching cycles, the HRS/LRS memory window shows no noticeable degradation, which exhibits good stability and repetitiveness. The yield for this GO RRAM device is near 80%, which is good news for mass manufacture. It is noticed that the device shows negative voltage SET and positive voltage RESET and forming free characters. The resistive switching voltage is ∼ 2.5 V, a little higher for graphene-based RRAM.

Fig. 8. (color online) (a) The structure of flexible Al/GO/Al RRAM. (b) I–V curve of the device. (c) The bending time test for the RRAM. (d) Bending radius test in HRS and LRS. (e) The retention properties of the RRAM. (f) The endurance properties of GO RRAM. The figure is reproduced from Ref. [108].

The device shows good flexibility. When the bending radius enters the extremely flexed state (∼ 7 mm for the device), the current ON/OFF ratio can also be maintained as the degree of flat device as shown in Fig. 8(d). The resistive switching also results from filament growth. By the assistance of atomic-resolution TEM and energy-filtered TEM (EFTEM),[110] the filament is directly observed to be formed and de-formed in the very thin insulating layer at the top Al/GO interface under different applied voltages, rather than in the bulk film.

Other GO RRAMs with MIM structure have been deeply researched as well, such as metal/GO/Pt.[111] In a relatively large proportion of these RRAMs, the metal electrode ions diffusion dominates the filament growth and anticlockwise (AW) bipolar switching. They show good retention time (over 103 s) and cycle number (> 100), exhibiting high stability and well repeatability.[112] However, the type of electrode has a profound effect on the switching voltage and yield of GO RRAM. Taking metal/GO/Pt RRAM as an example, the metal here can be Ag, Ti, Cu, Au, etc. The forming voltages for these four metal RRAMs vary as follows: VAg < VTi < VCu < VAu. Compared to other metals (Ag, Ti, Cu), gold needs a high voltage to be oxidized to ions and activate Au/GO/Pt RRAM. The yields of these RRAMs vary as follows: YAg > YCu > YAu > YTi.[111] Ag-based GO RRAM has the maximum yield (reach up to ∼ 70%) and the lowest forming voltage among the four devices, indicating that Ag can be treated as the selective electrode for GO RRAM application.

3.2.3. RRAM with modified GO

For metal/GO/ITO RRAM, many modifications of GO have been made to improve the RRAM properties such as lower the SET voltage. One modification method is adding one more layer above or below the GO film. For example, reference [113] proposed Pr0.7Ca0.3MnO3 (PCMO)/GO RRAM by inserting PCMO between GO and the bottom electrode to realize oxygen exchange between GO and the PCMO layer, as shown in Fig. 9(a). PCMO owns a higher oxygen vacancy concentration at the surface region and can be used as a good ion conductor.[114] Both GO and PCMO films are deposited by a spin coating method at low temperature (300 ° ), annealed at 250 °C for 1 h and 150 °C for 30 min respectively after spin-coating. After that, Pt is deposited as the top electrode. This device shows reverse bipolar switching behavior. Due to PCMO high oxygen vacancy concentration, the oxygen in GO is inclined to move to the surface of PCMO, which is assumed to contain 5% higher oxygen vacancies than the bulk region,[115] making both GO and PCMO more conductive and therefore less voltage (−0.75 V for SET as Fig. 9(b) shows) is needed to achieve switching. Apart from that, the ON/OFF current ratio of the device is ∼ 102, the normal operating cycles can be up to 150, exhibiting good stability. As Fig. 9(d) shows, the retention time is up to 104 s even at test temperature 85 °C, showing good store retention property at high temperature.

Fig. 9. (color online) (a) The structure of PCMO/GO RRAM. (b) I–V characteristics of this device. The inset is I–V properties of RRAM without PCMO layer, which has a large SET voltage. (c), (d) Repeatability and endurance properties of PCMO/GO RRAM in HRS and LRS. The figure is reproduced from Ref. [113].

Another modification method is to make a GO composite film with other materials. The added materials can be classified as three types: nanoparticle (NP), polymer, and 2D materials such as MoS2. The following part will introduce the three types of materials in detail.

4. Concluding remarks

In this review, we have shown 8 main merits and roles of graphene as the electrode for RRAM, multilevel switching, and improvement of RRAM with graphene as the sandwiched middle layer, and graphene-based GO/rGO RRAM. However, the graphene RRAMs are so diverse that the lists are not able to include all of them. Compared to ordinary RRAM, the low power consumption, higher density, transparency, SET voltage controllable, etc. properties of graphene RRAM and lower SET voltage, flexible, easy to fabricate, etc. properties of rGO/GO RRAM make them have strong competitive power in the memory field, even though many issues for graphene are currently existing, such as growth and pattern of graphene are hard to integrate to the standard silicon manufacturing process, and the ON/OFF current ratio of GO RRAM is not high enough. With the fabrication process technology improving, we believe these issues will be solved and graphene RRAM will exhibit great potential on memory and neuromorphic applications separately in the further scaled down, flexible, or transparent circuit.

Reference
[1] Russo U Ielmini D Cagli C Lacaita A L 2009 IEEE Trans. Electron Devices 56 193
[2] Wong H-S P Lee H Y Yu S Chen Y S Wu Y Chen P S Lee B Chen F T Tsai M J 2012 Proc. IEEE 100 1951
[3] Bersuker G Gilmer D Veksler D Kirsch P Vandelli L Padovani A Iglesias V 2011 J. Appl. Phys. 110 124518
[4] Yu S Guan X Wong H S P 2011 IEEE International Electron Devices Meeting December 5–7, 2011 Washington, D.C., USA 17.3.1
[5] Hosoi Y Tamai Y Ohnishi T Ishihara K Shibuya T Inoue Y Yamazaki S Nakano T Ohnishi S Awaya N 2006 International Electron Devices Meeting December 11–13, 2006 San Francisco, USA 1
[6] Ielmini D 2011 IEEE Trans. Electron Devices 58 4309
[7] Wang L Yang C H Wen J 2015 Electron. Mater. Lett. 11 505
[8] Meena J S Sze S M Chand U Tseng T Y 2014 Nanoscale Res. Lett. 9 1
[9] Choi B J Torrezan A C Norris K J Miao F Strachan J P Zhang M X Ohlberg D A Kobayashi N P Yang J J Williams R S 2013 Nano Lett. 13 3213
[10] Lee M J Lee C B Lee D Lee S R Chang M Hur J H Kim Y B Kim C J Seo D H Seo S 2011 Nat. Mater. 10 625
[11] Wei Z Kanzawa Y Arita K Katoh Y Kawai K Muraoka S Mitani S Fujii S Katayama K Iijima M 2008 IEEE International Electron Devices Meeting December 15–17, 2008 San Francisco, USA 1
[12] Guan X Yu S Wong H S P 2012 IEEE Trans. Electron Devices 59 1172
[13] Yu S Guan X Wong H S P 2012 IEEE Trans. Electron Devices 59 1183
[14] Sheridan P Kim K H Gaba S Chang T Chen L Lu W 2011 Nanoscale 3 3833
[15] Celano U Chen Y Y Wouters D J Groeseneken G Jurczak M Vandervorst W 2013 Appl. Phys. Lett. 102 121602
[16] Chang W Y Lai Y C Wu T B Wang S F Chen F Tsai M J 2008 Appl. Phys. Lett. 92 2110
[17] Wu X Zhou P Li J Chen L Lv H Lin Y Tang T 2007 Appl. Phys. Lett. 90 183507
[18] Jeong H Y Kim Y I Lee J Y Choi S Y 2010 Nanotechnology 21 115203
[19] Son J Y Shin Y H Kim H Jang H M 2010 ACS Nano 4 2655
[20] Zhang H Liu L Gao B Qiu Y Liu X Lu J Han R Kang J Yu B 2011 Appl. Phys. Lett. 98 042105
[21] Wu Y Yu S Lee B Wong P 2011 J. Appl. Phys. 110 094104
[22] Wang Z Zeng F Yang J Chen C Yang Y Pan F 2010 Appl. Phys. Lett. 97 253301
[23] Chen Y C Chen C Chen C Yu J Wu S Lung S Liu R Lu C Y 2003 International Electron Devices Meeting December 8–10, 2003 Washington, D.C., USA 37.4.1
[24] Novoselov K Geim A K Morozov S Jiang D Katsnelson M Grigorieva I Dubonos S Firsov A 2005 Nature 438 197
[25] Orlita M Faugeras C Plochocka P Neugebauer P Martinez G Maude D K Barra A L Sprinkle M Berger C De Heer W A 2008 Phys. Rev. Lett. 101 267601
[26] Peres N Guinea F Neto A C 2006 Phys. Rev. B 73 125411
[27] Nair R R Blake P Grigorenko A N Novoselov K S Booth T J Stauber T Peres N M Geim A K 2008 Science 320 1308
[28] Lee C Wei X Kysar J W Hone J 2008 Science 321 385
[29] Sun Z S Zhang L Gao F 2016 Chin. Phys. B 25 0108701
[30] Jiang R Wu Z R Han Z Y Jung H S 2016 Chin. Phys. B. 25 0106803
[31] Lin Y M Jenkins K A Valdes-Garcia A Small J P Farmer D B Avouris P 2008 Nano Lett. 9 422
[32] Feng T Xie D Lin Y Zang Y Ren T Song R Zhao H Tian H Li X Zhu H 2011 Appl. Phys. Lett. 99 233505
[33] Tian H Xie D Yang Y Ren T L Wang Y F Zhou C J Peng P G Wang L G Liu L T 2012 Nanoscale 4 2272
[34] Schedin F Geim A Morozov S Hill E Blake P Katsnelson M Novoselov K 2007 2007 Nat. Mater. 6 652
[35] Liu N Tian H Schwartz G Tok J B H Ren T L Bao Z 2014 Nano Lett. 14 3702
[36] Li X Cai W An J Kim S Nah J Yang D Piner R Velamakanni A Jung I Tutuc E 2009 Science 324 1312
[37] Losurdo M Giangregorio M M Capezzuto P Bruno G 2011 PCCP 13 20836
[38] Hirata M Gotou T Ohba M 2005 Carbon 43 503
[39] Strong V Dubin S El-Kady M F Lech A Wang Y Weiller B H Kaner R B 2012 ACS Nano 6 139
[40] El-Kady M F Strong V Dubin S Kaner R B 2012 Science 335 1326
[41] Gómez-Navarro C Weitz R T Bittner A M Scolari M Mews A Burghard M Kern K 2007 Nano Lett. 7 3499
[42] Eda G Fanchini G Chhowalla M 2008 Nat. Nanotechnol. 3 270
[43] Tian H Shu Y Wang X F Mohammad M A Bie Z Xie Q Y Li C Mi W T Yang Y Ren T L 2015 Sci. Rep. 5
[44] Tian H Xie D Yang Y Ren T L Zhang G Wang Y F Zhou C J Peng P G Wang L G Liu L T 2012 Sci. Rep. 2 523
[45] Tian H Yang Y Xie D Ren T L Shu Y Zhou C J Tao L Q Liu L T 2013 IEEE 26th International Conference September 4–6 2013 Erlangen, Germany 709
[46] Tian H Yang Y Xie D Ge J Ren T L 2013 RSC Adv. 3 17672
[47] Tian H Shu Y Cui Y L Mi W T Yang Y Xie D Ren T L 2014 Nanoscale 6 699
[48] Wang X Tian H Mohammad M A Li C Wu C Yang Y Ren T L 2015 Nat. Commun. 6 7767
[49] Xia F Perebeinos V Lin Y M Wu Y Avouris P 2011 Nat. Nanotechnol. 6 179
[50] Chen H Y Tian H Gao B Yu S Liang J Kang J Zhang Y Ren T L Wong H S P 2012 IEEE International Electron Devices Meeting December 10–13, 2012 San Francisco, USA 20.5.1
[51] Tian H Chen H Y Gao B Yu S Liang J Yang Y Xie D Kang J Ren T L Zhang Y 2013 Nano Lett. 13 651
[52] Ferrari A C 2007 Solid State Commun. 143 47
[53] Ryu S Liu L Berciaud S Yu Y J Liu H Kim P Flynn G W Brus L E 2010 Nano Lett. 10 4944
[54] Jeon H Park J Jang W Kim H Ahn S Jeon K J Seo H Jeon H 2014 Carbon 75 209
[55] Chen H Y Yu S Gao B Liu R Jiang Z Deng Y Chen B Kang J Wong H P 2013 Nanotechnology 24 465201
[56] Sohn J Lee S Jiang Z Chen H Y Wong H S P 2014 IEEE International Electron Devices Meeting December 15–17, 2014 San Francisco, USA 5.3.1
[57] Deng Y Chen H Y Gao B Yu S Wu S C Zhao L Chen B Jiang Z Liu X Hou T H 2013 IEEE International Electron Devices Meeting December 9–11, 2013 Washington, D.C., USA 25.7.1
[58] Cheng H Chen S Yu P Duan X Xie B Tian J 2013 Appl. Phys. Lett. 103 203112
[59] Bae S Kim H Lee Y Xu X Park J S Zheng Y Balakrishnan J Lei T Kim H R Song Y I 2010 Nat. Nanotechnol. 5 574
[60] Yang P K Chang W Y Teng P Y Jeng S F Lin S J Chiu P W He J H 2013 Proc. IEEE 101 1732
[61] Sato Y Takai K Enoki T 2011 Nano Lett. 11 3468
[62] Chang W Chu J Wang S 2012 Appl. Phys. Lett. 101 012102
[63] Zhao H Tu H Wei F Xiong Y Zhang X Du J 2013 RRL-Rapid Res. Lett. 7 1005
[64] Cao X Li X Gao X Yu W Liu X Zhang Y Chen L Cheng X 2009 J. Appl. Phys. 106 073723
[65] Zhao H Tu H Wei F Du J 2014 IEEE Trans. Electron Devices 61 1388
[66] Thomas G 1997 Nature 389 907
[67] Lu H Sun K Zhang K Wang F Miao Y Wang B 2014 ECS Trans. 60 545
[68] Seo S Lee M Kim D Ahn S Park B H Kim Y Yoo I Byun I Hwang I Kim S 2005 Appl. Phys. Lett. 87 263507
[69] Son J Y Shin Y H Kim H Jang H M 2010 ACS Nano 4 2655
[70] Nagaraj B Wu T Ogale S Venkatesan T Ramesh R 2002 J. Electroceram. 8 233
[71] Kim W H Park C S Son J Y 2014 Carbon 79 388
[72] Li C X Deng X Q Sun L 2016 Acta Phys. Sin. 65 068503 in Chinese
[73] McCann E 2006 Phys. Rev. B 74 161403
[74] Ohta T Bostwick A Seyller T Horn K Rotenberg E 2006 Science 313 951
[75] Castro E V Novoselov K Morozov S Peres N Dos Santos J L Nilsson J Guinea F Geim A Neto A C 2007 Phys. Rev. Lett. 99 216802
[76] McCann E Koshino M 2013 Rep. Prog. Phys. 76 056503
[77] Tian H Zhao H Wang X F Xie Q Y Chen H Y Mohammad M A Li C Mi W T Bie Z Chao H Y Yang Y Wong H P Chiu P W Ren T L 2015 Adv. Mater. 27 7767
[78] Rozenberg M Inoue I Sanchez M 2004 Phys. Rev. Lett. 92 178302
[79] Waser R Aono M 2007 Nat. Mater. 6 833
[80] Lee S R Kim Y B Chang M Kim K M Lee C B Hur J H Park G S Lee D Lee M J Kim C J 2012 Symposium on VLSI Technology June 12–14, 2012 Honolulu, USA 71
[81] Wu M C Lin Y W Jang W Y Lin C H Tseng T Y 2011 IEEE Electron Device Lett. 32 1026
[82] He C Shi Z Zhang L Yang W Yang R Shi D Zhang G 2012 ACS Nano 6 4214
[83] Liao A D Wu J Z Wang X Tahy K Jena D Dai H Pop E 2011 Phys. Rev. Lett. 106 256801
[84] Yu S Guan X Wong H S P 2011 Appl. Phys. Lett. 99 063507
[85] Valanarasu S Kathalingam A Senthilkumar V Kannan V Rhee J 2012 Electron. Mater. Lett. 8 649
[86] Khurana G Misra P Katiyar R S 2014 Carbon 76 341
[87] Kane B 2010 Phys. Rev. B 82 115441
[88] Yang R Zhu C Meng J Huo Z Cheng M Liu D Yang W Shi D Liu M Zhang G 2013 Sci. Rep. 3 2126
[89] Rose A 1955 Phys. Rev. 97 1538
[90] Tian H Chen H Y Ren T L Li C Xue Q T Mohammad M A Wu C Yang Y Wong H S P 2014 Nano Lett. 14 3214
[91] Fang Z Yu H Li X Singh N Lo G Kwong D 2011 IEEE Electron Device Lett. 32 566
[92] Miao F Strachan J P Yang J J Zhang M X Goldfarb I Torrezan A C Eschbach P Kelley R D Medeiros-Ribeiro G Williams R S 2011 Adv. Mater. 23 5633
[93] Panin G N Kapitanova O O Lee S W Baranov A N Kang T W 2011 Jpn. J. Appl. Phys. 50 070110
[94] Jilani S M Gamot T D Banerji P Chakraborty S 2013 Carbon 64 187
[95] Yuan F Ye Y R Wang J C Zhang Z Pan L Xu J Lai C S 2013 IEEE 5th International Nanoelectronics Conference January 2–4 2013 Singapore 288
[96] Kapitanova O Panin G Kononenko O Baranov A Kang T 2014 J. Korean Phys. Soc. 64 1399
[97] Yi M D Guo J L Hu B Xia X H Fan Q L Xie L H Huang W 2015 Chin. Phys. Lett. 32 077201
[98] Eda G Mattevi C Yamaguchi H Kim H Chhowalla M 2009 J. Phys. Chem. C 113 15768
[99] Ekiz O O Urel M Guner H Mizrak A K D’ana A 2011 ACS Nano 5 2475
[100] Lim E W Ahmadi M T Ismail R 2016 J. Comput. Electron. 15 602
[101] Herrmann M Schenk A 1995 J. Appl. Phys. 77 4522
[102] Bocquet M Deleruyelle D Muller C Portal J M 2011 Appl. Phys. Lett. 98 263507
[103] Grierson D S Sumant A Konicek A Friedmann T Sullivan J Carpick R W 2010 J. Appl. Phys. 107 033523
[104] Laidler K J 1984 J. Chem. Educ. 61 494
[105] Lampert M A 1956 Phys. Rev. 103 1648
[106] Zhuang X D Chen Y Liu G Li P P Zhu C X Kang E T Noeh K G Zhang B Zhu J H Li Y X 2010 Adv. Mater. 22 1731
[107] He C Zhuge F Zhou X Li M Zhou G Liu Y Wang J Chen B Su W Liu Z 2009 Appl. Phys. Lett. 95 232101
[108] Jeong H Y Kim J Y Kim J W Hwang J O Kim J E Lee J Y Yoon T H Cho B J Kim S O Ruoff R S 2010 Nano Lett. 10 4381
[109] Lee D H Kim J E Han T H Hwang J W Jeon S Choi S Y Hong S H Lee W J Ruoff R S Kim S O 2010 Adv. Mater. 22 1247
[110] Kim S K Kim J Y Choi S Y Lee J Y Jeong H Y 2015 Adv. Funct. Mater. 25 6710
[111] Zhuge F Hu B He C Zhou X Liu Z Li R W 2011 Carbon 49 3796
[112] Wang Z Tjoa V Wu L Liu W Fang Z Tran X Wei J Zhu W Yu H 2012 J. Electrochem. Soc. 159 K177
[113] Kim I Siddik M Shin J Biju K P Jung S Hwang H 2011 Appl. Phys. Lett. 99 042101
[114] Li S L Liao Z Li J Gang J Zheng D 2009 J. Phys. D: Appl. Phys. 42 045411
[115] Nian Y Strozier J Wu N Chen X Ignatiev A 2007 Phys. Rev. Lett. 98 146403
[116] Hu B Quhe R Chen C Zhuge F Zhu X Peng S Chen X Pan L Wu Y Zheng W 2012 J. Mater. Chem. 22 16422
[117] Wu C Li F Zhang Y Guo T Chen T 2011 Appl. Phys. Lett. 99 042108
[118] Li Y Ni X Ding S 2015 J. Mater. Sci. Mater. Electron. 26 9001
[119] Myung S Park J Lee H Kim K S Hong S 2010 Adv. Mater. 22 2045
[120] Wang W Ma D G 2010 Chin. Phys. Lett. 27 018503
[121] Midya A Gogurla N Ray S K 2015 Curr. Appl Phys. 15 706
[122] Pinto S Krishna R Dias C Pimentel G Oliveira G Teixeira J Aguiar P Titus E Gracio J Ventura J 2012 Appl. Phys. Lett. 101 063104
[123] Radisavljevic B Radenovic A Brivio J Giacometti i V Kis A 2011 Nat. Nanotechnol. 6 147
[124] Shao P Z Zhao H M Cao H W Wang X F Pang Y Li Y X Deng N Q Zhang J Zhang G Y Yang Y Zhang S Ren T L 2016 Appl. Phys. Lett. 108 203105
[125] Cao H W Zhao H M Xin X Shao P Z Qi H Y Jian M Q Zhang Y Y Yang Y Ren T L 2016 Mod. Phys. Lett. B 30 1650267
[126] Podzorov V Gershenson M Kloc C Zeis R Bucher E 2004 Appl. Phys. Lett. 84 3301
[127] Liu W Kang J Sarkar D Khatami Y Jena D Banerjee K 2013 Nano Lett. 13 1983
[128] Yin Z Zeng Z Liu J He Q Chen P Zhang H 2013 Small 9 727