INTRODUCTION
Over the past few decades, computing systems based on the von Neumann architecture, which rely on centralized and sequential operations determined by a clock, have proven to be more powerful than the human brain in solving complex and well-structured mathematical problems1-3. However, due to increasing computational complexity and power consumption, computer systems based on the von Neumann architecture have encountered the bottleneck of the separation of memory and central processing units4,5. At the same time, the performance gap between memory and central processing units is gradually increasing, which is not suitable for solving unstructured problems and consumes a lot of energy in the process of processing large amounts of data6-9.
As an emerging passive electronic device that is different from traditional resistors, capacitors, and inductors, memristive devices exhibit the characteristics of both storage and computing, and are endowed with outstanding advantages in terms of integration density, computing speed and power consumption10-14. The development of high-performance memristors is also urgently needed for the development of advanced brain-like neuromorphic systems15-20. Neuromorphic systems are seen as an effective solution for processing large amounts of complex data20,21. Materials used for memristor processing include ferroelectric materials22-24, electrolytes25,26, organic semiconductors27-29, phase-change materials30,31, metal oxide semiconductors32,33, perovskites34-37, quantum dots38,39, covalent organic frameworks thin films40, supramolecular materials41, transition metal dichalcogenides42,43, polymers44-46, peptides47,48, DNA49, zeolites50, liquid metals51, protein52,53, covalent organic frameworks54, carbon nanomaterials55-57, etc. Among them, carbon nanomaterials have multiple dimensions, covering zero-dimensional (0D), one-dimensional (1D), and two-dimensional (2D). In particular, memristors based on carbon nanomaterials have become strong candidates for realizing high-quality information storage, high-performance neuromorphic applications, and high-sensitivity bionic sensing due to their excellent device stability and weight tunability58,59.
In this paper, emphasis was laid upon the latest research progress of carbon nanomaterial memristors for resistive random access memory (RRAM) and neuromorphic applications. The performance characteristics of 0D, 1D, and 2D nanomaterials and synthesis methods were firstly introduced. Then, the processing methods, performance, intrinsic mechanisms, and application scenarios of RRAM and neuromorphic applications based on carbon nanomaterial memristors were well discussed. Finally, the future development directions of carbon-based memristors were also discussed. It is believed that this paper can provide a comprehensive overview of the existing research results and future challenges of memristors based on carbon nanomaterials (Fig. 1) and open up a new perspective for the future development of carbon-based memristors that integrate sensing, memory, and computing.
Fig. 1. RRAM and neuromorphic devices based on carbon nanomaterial memristors60-69. Reprinted with permission from refs.60-69. © 2015 American Chemical Society. Reprinted with permission from ref.61. © 2023 The Author(s). © 2021 The Minerals, Metals & Materials Society. © 2012 Royal Society of Chemistry. © 2021 Royal Society of Chemistry. © 2021, 2022, 2023 American Chemical Society. © 2023 Elsevier Ltd. Abbreviation: RRAM, resistive random access memory. |
CARBON NANOMATERIALS
Carbon nanomaterials are usually obtained through various hybridisations of pure carbon elements and are allotropes of carbon. Carbon nanomaterials can be divided into 0D, 1D, and 2D in terms of dimensions, as shown in Fig. 2. 0D carbon nanomaterials can be divided into carbon quantum dots (CQDs)60, graphene quantum dots (GQDs), graphene oxide (GO) quantum dots, fullerenes70, etc. In 1D carbon nanomaterials, electrons move freely only in non-nanoscale directions and the motion is linear. 1D carbon nanomaterials mainly include single-walled carbon nanotubes (SWCNTs), multi-walled carbon nanotubes (MWCNTs)61, and carbon nanofibers71. 2D carbon nanomaterials mainly including graphene, GO, reduced GO72, and MXene73, composed of transition metal carbides, have mechanical, electrical, and chemical properties that surpass those of traditional materials and are used in photodetectors, sensors and memristors, which have attracted extensive research in the field of instrumentation.
Fig. 2. Carbon nanomaterials. a, Carbon quantum dots60. Reprinted with permission from ref.60. © 2015 American Chemical Society. b, Graphene quantum dots. c, Graphene oxide quantum dots70. Reprinted with permission from ref.70. © 2013 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. d, Fullerene74. Reprinted with permission from ref.74. © 2023 The Author(s). e, Single-walled carbon nanotubes. f, Multi-walled carbon nanotubes61. Reprinted with permission from ref.61. © 2023 The Author(s). g, Carbon nanotubes fiber71. Reprinted with permission from ref.71. © 2021 Elsevier B.V. h, Graphene. i, Graphene oxide. j, Reduced graphene oxide72. Reprinted with permission from ref.72. © 2023 Elsevier B.V. k, MXene73. Reprinted with permission from ref.73. © 2023 The Author(s). |
Zero-dimentional carbon nanomaterials
CQDs refer to carbon nanoparticles (NPs) with a size of less than 10 nm. They are sp3 hybridized and endowed with the advantages of excellent conductivity, environmental friendliness, low toxicity, and simplicity of modification75. At the same time, the surface defects of CQDs can control the generation and recombination of electron-hole pairs, thereby realizing the memristor function of the memristor. The top-down method mainly refers to the use of chemical or physical methods to decompose larger carbon structures into CQDs, which is beneficial for the synthesis of CQDs with complete nanostructures and high crystallinity76. Top-down methods include pyrolysis, arc discharge, laser ablation, chemical oxidation, electrochemistry, etc. Thamankar et al. obtained CQD solutions dispersed in methanol by annealing, grinding, dissolving (methanol solution), ultrasonicating and filter paper filtration at different temperatures (200 °C, 220 °C, and 240 °C). The specific steps are shown in Fig. 3a. The size of the synthesized CQDs is between 4.5 and 2.1 nm and exhibits fluorescent properties under ultraviolet (UV)-light irradiation77. Liu et al. used pyrolysis of Ginkgo leaves, as shown in Fig. 3b, to obtain highly crystalline CQDs. The CQDs yield reached 21.7%, proving that it is feasible to use leaf decomposition to produce CQDs on a large scale78. Scrivens et al. used arc-discharge single-walled nanotubes to synthesize CQDs79. However, CQDs prepared by the arc-discharge method have the disadvantage of being difficult to purify. Vega et al. used a laser with a wavelength of 1064 nm to irradiate a carbon particle jet solution with polyethylene glycol as the solvent, and the irradiation time was 3 h. The colour of the solution changed from grey to caramel colour, indicating the formation of CQDs, and the average size of the prepared CQDs was 3.57 nm80. A novel method of synthesis CQDs using a laser to ablate graphite under vacuum conditions has been proposed by Malekfar et al. The laser ablation product is then acid treated and surface passivated using nitric acid and ethanol, and CQDs are finally obtained. Laser ablation requires the construction of a specialized laser ablation system, as shown in Fig. 3c, and is usually accompanied by complex post-processing processes81.
Fig. 3. Carbon quantum dots. a, Pyrolysis of jaggery77. Reprinted with permission from ref.77. © 2023 Author(s). b, Pyrolysis of leaves78. Reprinted with permission from ref.78. © 2018 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. c, Laser ablation81. Reprinted with permission from ref.81. © 2017 Elsevier Ltd. d, Mixed acid oxidation82. Reprinted with permission from ref.82. © 2017 Science China Press and Springer-Verlag GmbH Germany, part of Springer Nature 2021. e, Microwave synthesis83. Reprinted with permission from ref.83. © 2009 The Royal Society of Chemistry. f, UV-visible absorption spectrum of carbon quantum dots84. Reprinted with permission from ref.84. © 2014 American Chemical Society. g, Hydrothermal synthesis85. Reprinted with permission from ref.85. © 2017 The Royal Society of Chemistry. h, MOF-assisted synthesis86. Reprinted with permission from ref.86. © 2017 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim. Abbreviations: MOF, metal organic framework; UV, ultraviolet. |
Therefore, the complex process and high cost restrict the application of laser ablation in preparing CQDs. In chemical oxidation technology, strong oxidants (hydrogen peroxide, nitric acid, sulfuric acid) are used to treat carbon compounds, allowing for the quick and convenient acquisition of large quantities of CQDs. Li et al. used a mixed acid solution of nitric acid and sulfuric acid to oxidize fullerene and obtain CQDs with high fluorescence quantum yield (3%-5%), as shown in Fig. 3d82. In addition, electrochemical synthesis of CQDs is an efficient and economically feasible technology. Zhou et al. used a chemical vapor deposition (CVD) method to prepare MWCNTs on carbon paper and then cut the carbon-nanotube-covered carbon paper into appropriate sizes and placed them in a Teflon sheath for use as electrodes. 0.1 M tetrabutylammonium perchlorate was added to a three-electrode system consisting of a carbon paper working electrode, a platinum wire electrode, and a Ag/AgClO reference electrode, and CQDs were synthesized in a degassed acetonitrile solution. The synthesized CQD solution appears bright blue under the excitation of UV light87.
The bottom-up approach is beneficial for the generation of CQDs with rich doping sites and amorphous structures. In the bottom-up method of synthesizing CQDs, organic molecules such as sugars and citric acid are typically used as raw materials for the synthesis of CQDs. The main methods for bottom-up synthesis of CQDs include microwave method, hydrothermal synthesis, and metal organic framework (MOF)-template-assisted method. Zhu et al. used polyethylene glycol, glucose, and fructose as raw materials, mixed polyethylene glycol, and heated it in a 500-W microwave. The color of the solution changed from colorless to yellow and finally to dark brown, proving that CQDs have been synthesized, as shown in Fig. 3e83. Zhang et al. prepared a mixed solution by using citric acid, ethylenediamine, and ultrapure water. The mixed solution was added into the high-pressure hydrothermal reaction kettle and was placed in an oven with an oven at a temperature of 200 °C for a heating time of 5 h. Finally, CQDs were obtained through a series of methods such as dialysis, rotary evaporation, and freeze-drying, and under UV-light irradiation, CQDs appear blue, as shown in Fig. 3f84. Mohapatra et al. used orange juice as the raw material and the hydrothermal method to synthesize CQDs. The specific process is shown in Fig. 3g. Low rotation speed can synthesize large particles with low fluorescence, and high rotation speed can synthesize small particles with high fluorescence85. Zhang et al. used MOF as a template to synthesize ultra-small-sized CQDs by impregnating glucose and combining low-temperature calcination. The synthesis process is shown in Fig. 3h. The size of the synthesized CQDs is very close to the pore size of the MOF-template, proving that the method of MOF-template-assisted synthesis of CQDs can well control the size of CQDs86.
GQDs are nanocrystals composed of sp2 hybridized graphene particles. They exhibit the properties of variable energy levels, electroluminescence, charge storage and high crystallinity. Top-down methods include oxidative lysis, hydrothermal, ultrasonic, and microfluidic methods. Oxidative cleavage, also known as oxidative cutting, achieves the synthesis of GQDs through the destruction of chemical bonds in graphene by strong oxidants. Li et al. used HNO3 to cut GO into small pieces and then used ethylene glycol passivation and hydrazine hydrate reduction treatment to obtain GQDs with a diameter of 5-19 nm88. Zhu et al. used concentrated H2SO4 and concentrated HNO3 solutions with a volume ratio of 3 : 1 to oxidize carbon fibers, as shown in Fig. 4a. The particle size of GQDs is controlled by controlling the reaction temperature. The thickness of the synthesized GQDs is between 0.4 and 2 nm89. In addition, graphite90 and fullerene91 can be used as raw materials for oxidative cracking to produce GQDs. The processing flow is shown in Fig. 4b and c. The synthesis of GQDs by oxidative pyrolysis method can achieve large-scale synthesis, but the synthesis process requires the use of strong oxidants, and excess strong oxidants are difficult to remove. The hydrothermal method is a simple, economical, and environmentally friendly method. The thermal shear force generated by high temperature and high pressure is used to gradually peel off the bulk carbon materials into GQDs. Correa et al. achieved the cleavage of GO by optimizing the GO concentration, pH value, and hydrothermal temperature and obtained GQDs with a quantum yield of 8.9%. The specific steps are shown in Fig. 4d92. The most obvious advantage of the hydrothermal method is that the precursors for synthesizing GQDs are widely available, but the disadvantage is that other impurities may be prensent in the produced GQDs. The energy generated by ultrasound can also be used to break bulk carbon materials into low-dimensional nanostructures. However, the energy of ultrasound is limited and is not enough to completely break the bulk carbon materials into GQDs93. Therefore, the method of synthesizing GQDs using ultrasound in combination with hydrothermal, oxidation, and other methods has been applied. Shao et al. used concentrated H2SO4 and HNO3 to oxidize graphene at room temperature and then used an ultrasonic instrument to ultrasonicate the mixed solution for 12 h. The solution was calcinated at 350 °C for 20 min to remove concentrated H2SO4 and HNO3 from the mixed solution. The calcined product was dispersed into water and filtered using a 0.22-μm microporous membrane to obtain a brown solution. Finally, the brown solution was placed in a dialysis bag overnight to obtain GQDs94. In order to reduce contamination during the synthesis of GQDs, Regev et al. proposed a strategy for the physical synthesis of GQDs based on microfluidic technology. The graphite aqueous solution flows through the Z-shaped pipe under the condition of pressurization by a high-pressure water pump. The internal crushing process of the Z-shaped pipe is shown in Fig. 4e. Graphite flakes are exfoliated into graphene sheets and are further fragmented into nano-sized GQDs without the introduction of chemicals95.
Fig. 4. Graphene quantum dots. a, Carbon-fiber oxidative pyrolysis89. Reprinted with permission from ref.89. © 2012 American Chemical Society. b, Graphite oxidative pyrolysis90. Reprinted with permission from ref.90. © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. c, Fullerene oxidative pyrolysis91. Reprinted with permission from ref.91 © 2015 American Chemical Society. d, Graphene oxide hydrothermal92. Reprinted with permission from ref. 92. © 2021 Elsevier Ltd. e, Graphite microfluidic95. Reprinted with permission from ref.95. © 2015 American Chemical Society. f, Pyrene solvothermal96. Reprinted with permission from ref.96. © 2014 Macmillan Publishers Limited. g, Glucose microwave97. Reprinted with permission from ref.97. © 2012 American Chemical Society. h, Norepinephrine microwave98. Reprinted with permission from ref.98. © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. |
Methods for bottom-up synthesis of GQDs include thermal decomposition as well as solvothermal and microwave methods. Wu et al. obtained GQDs by direct pyrolysis of citric acid at the temperature of 200 °C. During the pyrolysis process, citric acid liquefies first, and the color of the liquid changes from colorless to light yellow. Finally, the liquid turned orange, indicating the formation of GQDs99. The advantages of the pyrolysis method are low cost and ease of operation, but the synthesized GQDs have the problem of low purity. Pan et al. used a solvothermal method to synthesize GQDs. The specific process included nitration of pyrene and alkaline hydrothermal treatment and achieved a gram-level preparation process of single-crystal GQDs, as shown in Fig. 4f96. The microwave method uses the heat energy generated by microwave irradiation to destroy the chemical bonds in the GQDs’ precursor, thereby triggering polymerization and carbonization and achieving the goal of uniform and rapid synthesis of GQDs. Lau et al. added a series of glucose solutions of different concentrations into glass bottles and found that the growth of GQDs could be controlled by controlling the experimental parameters such as the heating power and heating time of the microwave oven, as shown in Fig. 4g97. Kim et al. used a household microwave oven to heat an aqueous solution of norepinephrine and obtained N-GQDs with a diameter of 3.4 nm. After changing the solution from water to dimethyl sulfoxide, NS-GQDs can be synthesized, as shown in Fig. 4h98. However, the microwave method has certain disadvantages, such as high energy consumption, and is not suitable for large-scale commercial production.
One-dimensional carbon nanomaterials
Carbon nanotubes (CNTs) can be simply understood as rolled honeycomb structures of sp2 hybridized carbon, as shown in Fig. 5a100. CNTs have the advantages of large elastic modulus, high strength, high thermal stability, and high electrical conductivity and can be used in the development of electronic devices101,102. The main synthesis methods of CNTs include arc discharge, laser ablation, CVD, and flame synthesis. The arc-discharge method to synthesize CNTs uses high temperature (>3000 °C) to evaporate carbon atoms into plasma, thereby forming SWCNTs and MWCNTs. The arc-discharge method is used to synthesize SWCNTs, which requires the use of catalysts such as yttrium, nickel, and iron. Synthesis of MWCNTs is done using the arc-discharge method without mandatory use of catalysts103.
Fig. 5. Carbon nanotubes. a, Single-wall and multi-wall carbon nanotube structures100. b, Arc-discharge method104. Reprinted with permission from ref.104. © 2014 Elsevier B.V. c, Laser ablation105. Reprinted with permission from ref.105. © 2019 by the authors. d, Dense array of SWCNTs106. Reprinted with permission from ref.106. © 2009 American Chemical Society. e, Preparation of W–Co nanocrystal catalyst and template growth of SWCNTs107. Reprinted with permission from ref.107. © 2014 Macmillan Publishers Limited. f, Metal molybdenum and rhenium oxides are used as catalysts to synthesize SWCNTs108. Reprinted with permission from ref.108. © 2020 The Authors, g, Rapid annealing109. Reprinted with permission from ref.109. © 2012 Elsevier B.V. h, Flame synthesis method110. Reprinted with permission from ref.110. © 2019 Taylor & Francis Group, LLC. |
The schematic diagram of the arc-discharge chamber is shown in Fig. 5b. The arc generates very high-temperature plasma, which sublimates the carbon precursor filled in the anode. After the discharge stopped, CNTs can be collected from the chamber walls104. Laser ablation is considered a green, environmentally friendly, and low-cost method effective for patterning CNTs. Ho et al. used an Nd:YAG laser with a wavelength of 1604 nm to perform laser ablation of MWCNT films deposited on polyethylene terephthalate (PET). The scanning electron microscopy results showed that laser ablation can obtain clear patterned CNT films111. Wu et al. built a femtosecond laser system to ablate CNTs, as shown in Fig. 5c. When the laser energy is 100 nJ and the scanning speed is 0.1 mm s−1, the CNT patterning has the optimal effect, which provides a reliable method for the subsequent manufacturing of CNT thin-film transistors105. CVD is used in the field of synthesizing CNTs because of its low cost, high purity, and controllable growth. Liu et al. proposed a method to prepare SWCNTs by CVD, using ethanol/methanol as the carbon source to prepare SWCNTs and Cu NPs as the catalyst and growing a dense array of SWCNTs on a quartz substrate, as shown in Fig. 5d. The introduction of methanol combined with the interaction between SWCNTs and quartz lattice is the key to directional growth of nanotubes106. Li et al. synthesized SWCNTs using tungsten-based bimetal alloy nanocrystals as catalysts and ethanol as the carbon source, as shown in Fig. 5e. The addition of a catalyst significantly increased the abundance of SWCNTs produced, reaching an abundance of 92%107. In addition, metal molybdenum and rhenium oxides can also be used as catalysts to promote the synthesis of SWCNTs. The specific process is shown in Fig. 5f108.
In order to improve the consistency of MWCNTs generated by the CVD method, Zhang et al. proposed to improve the consistency of CNT growth through rapid annealing, as shown in Fig. 5g. The graphite tubes were wrapped in carbon felt for thermal insulation, and the MWCNTs were annealed using a DC voltage in an argon atmosphere. Compared to unannealed CNTs, annealed CNTs have significantly reduced disorder and fewer defects109. Flame synthesis is a method for large-scale synthesis of CNTs. Liu et al. used the reverse diffusion flame method to compensate for the poor flame stability and low carbon source of normal diffusion flames. The designed pre-evaporated gas reverse diffusion combustion system is shown in Fig. 5h. The burner is composed of three concentric tubes. The outer tube transports the protective gas N2, the middle tube transports a mixture of liquid fuel (ethanol, n-butanol) and N2, and the inner tube transports air. A copper-alloy mesh (substrate) and a nickel mesh (catalyst) are placed above the flame to collect CNTs. The characterization results prove that high-quality CNTs can be synthesized in an ethanol flame of 1273 K with a burning time of 15 min110.
Carbon fiber is a new type of fiber material with a carbon content of more than 90%. It has good electrical conductivity. There are many methods for carbon fiber growth, such as solution spinning, melt spinning, electrospinning, template method, wet spinning, dry-jet wet spinning, gel spinning, polymerization, CVD, etc. Solution spinning technology has been used to prepare various carbon fibers with excellent properties, including surfactant suspended CNT spinning. Poulin et al. used sodium dodecyl sulfate (SDS) as a surfactant to uniformly disperse SWCNTs in an aqueous solution and then injected the dispersion through a syringe into a coagulation solution containing polyvinyl alcohol (PVA, 5 wt.%), as shown in Fig. 6a. Due to the amphipathic nature of PVA, it is adsorbed on the CNT to replace some SDS molecules and then form CNT ribbons. CNT-based carbon fibers are prepared by twisting CNT ribbons due to the presence of capillary action. Melt spinning is a good method to prepare carbon fibers when pitch with a higher viscosity is selected as the precursor. Under inert gas conditions, the asphalt is heated to a molten state, and the asphalt is extruded from the nozzle to form fibers by applying air pressure. The asphalt fibers are rapidly cooled and solidified in the air and are finally wound on a spool, as shown in Fig. 6b112.
Fig. 6. Carbon fiber. a, Solution spinning. b, Melt spinning112. Reprinted with permission from ref.112. © 2022 Donghua University, Shanghai, China. c, Electrospinning113. Reprinted with permission from ref.113. © 2023 Published by Elsevier B.V. d, Template method114. Reprinted with permission from ref.114. © 2023 Elsevier B.V. e, Wet spinning, dry-jet wet spinning, and gel spinning equipment112. f, Polymerization law115. Reprinted with permission from ref.115. © 2019 Elsevier B.V. |
Electrospinning is used for carbon fiber production and has the advantages of low cost, high efficiency, and a wide range of precursors. Precursors that can be used for electrospinning include polyacrylonitrile (PAN), polyimide (PI), PVA, polyvinyl chloride, etc. Electrospinning equipment generally includes pumps, needles, high-voltage generators, and collectors, as shown in Fig. 6c113. A high-voltage generator creates an electric field between the needle and collector, which attracts precursors from the needle tip to the collector. Finally, non-carbon elements in the precursor are removed through heat treatment to obtain carbon fibers. On the basis of electrospinning, Guo et al. added polymethyl methacrylate (PMMA) and β-cyclodextrin as templates into the electrospinning precursor solution and subsequently used electrospinning to silk, muffle furnace oxidation, and high-temperature carbonization in a tube furnace at 800 °C (under N2 atmosphere) to prepare black porous carbon nanofibers, as shown in Fig. 6d. The introduced double template is completely decomposed at a high temperature to form a porous structure in the carbon fiber, which significantly enhances the mechanical flexibility of the carbon fiber114.
When PAN is selected as the precursor, wet spinning, dry-jet wet spinning, and gel spinning can be used, as shown in Fig. 6e. After completing the PAN precursor treatment, carbon fibers are subsequently prepared through high-temperature carbonization112. The polymerization method has attracted the interest of researchers due to its advantages such as fast reaction speed, simple process flow, and mild reaction conditions. Li et al. used a structural guide (methyl orange) and an oxidizing agent (ferric chloride) to form fiber precipitates to synthesize polypyrrole nanofibers. Then using methods such as microwave carbonization, fiber surface functionalization, and secondary polymerization, sea cucumber-like hollow CNTs were obtained, as shown in Fig. 6f115. In addition, CVD is a potential method for continuous batch production of carbon nanofibers. The specific process is divided into reactive gas diffusion, gas deposition on the substrate surface, and chemical reaction on the substrate surface. Chen et al. thermally decomposed copper tartrate under an argon atmosphere to obtain copper NPs. Using copper NPs as a catalyst to promote the thermal decomposition of acetylene, carbon fibers with a smooth surface and a diameter of 100-200 nm were obtained116.
Two-dimensional carbon nanomaterials
Since its discovery in 2004, graphene has has been the subject of intense researches due to its excellent mechanical properties, ultra-wide absorption spectrum, ultra-large specific surface area, ultra-high thermal conductivity, and electrical conductivity117,118. There are many methods for the preparation of graphene, including mechanical exfoliation, epitaxial growth, CVD, and other methods. Geim et al. used tape to mechanically exfoliate graphene from graphite and successfully prepared graphene with a thickness of only a few atomic layers thick. The graphene obtained by mechanical exfoliation has high purity, but the efficiency is low and is not suitable for large-scale production119. Therefore, researchers proposed to use epitaxial growth method to obtain graphene, remove Si atoms from SiC under ultra-vacuum conditions, and use the reconstruction of C atoms on the SiC surface to form graphene120.
However, in the process of obtaining graphene using the epitaxial growth method, the cost of removing other elements in the substrate is high. The CVD technology has been shown to be able to easily fabricate singlelayer and multilayer graphene samples with an area of several square centimeters and be transferred to other substrates. The preparation of graphene using CVD is mainly divided into the following steps: (1) Deposit nickel on the substrate and heat the substrate. (2) The nickel-containing substrate is heated to 1000 °C, passed by a carbon-containing gas, and carbon atoms are deposited onto the substrate. (3) Graphene is crystallized by cooling the nickel-containing substrate, and the thickness of graphene is controlled by the cooling rate and the concentration of carbon-containing gas. (4) Chemical etching separates the graphene from the nickel-containing substrate, as shown in Fig. 7a121. However, graphene has no forbidden band, making it difficult to achieve switching behaviour and control current flow, which poses challenges in electronic device applications122.
Fig. 7. Graphene and its derivatives. a, Preparation of graphene oxide by CVD121. Reprinted with permission from ref.121. © 2009 Macmillan Publishers Limited. b, Preparation of graphene oxide by improved Hummers’ method123. Reprinted with permission from ref.123. © 2016 The Author(s). c, Preparation of graphene oxide assisted by electric field124. Reprinted with permission from ref.124. © 2019 IOP Publishing Ltd. d, Preparation of reduced graphene oxide by hydrothermal method62. Reprinted with permission from ref.62. © 2021 The Minerals, Metals & Materials Society. e, Reduction of graphene oxide by laser irradiation125. Reprinted with permission from ref.125. © 2012 Elsevier Ltd. f, Thermal decomposition of DMF as reducing agent to prepare reduced graphene oxide126. Reprinted with permission from ref.126. © 2010 The Royal Society of Chemistry. g, Ginger extract to reduce graphene oxide127. Reprinted with permission from ref.127. © 2021 Indian Academy of Sciences. Abbreviations: CVD, chemical vapor deposition; DMF, dimethylformamide. |
In order to overcome the disadvantage that graphene has no bandgap and is difficult to be applied in the field of electronic devices, researchers have obtained GO by strongly oxidizing graphene and have modified the graphene128. GO is generally produced by oxidation of graphene through a strong acid. The traditional methods for preparing graphite oxide are as follows: Brodie method, Staudenmaier method, and Hummers' method. Among the three methods, the Hummers' method has the highest efficiency and is therefore widely used. The Hummers' method refers to mixing graphite with nitric acid and sulfuric acid to cause an oxidation reaction on the graphite surface under heating conditions. Then, a strong oxidant such as potassium permanganate is introduced to further oxidize the graphite, and a large amount of water is added to obtain a GO dispersion. Finally, filtering, washing, drying, and other treatments are performed to obtain GO. Although the Hummers' method can produce a large amount of GO, the surface of the prepared GO is prone to impurities and defects, and the preparation process will release toxic gases such as NO2 and N2O4.
Researchers have improved the Hummers' method and proposed a series of improved Hummers’ methods for preparing GO. Shi et al. mixed graphite and sulfuric acid and slowly added potassium permanganate as an oxidant. Then, the reaction was transferred to an oil bath and stirred vigorously; water and hydrogen peroxide were added in sequence, and the color of the solution changed from dark brown to yellow. Finally, the solution is filtered, HCl solution is added to remove metal ions, and GO is obtained by drying. This improved Hummers' method eliminates toxic gases, simplifies the purification procedures of waste liquid, and reduces the synthesis cost of GO129. Xing et al. used flake graphite as the carbon source, potassium permanganate and potassium ferrate as the oxidant, and boric acid as the stabilizer, which were dispersed into concentrated sulfuric acid and were stirred at 5 °C for 1.5 h. Then, 5 g of potassium ferrate was added and stirring was continued for 3 h at 35 °C for deep oxidation. Finally, hydrogen peroxide is added to treat the remaining oxidant, and GO is obtained by repeatedly washing with deionized water and hydrochloric acid. The process is shown in Fig. 7b123. By introducing an electric field, the oxidation efficiency of graphene can be improved. Guo et al. fixed a copper plate on the outer wall of the beaker as the working electrode, and the distance between the copper plates was 60 mm, as shown in Fig. 7c. Under electric field conditions,0.5 g of graphite and concentrated sulfuric acid were added into a 400-mL square beaker; 1.5 g of KMnO4 was slowly added and stirred vigorously for 10 min. Then, the mixture was stirred in a 40-°C water bath for 30 min and transferred to a 70 °C water bath, and 25 mL of deionized water was added; stirring was continued for 5 min. Finally, H2O2 was added, and the color of the solution changed from dark brown to yellow; hydrochloric acid is added to remove metal ions, the mixture is purified using a dialysis membrane, and GO is obtained by ultrasonic treatment. The electric field force generated by the external electric field drives the ions to insert into the graphite layer, reducing the van der Waals force between the graphite layersand improving the graphite peeling efficiency. Moreover, the method of electric-field-driven ion insertion reduces the use of oxidants and reduces the damage of oxidants to the graphene structure124. By reducing GO, the material properties are made more stable, and reduced GO is obtained. It is a feasible method to obtain reduced GO through thermal reduction. Saleem et al. directly put GO into a muffle furnace and annealed it at 500 °C for 2 h until the temperature dropped to 50 °C, and GO was completely converted into reduced GO130.
The reduction of GO can also be achieved through the hydrothermal method, as shown in Fig. 7d. Zhang et al. used a GO solution mixed with nickel nitrate and cobalt nitrate, and the hydrothermal conditions were set to 140 °C/9 h. After cooling, the precipitate was cooled, dried, and calcined at 300 °C for 2 h to obtain the NiCo2O4/rGO composite material62. In addition to direct heating reduction, the reduction of graphene can also be achieved by using laser irradiation to generate local high temperatures and decompose oxygen-containing functional groups on GO, as shown in Fig. 7e125. Reducing agents produced by thermal decomposition of industrial solvents can also be used to produce reduced GO, as shown in Fig. 7f. The thermal decomposition of dimethylformamide (DMF) to reduce GO has the advantages of simple steps and low cost, which is conducive to large-scale production of reduced GO126. Extracting natural reducing agents from vegetables for the reduction of GO is a novel and green method. Swain et al. cut ginger into pieces, dried it, ground it into powder, refluxed it, and filtered it to obtain ginger extract. The ginger extract was added to the GO dispersion, and the mixture was refluxed at 90 °C for 4 h. The color of the solution changed from brown to black. The process is shown in Fig. 7g. The attack of the hydroxyl group of gingerol on the epoxy group of GO may be an important reason for the reduction of GO127.
MXene is a new 2D material, which refers to the general name of 2D transition metal carbides, carbonitrides, and nitride materials131-133. The synthesis methods of MXene mainly include hydrofluoric (HF) acid etching, alkali etching, fluoride salt etching, molten fluoride salt etching, electrochemical etching, and molten-salt assisted electrochemical etching. In 2011, Gogotsi et al. used Ti3AlC2 as a raw material for the first time and prepared MXene with HF acid etching. The typical HF acid etching method to prepare MXene is shown in Fig. 8a134. Zhang et al. used NaOH as an etchant to support the hydrothermal method and prepared MXene with a purity of 92 wt%. The preparation process is shown in Fig. 8b. Under the conditions of high temperature (270 °C) and high concentration of NaOH (27.5 M) solution, it is beneficial to the dissolution of Al. During the alkaline etching process, Al is oxidized into hydroxide and then dissolved in the alkali135. In addition, KOH136 and tetramethylammonium hydroxide137 can also be used as alkali etchants, as shown in Fig. 8c, d.
Fig. 8. MXene. a, HF acid etching method134. b, NaOH-alkali-etching-assisted hydrothermal method135. Reprinted with permission from ref.135. © 2018 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim. c, KOH-alkali-etching-assisted hydrothermal method136. Reprinted with permission from ref.136. © 2017 American Chemical Society. d, TMAOH-alkali-etching-assisted hydrothermal method137. Reprinted with permission from ref.137. © 2016 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim. e, Fluoride salt etching138. Reprinted with permission from ref.138. © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. f, Molten fluorine chemical salt etching139. Reprinted with permission from ref.139. © 2019 American Chemical Society. g, Electrochemical etching140. Reprinted with permission from ref.140. © 2018 Wiley-VCH Verlag GmbH & Co. KGaA, Weinheim. h, Molten salt assisted electrochemical etching141. Reprinted with permission from ref.141. © 2021 Wiley-VCH GmbH. Abbreviation: HF, hydrofluoric; TMAOH, Tetramethylammonium hydroxide. |
In addition, researchers have proposed a new idea of mild preparation of MXene through fluoride salt etching. Gogotsi et al. discussed the effects on MXene preparation when the molar ratio of LiF to MXene was 5 : 1 and 7.5 : 1, as shown in Fig. 8e. The results show that when excess LiF is introduced, it is beneficial for the etching of Al during the preparation of MXene138. MXene materials can be prepared by etching the MAX phase with transition metal halides in the molten state, as shown in Fig. 8f. Taking Ti3AlC2 MAX as an example, the new MAX phase Ti3C2Cl2 is synthesized through the substitution reaction between the Zn element in the molten ZnCl2 and the Al element in the MAX phase precursor (Ti3AlC2)139. Although molten fluoride salt etching is a safe etching method, it also has the disadvantages of high etching temperature and low yield rate142. Feng et al. achieved fluorine-free preparation of MXene through electrochemical etching method, as shown in Fig. 8g. Two pieces of Ti3AlC2 were selected as the anode and cathode of the two-electrode system. During the electrochemical etching process, only the Ti3AlC2 of the anode reacts. In binary aqueous electrolytes, chloride ions rapidly etch the anode Al and destroy the Ti-Al bond. Electrochemical etching achieves high-yield preparation of MXene, and the etching effect is better than that of the traditional etching method140. Wang et al. combined molten salt etching with electrochemical etching to prepare MXene. Ti3AlC2 and nickel were selected as the anode and cathode, respectively. By maintaining the electrolysis voltage at 2.0 V in the LiCl-KCl salt, due to the combined effects of electrochemistry and thermochemistry, Ti3C2Cl2 was obtained as the final product, as shown in Fig. 8h. The process of preparing MXene by molten-salt-assisted electrochemical etching does not produce acidic liquid waste and is a green and sustainable synthesis method141.
APPLICATION OF CARBON NANOMATERIALS IN MEMRISTORS
Before 1971, the three basic circuit components were resistors, capacitors, and inductors. In 1971, Leon O. Chua theoretically proposed the fourth basic component, the memristor, based on the relationship between electric charge and magnetic flux. A memristor has structural similarities with biological systems, in which the top electrode, functional layer, and lower electrode of the memristor correspond to the axon, synaptic cleft, and dendrites of biological synapse, respectively143,144. Memristors can be divided into mutation memristors and slowly changing memristors. Mutation memristors are also known as digital memristors. These devices often have two distinct high and low conductivity states, which can be reversibly transformed under external signal stimulation. Slowly changing memristor devices are also called analog memristors. The conductivity of the device will change when stimulated by external signals145.
Electrode
Nowadays, researchers have explored a variety of materials as electrodes for memristors, including metal elements (Ag, Au, Ti, Pt, Al, Ru, W, Cu), metal nitrides (TiN), metal oxides (ITO, IGZO), doped silicon, etc145,146. In addition, carbon nanomaterials, due to their nanoscale size, are suitable as electrodes for devices that do not require the use of traditional photolithography techniques147. Using carbon nanomaterials as crossbars or edge electrodes can achieve ultra-high-density RRAM with excellent switching performance, high thermal and chemical stability, good mechanical properties, and transferability can be achieved, compensating for the shortcomings of metal electrodes148,149.
Memristor functional layer material
The resistance transition characteristics of memristors are closely related to the type of functional layer materials. The resistance transition mechanisms of different types of functional materials are different150. From 0D materials such as carbon nanodots and GQDs (inducing directional growth of conductive filaments), to 1D materials such as SWCNTs and MWCNTs and carbon fibers (ideal electron transport properties), and then to graphene, Mxene, and other 2D materials (excellent mechanical strength and chemical stability), carbon materials of different dimensions construct memristive devices with different functional characteristics and promote the development of RRAM and neuromorphic devices.
CARBON-BASED MEMRISTORS FOR RESISTIVE RANDOM ACCESS MEMORY AND NEUROMORPHIC APPLICATIONS
Based on carbon nanomaterial memristors, researchers have developed a series of RRAM and neuromorphic applications. The working principle of RRAM is to achieve reversible control of the resistance state of the device through changes in external voltage and illumination151-154. For neuromorphic applications, the aim is to replicate brain functions with low power consumption, high parallelism, and high throughput155,156. This section will present the structural characteristics and performance indicators of RRAM and neuromorphic applications based on carbon nanomaterial memristors with different material dimensions and analyze the working mechanism.
Resistive random access memory
Pure carbon dots are not suitable for use as memory devices because pure carbon dots tend to aggregate and have poor dispersion. Therefore, for the use of pure carbon dot materials, the method of compounding carbon dot materials with other materials is most commonly used157. Tang et al. embedded CQDs into PMMA to prepare an RRAM device with a Ag/PMMA:CQDs/fluorine-doped tin oxide (FTO) structure, as shown in Fig. 9a. Citric acid and ethylenediamine were dissolved in deionized water, and a black-yellow CQD solution was prepared by hydrothermal method. By comparing the content of CQDs in PMMA from scratch, it can be seen that the stability of the device has been improved. COMSOL multiphysics software was used to simulate the changes in the interlayer electric field model, and it was found that the places where CQDs appeared showed obvious local electric field enhancement, as shown in Fig. 9b. The device has a long on/off retention time of 10,000 s, as shown in Fig. 9c. The electric field near the quantum dots induces the growth position of conductive filaments (CFs) along the direction of the CQDs, which can accurately affect the formation and breakage of conductive filaments, thereby changing the growth path of the conductive filaments and affecting the entire process of switching the RRAM device between the high-resistance state (HRS) and low-resistance state (LRS)158,159. Wang et al. spin-coated an active layer mixed with egg white and CQDs on the top of the ITO electrode and thermally evaporated the Al electrode on the top of the active layer to form an RRAM device with an Al/egg white:CQDs/ITO structure, as shown in Fig. 9d, e. Egg white is used as the active layer, and CQDs serve as charge-trapping sites. Compared to the RRAM that only uses egg white as the active layer, the addition of CQDs increases the on-off ratio to 1.19 × 104, a 10-fold increase. When a positive voltage is applied, carriers fill the trap. When the trap is completely filled and the threshold voltage of the RRAM is reached, the excess carriers form a conductive path, and the RRAM exhibits a LRS. A large number of carriers form an internal electric field. When a reverse voltage is applied, the charges in the trap are difficult to release, and the conductive path is not interrupted. Therefore, the Al/egg white:CQDs/ITO device exhibits the characteristics of once writing and reading multiple times. In addition, the device still works when it is bent from a planar state to a diameter of 15 mm, as shown in Fig. 9f160.
Fig. 9. 0D carbon-based storage devices. a, RRAM device with Ag/PMMA&CQDs/FTO structure. b, Interlayer electric field distribution simulated by COMSOL multiphysics. c, Ag/PMMA&CQDs/FTO on/off state data retentio158. Reprinted with permission from ref.158. © 2021 AIP Publishing. d, Al/egg white: RRAM device with CQD/ITO structure. e, Al/egg white:CQD/ITO structure RRAM device processing process. f, I–V characteristic curves of the device under different bending conditions160. Reprinted with permission from ref.160. © 2022 by the authors. g, Ag/GQD: RRAM device structure with PVP/Ag structure. h, Ag/GQD:PVP/Ag structure RRAM device (visible silver electrode and transparent active layer). i, The device was bent 1000 times in diameter at 8 mm for stability161. Reprinted with permission from ref.161. © 2015 Elsevier B.V. j, Ag/ZHO/GOQDs/ZHO/Pt structure RRAM device. k, I–V curve of RRAM device with Ag/ZHO/GOQDs/ZHO/Pt structure63. Reprinted with permission from ref.63. © 2017 The Royal Society of Chemistry. Abbreviations: 0D, zedo-dimensional; CQD, carbon quantum dot; GOQD, graphene oxide quantum dot; PMMA, polymethyl methacrylate; PVP, polyvinyl pyrrolidone; RRAM, resistive random access memory; ZHO, Zr0.5Hf0.5O2. |
In the field of RRAM, GQDs are usually used in combination with other materials162. Bae et al. used electrodynamic fluid technology to fabricate a composite material of GQDs and polyvinyl pyrrolidone (PVP) as a resistive switching layer. The top and bottom electrode materials were both Ag, forming a flexible RRAM with a Ag/GQD:PVP/Ag structure, as shown in Fig. 9g. The introduction of PVP is conducive to the electrodynamic fluid technology deposition of GQDs, and the mixture of graphene and PVP obtains semiconductor properties and achieves a fully organic active layer. The durability of the device is tested 1000 times with a diameter of 8 mm, as shown in Fig. 9i. The resistance of the high- and low-state resistors of the device changes slightly, which proves the stability of the device161. Yan et al. fabricated an RRAM device with a structure of Ag/Zr0.5Hf0.5O2 (ZHO)/GO quantum dots (GOQDs)/ZHO/Pt by inserting GOQDs into a ZHO film, as shown in Fig. 9j. After 106 switching cycles of the RRAM device, the high- and low-resistance ratios did not change significantly. The local electric field generated by GO quantum dots can induce the formation of Ag conductive filaments, thereby improving the stability of RRAM device. At the same time, the device also exhibits bipolar characteristics and a lower threshold voltage, as shown in Fig. 9k63.
1D CNTs have ideal electron transport properties, scalability, and excellent electrical properties at low voltages and are considered to be ideal materials for the fabrication of electronic devices163,164. Esqueda et al. used floating self-assembly technology to successfully fabricate highly aligned CNT RRAM at the wafer level, as shown in Fig. 10a. Test results show that the addition of CNTs improves the uniformity and non-volatile conductivity of the device. CNT RRAM and complementary metal oxide semiconductor (CMOS) circuits can be integrated three-dimensionally, as shown in Fig. 10b. The performance of the device has been tested using 20 enhancement pulses and 20 suppression pulses, as shown in Fig. 10c, and the curves show that the device has good repeatability165. The surface of MWCNTs contains a large number of surface groups that can increase defects and chemical reactivity. GQDs have excellent resistance adjustment capabilities and UV sensitivity. By using synthetic composites of MWCNT and GQD as resistive switching layers, the advantages of both materials can be combined to achieve high-performance RRAM devices. Wang et al. used ultrasonic oscillation to mix GQDs and MWCNTs in deionized water, spin-coated the mixed material on the ITO substrate, and used vacuum evaporation onto deposit the top Al electrode. The structure of the RRAM device is shown in Fig. 10d. By controlling the mass ratio of MWCNTs and GQDs, the switching ratio of RRAM devices can be controlled. When the mass ratio of MWCNTs and GQDs is 1:0.5, the switching ratio of the RRAM device is 3.3 × 103, and the retention time of the device state exceeds 104 s, as shown in Fig. 10e. At the same time, the device exhibits excellent resistance uniformity, as shown in Fig. 10f166. In addition, electrodes for RRAM devices can also be fabricated from 1D CNTs (carbon nanotubes). Shim et al. prepared an RRAM with a CNT/AlOx/CNT structure, with a programming current as low as 1 nA and an on/off ratio as high as 5 × 105. By studying the effects of CNT electrodes (semiconductor or metal) with different properties on RRAM devices, it was found that the resistance under high resistance is mainly determined by the resistance of the AlOx film in the off state, whereas low resistance is more dependent on CNT electrodes with greater resistance. Metallic CNT electrodes will cause breakdown of the device, whereas CNT electrodes with too high resistance will sacrifice power consumption and switching ratio too much. Therefore, selecting the appropriate type of CNT electrodes is the key to preparing CNT-electrode-based RRAM. In addition, the rapid development of CNTs in large-scale preparation, transfer, and device integration has opened the way for the use of CNT electrodes to prepare ultra-high density RRAM arrays167,168.
Fig. 10. 1D carbon-based storage devices. a, Wafer-level highly aligned carbon nanotube RRAM. b, Three-dimensional integration of carbon nanotube RRAM and CMOS circuits. c, Positive- and negative-pulse testing of the device165 Reprinted with permission from ref.165. © 2018 American Chemical Society. d, RRAM based on Al/MWCNTs:GQDs/ITO structure. e, The time curve for which the device is in a state of hold. f, Cumulative probability distribution of RRAM resistance based on Al/MWCNTs:GQDs/ITO166. Reprinted with permission from166. © 2021 by the authors. g, Textile carbon fiber RRAM. h, Resistance stability test of textile carbon fiber RRAM64. Reprinted with permission from ref.64. © 2021 The Royal Society of Chemistry. i, RRAM based on TiO2@cf composite fiber. j, I–V characteristics of the device169. Reprinted with permission from ref.169. © 2019 Elsevier Ltd and Techna Group S.r.l. Abbreviations: 1D, one-dimensional; CMOS, complementary metal oxide semiconductor; GQD, graphene quantum dot; MWCNT, multiwallled carbon nanotube; RRAM, resistive random access memory. |
Carbon fibers can solve the problem that traditional insulating textile fibers cannot be directly used as conductive substrates. Jiang et al. used magnetron sputtering to deposit Ba0.6Sr0.4TiO3 material on the surface of carbon fibers to form composite carbon fibers. Flexible RRAM was obtained by assembling it on a PI film through the cross method. The device has a high switching ratio of 106, and its performance remains stable after 1000 switching cycles170. Lee et al. successfully developed a textile-type RRAM, which is made of aluminum-coated yarn (solution method), carbon fiber yarn, and cotton thread, woven through a knitting machine. The natural alumina formed on the surface of the aluminized yarn serves as a resistive layer, and aluminum and carbon fibers serve as electrodes at both ends, forming an RRAM structure, as shown in Fig. 10g. Textile RRAM has high and low resistance characteristics, a switching ratio of 10, and a set/reset voltage of +3/−3 V. Within 40 switching cycles, the high and low resistance values of the textile RRAM remain stable. Within 40 cycles, the HRS and LRS values of the device remain stable, as shown in Fig. 10h64. Yue et al. assembled TiO2-nanorod-coated carbon fibers (prepared by hydrothermal method) on the top of the PI film to form a flexible RRAM with interdigitated fibers, as shown in Fig. 10i. Unlike the ohmic properties of pure carbon fibers, TiO2-nanorod-coated carbon fibers exhibit good I-V characteristics, as shown in Fig. 10j. The TiO2-nanorod-coated carbon fiber exhibits bidirectional threshold switching behavior with set voltage/recovery voltage of +5/−5 V. The device showed good stability and remained stable after 1500 switching cycles. The work functions of carbon fibers and TiO2 nanorods are different, and a Schottky barrier is formed between them. Under the condition of applying a positive voltage, oxygen vacancies accumulate at the barrier interface, the width of the depletion region becomes smaller, and the device transitions to an LRS. Carbon fibers give the device better flexibility, and TiO2 nanorods provide more oxygen vacancies for the device state transitions169.
Graphene electrodes have high contact resistance and weak van der Waals forces, which can reduce operating current compared to some metal electrodes171,172. As an electrode material for RRAM, graphene can significantly reduce the operating current of RRAM devices173,174. Yang et al. used radio frequency magnetron sputtering to prepare RRAM with a structure of Ag/ZnO/graphene. Compared to devices with an Ag/ZnO/Au structure, the HRS and LRS fluctuations of RRAM using graphene electrodes after 3 × 103 cycles are small175 after 3 × 103 cycles. Graphene has a low longitudinal electron mobility. The relationship between graphene and the memristor resistive material will produce a series-resistance-like effect, which can improve the self-current compliance of RRAM devices176,177. At the same time, RRAM devices using graphene as electrode materials have good biocompatibility and have broad applicability in future implantable biomedical devices178. Graphene can also be used as a transparent electrode, which can bring significant photovoltaic effects to devices. 0.7 wt.% Nb-doped SrTiO3 was used as the substrate, and a 5-mm-wide polytetrafluorethylene (PTFE) film was used as the insulating layer. The top of the PTFE film is graphene and Ag top electrodes, and the Al wire bonded on the Nb-doped SrTiO3 substrate serves as the bottom electrode, as shown in Fig. 11a. During the process of turning the light source on and off, the photovoltage of the device was 38 mV and 7 mV, respectively. The stable photovoltage measurements of the device are attributed to the dominant effect of photothermal electrons in the graphene electrode on the photocurrent. At the same time, the high carriers in graphene can greatly improve quantum efficiency and generate larger photocurrents than conventional metal electrodes. In addition, the device has memristive properties even when only voltage is applied. At a positive voltage of 1.5 V, the device enters the LRS, whereas at a negative voltage of −1.5 V, the device enters the HRS, as shown in Fig. 11b179. By inserting graphene into the resistive layer material and designing nanopores in the graphene, the regulation of ion transport can be achieved180. Lu et al. inserted graphene between Ta and Ta2O5 to control the movement of oxygen vacancies. As the size of the nanopores on graphene increases, the device current value increases, indicating that the nanopores on graphene can control the movement of oxygen vacancies, thereby controlling the formation position of conductive filaments, which is beneficial to improving the stability of RRAM devices, as shown in Fig. 11c. As the pore size of graphene increases, the conductive filament gradually becomes thicker and the current value increases, as shown in Fig. 11d181.
Fig. 11. 2D carbon-based storage devices. a, Interface RRAM based on graphene transparent electrode. b, The I–V characteristic curve of the device179. Reprinted with permission from ref.179. © 2018 Elsevier B.V. c, Schematic diagram of the memristor structure inserted into the graphene layer. d, Effect of graphene with different pore sizes on memristive properties181. Reprinted with permission from ref.181. © 2016 American Chemical Society. e, RRAM device with Al/GO/ITO structure. f, Stability testing of Al/GO/ITO devices182. Reprinted with permission from ref.182. © 2018 Elsevier B.V. g, RRAM device with Ag/GO/ITO structure183. Reprinted with permission from ref.183. © 2019 Elsevier Ltd. h, Energy band changes of RRAM with Au/Ni–Co layered double hydroxide/GO/ITO structure184. Reprinted with permission from ref.184. © 2022 Elsevier B.V. i, Preparation process of RRAM based on Ag/PVA-GO/FTO structure185. Reprinted with permission from ref.185. © 2020 The Authors. j, Based on Al/Ti3C2TX:PI/ITO structure179. Reprinted with permission from ref.186. © 2022 Elsevier Ltd. Abbreviations: 2D, two-dimensional; GO, graphene oxide; RRAM, resistive random access memory. |
GO has good dielectric properties and is widely used in the field of RRAM devices187. Zhang et al. used electrophoresis technology to deposit GO on the ITO substrate and evaporated the Al electrode by electron beam at the top to form an RRAM device based on Al/GO/ITO. The device has a low switching voltage (1.7 V), a switching ratio of more than 10, and a hold time of more than 100 s188. In addition, Jeganathan et al. also used hydrogen plasma to treat GO to reduce the number of oxygen groups on the lattice, promoting the formation of the interfacial layer between GO and the electrode (Al) and increasing the on-off ratio of the RRAM device to 1000. The device structure and stability tests are shown in Fig. 11e and f182. GO is also dependent on the wavelength of light. Long-wavelength light or short-wavelength light will trigger reversible and irreversible photoconductive effects of GO. Kemp et al. found that the photoconductive effect of the GO-based RRAM device is completely reversible under low-intensity-light irradiation with a longer wavelength. The device structure is shown in Fig. 11g. The GO has a bandgap of 4.2 to 5.5 eV, which is consistent with the optical response of the device. UV light irradiates devices to produce a 10-fold higher light respontaneity than red light. The GO prepared by light-induced heating reduction is more conductive, and the wavelength and intensity of light will affect the performance of the device. However, under shorter-wavelength-light irradiation, the photoconductive effect of the RRAM device is only partially reversible, which is attributed to the photoinduced reduction of GO. Therefore, the performance of the RRAM device can be modulated by modulating the wavelength and intensity of the irradiation light183.
However, the ratio of high resistance to low resistance of pure GO RRAM reported in the existing literature is less than 104, which limits the application of GO RRAM in large-capacity and low-power storage. Ren et al. improved the high-to-low resistance ratio of RRAM by introducing all-inorganic perovskite quantum dots (CsPbBr3) into 2D GO. At the same time, the introduction of CsPbBr3 quantum dots gives RRAM devices optoelectronic properties. The high-to-low resistance ratio of CsPbBr3 quantum dots/GO RRAM is 260 times that of GO RRAM. The migration of Br in CsPbBr3 promotes the formation of conductive filaments189. Sun et al. employed an externally applied electric field to adjust the work function of GO, affecting the barrier height of GO and Ni-Co layered double hydroxide, to control charge transport in the resistive layer of the memristor. The aqueous solution of GO was coated on the ITO surface by spin coating, and the ethanol solution of Ni-Co layered double hydroxide (hydrothermal method) was coated on the GO surface by drop coating. Finally, an Al/Au electrode with a length of 800 μm and a width of 10 μm was prepared with the adoption of vacuum evaporation and deposition processes. Applying positive and negative voltages to the ITO gate, the energy band bending between Ni-Co layered double hydroxide and GO changes, as shown in Fig. 11h184. Pham et al. explored the effects of different GO concentrations (0.5 wt.%, 1.0 wt.%) on the RRAM performance of PVA/GO composite films. GO and PVA were configured according to the volume of 1 : 1 to form a mixed solution. A PVA/GO composite film was deposited on the FTO substrate by spin coating, and a magnetron-sputtered Ag electrode was used as the top electrode, as shown in Fig. 11i. When the set/reset voltage is 0.34 V/−0.28 V, the composite thin-film RRAM incorporating 0.5 wt.% GO has a high on-off ratio of 104. The coupling between the oxygen-containing groups of GO and the hydroxyl groups of PVA controls the carrier transport pathway through the structure. The appropriate ratio of PVA to GO can improve the switching ratio of RRAM devices185. In addition, reduced GO (rGO) has emerged as an outstanding RRAM candidate due to its excellent electrical conductivity and chemical stability190. Zhang et al. used a hydrothermal method to synthesize chitosan and rGO composites as the resistive switching layer of RRAM. Chitosan and rGO exhibit an interpenetrating network structure composed of hydrogen bonds and covalent bonds. Under the action of an external electric field, metal ions can migrate rapidly and stably in the interpenetrating network. Therefore, the RRAM device based on chitosan and rGO composite has clearly stable HRS and LRS states, and the state retention time is higher than 104 s191.
MXene has a layered structure and high packing density and has received widespread attention in various fields192. The surface of MXene contains rich functional groups such as -O, -F, or -OH, which can be used as a narrow-bandgap semiconductor193,194. Tong et al. controlled the growth of conductive filaments by spin-coating MXene with a thickness of 50 nm on the surface of SiO2 materials and greatly reduced the operating voltage of SiO2 RRAM devices (set/recovery voltage is 0.2 V/ to 0.2 V). Mxene bonds with amorphous SiO2 to form a compound, which makes conductive filaments more likely to grow along the nanostructure of MXene, reducing the operating voltage and power consumption of SiO2 RRAM devices195-197. In addition, Liu et al. combined the energy storage and memristive properties of MXene and introduced 0.5 wt.% Ti3C2TX nanosheets into the PI to achieve a RRAM device processing with resistive memory properties and energy storage performance, as shown in Fig. 11j. The Al/Ti3C2TX:PI/ITO device has non-volatile rewritable storage performance, with a set voltage of approximately 2 V and a switching ratio close to 100. At the same time, the discharge energy density of Ti3C2TX-doped PI reaches 0.4 J cm−3 at 150 kV mm−1, which is 30% higher than that of pure PI (0.3 J cm−3), showing excellent energy-storage characteristics. The excellent energy and information storage mechanism of Al/Ti3C2TX:PI/ITO devices can be briefly described as follows: PI chains embedded in the MXene layer to form a strong interaction interface and are the main reasons for the enhanced performance of energy storage of RRAM. The hydroxyl groups and aluminum oxide compounds released by the cleavage of the MXene surface (the hydroxyl groups are cleaved to release active O atoms and form a wide-bandgap aluminum oxide compound with the residual Al) change the energy band structure of MXene, achieving a balance between dielectric strength and appropriate resistance balance. Al/Ti3C2TX:PI/ITO devices with excellent energy and information storage will provide a reference for subsequent research on self-powered high-performance RRAM devices186.
There are many types of RRAM devices based on carbon-based materials, and due to the limitation of article length, a table of the performance of RRAM devices are compiled, as shown in Table 1. As can be seen from Table 1, the RRAM based on carbon-based materials has the advantages of low threshold voltage, high switching ratio, long data storage time, and high stability.
Table 1. RRAM performance. |
| Device | Threshold voltage (V) | Threshold voltage power (W) | ON/OFF ratio | Retention time (sec) | Endurance (cycles) | Refs. |
|---|---|---|---|---|---|---|
| Cu/MXene/Cu | 0.68/−0.61 | - | - | - | - | 155 |
| Ag/GQDs/PVP/Ag | 1.8/−1.8 | - | 14 | 1.8 × 107 | 5 × 102 | 161 |
| CNT/AlOx/CNT | 0.5/−0.5 | - | 5 × 105 | 1.25 × 104 | 50 | 167 |
| Ag/Al2O3/GQDs/Al2O3/ITO | 1.2/−1.2 | - | - | - | - | 198 |
| Pd/CQDs/Ga2O3/Pt | 1.7/−0.06 | 1.2 × 10−7/4.15 × 10−5 | 102 | 4.5 × 104 | - | 199 |
| Ag/Zr0.5Hf0.5O2:GQDs/Ag | 0.6/−0.6 | - | <10 | - | - | 200 |
| Ag/HfO2/GQDs/Pt | 0.15/−0.13 | - | 106 | 1 × 104 | - | 201 |
| ITO/MQD-PVP/Au | 1.6/−3 | - | 102 | 1.2 × 104 | 2 × 102 | 202 |
| Ag/N-GOODs/Pt | 0.4/−0.2 | - | 107 | 2.5 × 103 | 1.2 × 104 | 203 |
| Ag/N-GOQDs/Pt | 0.14/- | - | 106 | - | 30 | 204 |
| Pt/GQDs-FeOX/Pt | 1/−0.7 | - | 50 | - | 2 × 103 | 205 |
| ITO/5CB−MWCNT/ITO | 4/- | - | - | 6.7 × 106 | - | 206 |
| N:BST@Cf/BST@Cf/PI | 1.5/−1.5 | - | 106 | 7.87 × 102 | 1 × 103 | 207 |
| Al/graphene/parylene/W | 2.5/−3 | - | - | 104 | 1.2 × 102 | 208 |
| Ag/GO/Py-salt/GO/ITO | 5/−5 | - | - | 6 × 103 | - | 209 |
| ITO/graphene/ZnO/ITO | 4/- | - | 20 | - | 102 | 210 |
| hrGO/lrGO/hrGO | −/−13.2 | - | 102 | - | - | 211 |
| Ni/GO/Au | 1.5/0.5 | - | 102 | 104 | 3 × 102 | 212 |
| Cu/Ti3C2/BFO/Pt | 1/-0.6 | - | >103 | - | - | 213 |
| DTM MXene/GO/DTM MXene | 2/−2 | - | 102 | 105 | 5 × 103 | 214 |
| Pt/MXene/Pt | 5.03/−5.12 | - | 5.62 × 103 | 104 | - | 215 |
Abbreviations: BFO, BaFe12O19; CNT, carbon nanotube; GO, graphene oxide; CQD, carbon quantum dot; DTM, double transition metal; GOOD, graphene oxide quantum dots; GQD, graphene quantum dot; GOQD, graphene oxide quantum dot; MWCNT, multiwallled carbon nanotube; RRAM, resistive random access memory. |
Neuromorphic applications
CQDs are not as active as Ag quantum dots. The randomness of CQD diffusion is lower, and the device performance is more stable. Zhao et al. used magnetron sputtering and spin coating methods to produce neuromorphic devices based on CQDs. The device has a Au/CQDs/ITO structure. The working mechanism of the Au/CQDs/ITO device is different from that of the traditional conductive filament model, as shown in Fig. 12a. Under the action of an external electric field, the hybridization state of carbon in the resistive switching layer of CQDs changes. The sp2 hybridization of carbon mainly includes C=O, C=C bonds, and the sp3 hybridization of carbon mainly includes C-O bonds. Under negative voltage, the negatively charged functional groups are separated from CQDs, and the sp3 hybridization of carbon is changed to sp2 hybridization, which reduces the trap spacing, increases the conductance value, and displays the LRS. Under positive voltage, positively charged functional groups aggregate with CQDs, and the sp2 hybridization of carbon will be transformed into sp3 hybridization, which increases the trap spacing and appears as an HRS state. Under an external electric field, the memristive effect caused by the change of carbon hybridization state has lower randomness and good stability during the period. At the same time, by extracting the weight value of the device, the neural network is trained to achieve a recognition rate of 96.7% on the Mixed National Institute of Standards and Technology (MNIST) database, as shown in Fig. 12b216. Yan et al. used magnetron sputtering to deposit a 10-nm Ga2O3 film on a Pt substrate and drop-coated CQDs on the surface of the Ga2O3 film to form a Pd/CQDs:Ga2O3 film/Pt structure neuromorphic device. Carbon quanta form conductive filaments to participate in the resistance switching of the device, as shown in Fig. 12c. At the same time, the device can realize the functions of short-term plasticity), long-term plasticity), spike-timing-dependent plasticity (STDP), and paired-pulse facilitation (PPF) during performance testing. The Pavlovian conditioned reflex was successfully simulated through different voltage values, as shown in Fig. 12d199.
Fig. 12. 0D carbon-based neuromorphic devices. a, Resistive switching mechanism of Au/CQDs/ITO neuromorphic device based on carbon hybridization. b, Accuracy training curve of neural network on MNIST data set216. Reprinted with permission from ref.216. © 2023 The Royal Society of Chemistry. c, EELS mapping of device. d, Pavlovian conditioning199. Reprinted with permission from ref.199. © 2020 The Royal Society of Chemistry. e, Typical solution processing and drop casting of N-GOQDs. f, Ag/N-GOQDs/Pt/Ti/SiO2/p-Si memristor working mechanism203. Reprinted with permission from ref.203. © 2019 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. g, GQDs:FeOx durability of neuromorphic devices at different voltages217. Reprinted with permission from ref.217. © 2017 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. Abbreviations: 0D, zero-dimentional; CQD, carbon quantum dot; EELS, electron energy loss spectroscopy; GQD, graphene quantum dot; GOQD, graphene oxide quantum dot; MNIST, Mixed National Institute of Standards and Technology database. |
For the application of GOQDs in the field of neuromorphic devices, most of them are N-doped GOQDs. After doping GOQDs with nitrogen, the material itself is rich in functional groups that control the electron-hole coincidence process. In 2019, Choi et al. used an improved Hummers' method to synthesize GO, introduced NH4OH as a nitrogen source, and used a hydrothermal method to achieve N doping of GO. The N-GOQDs colloidal solution was drop-coated on the top of the Pt electrode using the drop-coating method, as shown in Fig. 12e. The device structure completed by drop coating is a Ag/N-GOQDs/Pt/Ti/SiO2/p-Si structure. The introduction of nitrogen doping increases the electron density near the Fermi level of the material, which is beneficial to the transition of electrons. The device exhibits good high-resistance and low-resistance switching characteristics. In the initial state, there is no Ag in GOQDs. Under weak positive voltage stimulation, a small amount of Ag+ ions are implanted into GOQDs to form unstable conductive filaments. After the weak positive-voltage stimulus is removed, the conductive filaments break, enabling short-term memory. Under strong positive-voltage stimulation, a large number of Ag+ ions are injected into GOQDs to form stable conductive filaments and achieve long-term memory. The conduction mechanism is shown in Fig. 12f203. In 2021, Choi et al. investigated the effect of UV-light irradiation on conductive filaments. Under the irradiation condition of UV light with a wavelength of 365 nm, when different voltages are applied, the device switches between HRS and LRS. Under the condition of no UV-light irradiation, the resistance of the device remains unchanged. The switching of the device resistance is related to the formation/breakage of conductive silver wires. Therefore, it can be speculated that UV irradiation can be used to promote the formation of silver filaments in devices218. Highly repeatable simulation of resistive states in neuromorphic devices is critical for realizing high-density artificial neural networks219. Research shows that the combination of GO and metal oxides can significantly improve the resistance state stability of neuromorphic devices. Taking advantage of the fact that GQDs are rich in oxygen ions, GQDs can be used as oxygen storage. By spin-coating GQDs on FeOx films and using GQDs to enhance the localization of conductive fibers, the simulated resistance state stability of neuromorphic devices can be significantly improved. Under the 2000-cycle pulse test, the resistance value of the neuromorphic device with the addition of GQDs showed excellent stability, as shown in Fig. 12g217.
Aligned CNTs have been proven to form n conductive states, which are related to the intensity of the external electric field and the degree of CNT deformation. CNTs and metal oxides are considered effective materials for the fabrication of neuromorphic devices. Il'ina et al. used enhanced plasma CVD to grow a CNT array on the Ti surface, choosing metal W as the upper electrode to form a neuromorphic device structure. The resistive switching characteristics of the CNT neuromorphic device were tested under atmospheric pressure and vacuum conditions, as shown in Fig. 13a. The atomic force microscopy (AFM) probe tip applies a force of 300 nN to produce CNT deformation, and the AFM probe tip provides a voltage of ±5 V to the device. In the initial state, CNTs have high resistance characteristics, and the AFM probe presses the CNT to deform. When a voltage of −5 V is applied, the CNT undergoes compressive deformation, and the initial tensile deformation of the CNT is redistributed and switches to a LRS. At the same time, the on-off ratio of CNTs measured in vacuum is higher than that in air environment. In a vacuum environment, there is no adsorption layer on the CNT surface, resulting in a reduction in the piezoelectric charge of the deformed CNTs220. Wan et al. fabricated a CNT-based neuromorphic device using photolithography process, aerosol deposition, and self-selected coating technology, as shown in Fig. 13b. The photolithography process was used to fabricate Au as the source and drain of the device, dispersed CNT as the channel layer, and patterned Ag as the lateral gate and top gate. The device can exhibit high on-off ratio and low leakage current characteristics at low operating voltage (±1 V), which is attributed to the effective tuning of the hole concentration of the CNT channel. By applying a voltage to the side gate of the device, the current value at both ends of the device channel can be tuned by six orders of magnitude, as shown in Fig. 13c221.
Fig. 13. 1D carbon-based neuromorphic devices. a, CNT neuromorphic device characteristic test curve and schematic diagram in vacuum and air environments220. Reprinted with permission from ref.220. © 2022 Elsevier B.V. b, CNT-based three-terminal neuromorphic device. c, Device channel current versus voltage curve device channel current versus voltage curve. d, Implementation of device logic operations221. Reprinted with permission from ref.221. © 2016 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. e, Photoresponsive transistor based on CsBi3I10/SWCNTs. f, Energy band diagram of CsBi3I10 and SWCNTs films222. Reprinted with permission from ref.222. © 2019 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. g, 4-cyano-4′-pentylbiphenyl liquid crystal (5CB): MWCNTs composite material nerve morphological device. h, Distribution of 5CB-MWCNT composite materials under different electric field conditions206. Reprinted with permission from ref.206. © 2020 American Chemical Society. Abbreviations: 1D, one-dimensional; CNT, carbon nanotube; MWCNT, multi-walled carbon nanotubes; SWCNT, single-walled carbon nanotubes. |
MWCNTs have been shown to enhance the dissociation and transport of photogenerated carriers and can significantly improve the photoresponse. Zhao et al. fabricated a CsBi3I10/SWCNTs photoresponsive transistor by spin-coating SWCNTs on the SiO2 surface, spin-coating the CsBi3I10 precursor solution on the nanotube surface, and finally completing the annealing process in a glove box, as shown in Fig. 13e. The addition of CNTs significantly improves photoresponsivity (6.0 × 104 A W−1) and external quantum efficiency (1.66 × 105%) of the transistor. The corresponding energy band diagrams of CsBi3I10 and SWCNTs are shown in Fig. 13f. Under photoexcitation, holes are transferred to the SWCNT film, whereas electrons are retained in CsBi3I10. The effective separation of photoexcited carriers ensures the device’s high performance. CsBi3I10/SWCNT transistors also have potential as synaptic applications, successfully simulating neuromorphic indicators such as PPF, excitatory postsynaptic currents, long-term memory, and short-term memory222. By controlling the arrangement of CNTs, the resistance value of neuromorphic devices can also be controlled. Lee et al. used 4-cyano-4′-pentylbiphenyl liquid crystal (5CB):MWCNT composite as the resistive switching layer of neuromorphic devices, as shown in Fig. 13g. The resistivity is changed by controlling the arrangement and aggregation state of CNTs in the composite material by applying an external electric field, as shown in Fig. 13h. Due to the interaction between π-π chemistry, the liquid crystal is anchored to the surface of the MWCNTs. Under the action of an electric field, the anchored 5CB molecules are oriented parallel to the electric field, which in turn drives the MWCNTs to be oriented parallel to the electric field. By controlling the voltage direction and intensity, the MWCNTs can be controlled. Finally, the adjustment of neural weights is achieved206. Flexible textile neuromorphic devices can also be processed by using CNT arrays to weave carbon fibers (bottom electrode), with Ag fibers (vertex electrodes) and MoS2/HfAlOx (resistive layer). The device has extremely low power consumption of 1.9 fJ, which is three orders of magnitude lower than the existing neuron power consumption223. In addition, CNTs are wrapped by polymers, for example, HiPCO/SWCNT224, polyethylene glycol (PEG)/CNT225, polydimethylsiloxane (PDMS)/CNT226, etc. As a resistive switching layer for neuromorphic devices, it also opens up new ideas for flexible neuromorphic devices.
The 2D material properties of graphene make its interlayers a good host for ions. Injecting ions into graphene can significantly improve the performance of neuromorphic devices227. Xiong et al. used an external current to drive Li+ ions from the reference electrode into graphene, as shown in Fig. 14a. Test results show that Li+ ion injection achieves effective regulation of graphene conductivity. Li+ ions-injected graphene neuromorphic devices have a conductance adjustment capability of 700%, which is significantly higher than the 150% change in biological synapses. At the same time, the device has good repeatability, as shown in Fig. 14b. Graphene's high out-of-plane resistivity and low out-of-plane thermal conductivity properties can enable the preparation of low-current and low-power neuromorphic devices when used as electrode materials in neuromorphic devices228. Lai et al. used graphene as the bottom electrode, AlOx as the resistive switching layer, and Al as the top electrode of a neuromorphic device, as shown in Fig. 14c. Compared to devices with Pt as the bottom electrode, devices with graphene as the bottom electrode exhibit ultra-low operating current and a resistance ratio of up to 1 × 106 between the HRS and the LRS due to its weak surface van der Waals interaction229. Ding et al. used a vacuum filtration method to prepare a vertically structured black phosphorus-graphene hybrid neuromorphic device composed of black phosphorus, GO quantum dots, and GO films. The introduction of graphene greatly improves the stability of black phosphorus in the air, and the quantum confinement effect of GO increases the electronic capacity of black phosphorus, thereby achieving low power consumption (62 pW) and high sensitivity (−20 mV). At the same time, neuromorphic functions such as PPF, STDP, and Pavlovian conditioning were successfully simulated230.
Fig. 14. 2D carbon-based deuromorphic devices. a, Device diagram for doping Li+ ions in graphene. b, Repeatability of neuromorphic devices doped with Li+ ions in graphene228. Reprinted with permission from ref.228. © 2018 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim. c, Comparison of neuromorphic devices with graphene as bottom electrode and human synaptic structure229. Reprinted with permission from ref.229. © 2018 American Chemical Society. d, Neuromorphic device structure based on Pt/MXene/ITO structure. e, Working mechanism of neuromorphic devices with Pt/MXene/ITO structure67. Reprinted with permission from ref.67. © 2023 Elsevier Ltd. f, Fabrication process of neuromorphic devices based on Al/Ag NPs@MXene-TiO2 nanosheets/ITO231. Reprinted with permission from ref.231. © 2021 American Chemical Society. Abbreviation: 2D, two-dimensional. |
The ultra-high conductivity of MXene limits its application in the field of neuromorphic devices232. By oxidizing MXene, the high conductivity of the device is modified and the device is given certain semiconductor properties, thereby improving the bipolar switching performance. Ning et al. used Ti3AlCx as a raw material and etched multilayer Ti3C2Tx using LiF and HCl solution methods. Ti3C2Tx was calcined in air at a temperature of 200 °C to obtain partially oxidized Ti3C2Tx. Finally, magnetron sputtering was used to deposit Pt electrodes on the MXene surface to form a neuromorphic device with a Pt/MXene/ITO structure, as shown in Fig. 14d. The oxidized MXene has higher oxygen vacancy content, which is more conducive to the formation of conductive filaments, as shown in Fig. 14e. The oxidized MXene neuromorphic device has long cycle life (>103 cycles) and retention time (>104 s), successfully simulating synaptic functions such as PPF and LTP67.
In addition, doping Ag into MXene can also optimize the performance of neuromorphic devices and eliminate the sudden change in current behavior. Doping Ag in MXene solves the problem that traditional Ag electrodes are dissolved due to electrochemical reactions under the action of an external electric field, causing the electrical performance of the device to deteriorate. Yan et al. used a hydrothermal method to prepare Ag-doped MXene as a resistive layer to form an Al/MXene:Ag/Pt neuromorphic device. The Ag NP doping method not only overcomes the drawbacks of poor electronic performance caused by the dissolution of traditional Ag electrodes but also solves the current mutation behavior of the device while reducing the power consumption of a single device (0.35 pj)233. Zhou et al. prepared an optoelectronic neuromorphic device based on Ag NPs@MXene-TiO2 nanosheets through a hydrothermal oxidation process, as shown in Fig. 14f. Test results show that RRAM has low off-current (<1 pA), high on-off ratio (∼106), and long retention time (>104 s). Under the action of UV irradiation and pulse voltage, it can reach a multi-value resistance state and has good repeatability, electrical properties, and stability231.
APPLICATIONS OF CARBON-BASED RESISTIVE RANDOM ACCESS MEMORY AND NEUROMORPHIC DEVICES
Based on carbon nanomaterial RRAM and neuromorphic devices, researchers have further expanded the functions and application scenarios of RRAM and neuromorphic devices by introducing photoelectric hybrid modulation or integration with other sensing devices, such as implementing logical operations, building neural networks, bionic vision, bionic touch, and multimodal perception systems234. Specific application scenarios of RRAM and neuromorphic devices based on carbon nanomaterials will be introduced.
Logic in computing and artificial neural networks
In the recent years, memory devices based on memristors have become an important part of the construction of parallel information-processing hardware because their conductance states represent logical values 0 and 1 and logical calculation results can be stored locally235-238. Han et al. used lead-free perovskite (CsBi3I10) mixed SWCNTs as channel materials and successfully developed a high performance multimodal photoelectric synaptic transistor. 0.2598 g of cesium iodide and 1.7691 g of bismuth iodide were added to a mixed solvent consisting of DMF and dimethyl sulfoxide (DMSO) with a volume ratio of 13 : 1, and then the mixture was stirred at 60 °C at 1000 rpm for 3 h, and the final color changed to reddish brown to obtain CsBi3I10 precursor. The aerosol jet printer deposited the SWCNT ink into the channel and then heated it at 70 °C for 2 min. The diluted CsBi3I10 solution was spin-coated on the SWCNT thin film transistor liquid crystal display (TFT) channel at 4000 rpm, and the film was annealed at 125 °C for 30 min. The specific processing process is shown in Fig. 15a239. The combination of CsBi3I10 and SWCNTs has more defects and significantly improves the photoresponse and electrical response current of the device. It is defined that 5 V represents the laser-on, which is considered “1”, and 0 V represents the laser-off, which is considered “0”. A gate voltage of 2 V is considered a “1”, and a gate voltage of 1.73 V is considered a “0”. When the drain current generated by the device is at a high level, it is called a binary “1”; otherwise, it is called a binary “0”. Fig. 15b shows the control effect of the optical signal and the electrical signal on the device. When the optical signal and the electrical signal are both “1”, the drain current is “0”. In all the other cases, the drain current is “0”, which is consistent with typical NOR logic239. Sun et al. prepared an Al/NiAl-LDHS/GO/SiO2/Si memristor to realize the D-type latch function. The device structure is shown in Fig. 15c. E1 serves as the clock signal of the D flip-flop, +20 V/−20 V represents the 1/0 clock signal. E2 serves as the input signal of the D flip-flop, and +25 V/+15 V represents the 1/0 input signal. A voltage of 0.5 V was applied between E3 and E4 to measure the resistance value between E3 and E4 as output. When the output resistance is greater than 104 Ω, the output is logic 1; otherwise it is logic 0. D represents the input signal, Qn represents the previous state of the D flip-flop, and Qn+1 represents the secondary state of the D flip-flop. When the input signal at D terminal is 5 V, D = 0, the output signals are all 0, as shown in Fig. 15e and f. When the gate voltage is 20 V, Qn+1 = D regardless of Qn state. When D = 1, Qn+1 = 1, as shown in Fig. 15g and h, meeting the definition of D flip-flop. When the gate voltage is −20 V, the output signal does not change with the input D signal. This is because when a negative voltage is applied to the gate, a conductive channel cannot be formed between the source and the drain. The large channel resistance between the source and drain makes the bias voltage on the memristor small and does not change the state between the conductive filaments240.
Fig. 15. Logic in computing. a, Processing flow of CsBi3I10/SWCNTs transistor. b, Implementation of NOR logic239. Reprinted with permission from ref.239. © 2021 Elsevier Ltd. c, Al/NiAl-LDHS/graphene oxide/SiO2/Si memristor. d, D-type latch principle. e, D = 0, Qn, = 1, Qn+1 = 0. f, D = 0, Qn = 0, Qn+1 = 0. g, D = 1, Qn = 1, Qn+1 = 1. h, D = 1, Qn = 0, Qn+1 = 1240. Reprinted with permission from ref.240. © 2022 Elsevier B.V. |
The connection between each two neurons in artificial neural network (ANN) represents the weight value of the signal passing through the connection, which is called weight, which is equivalent to ANN memory. Memristors are mainly used as synaptic devices in neural networks, and their multi-value regulation characteristics under pulses perfectly realize hardware mapping of synaptic weights. Fig. 16a illustrates the network architecture of a typical pattern recognition system, which consists of an input layer and an output layer. Each pixel of the image corresponds to an input neuron. The system has a total of 28 × 28 input neurons, and the intensity values of the image pixels are simulated by altering the timing values of synaptic spikes. Fig. 16b shows the emulation of synaptic functions through a combination of inverters and synaptic transistors, and the creeping current problem of the memristor array is addressed using a three-terminal synaptic device. By designing different synaptic spike timings, the perception of different image pixel values is achieved. The temporal correlation between pre-synaptic and post-synaptic spikes is translated into various pulses applied to the synaptic transistors, as shown in Fig. 16c. Different pulses stimulate the transistor, triggering an update of the transistor's nonlinear weights, akin to the weight updating process used to train neural networks. The recognition capability of the system was validated using 28 × 28 pixel handwritten digits, and Fig. 16d shows the final simulated conductance state of synaptic transistors connecting input neurons and each output neuron, where the randomly initialized synaptic weights ultimately learned to encode the input images. Fig. 16e shows the relationship between the recognition rate and the number of output neurons (N). It can be seen that increasing the number of output neurons (N) can improve classification accuracy, with the recognition rate reaching 60% to 70%, when there are 80 output neurons241.
Fig. 16. Artificial neural network. a, The network architecture of the pattern recognition system. b, The simulation circuit of synaptic functions. c, The temporal correlation between pre-synaptic and post-synaptic spikes and its conversion into various pulses applied to synaptic transistors. d, Rearranging the weights of the connections from input neurons to output neurons. e, The relationship between performance and the number of output neurons241. Reprinted with permission from ref.241. © 2017 American Chemical Society. |
Table 2. Recognition rate of neuromorphic devices on corresponding data sets. |
| Device | Structure | Data set | Accuracy rate | Refs. |
|---|---|---|---|---|
| Pt/MXene/Pt | Two-terminal | MNIST | 80.6% | 65 |
| Channel: CeOX-MXene-Zn2SnO4 Gate: n++ Si | Three-terminal | MNIST | 98.3% | 68 |
| Pd/CQDs/Ga2O3/Pt | Two-terminal | MNIST | 92.63％ | 199 |
| Ag/HfO2/GQDs/Pt | Two-terminal | MNIST | 90.91% | 201 |
| Cu/Ti3C2/BFO/Pt | Two-terminal | CIFAR-10 | 90% | 213 |
| Pt/MoTe2/CQDs/ITO | Two-terminal | MNIST | 96.87% | 245 |
| Au/graphene−TiO2/Au | Two-terminal | MNIST | 92.2% | 246 |
| Channel: HfO2/AlOx/HfO2/CNT Gate: | Three-terminal | MNIST | 94.5％ | 247 |
| Channel: P(VDF-TrFE/SWCNT) Gate:n++ Si | Three-terminal | MNIST | 86.8％ | 248 |
| Channel: CNT/SiOx/Au/SiOx Gate: Pd | Three-terminal | MNIST | 90％ | 249 |
| GO/SF/GO | Two-terminal | MNIST | 92.3％ | 250 |
| Al/MXene-ZnO/ITO | Two-terminal | MNIST | 82.96% | 251 |
| Au/MXene/Cu | Two-terminal | CIFAR-10 | 87.5% | 252 |
| Cu/MXene/PZT/Pt | Two-terminal | MNIST | 95.13% | 253 |
| Channel: SWCNTs/F8T2/Al2O3 Gate: Al | Three-terminal | MNIST | 94.94% | 254 |
| Cu/TaOx/CNT | Two-terminal | MNIST | 95.49% | 255 |
| Channel: S-MXene Gate:Au | Three-terminal | Alphabet dataset | 99.3% | 256 |
Abbreviations: BFO, BaFe12O19; CIFAR, Canadian Institutes for Advanced Research; CNT, carbon nanotube; CQD, carbon quantum dot; MNIST, Mixed National Institute of Standards and Technology database; PZT, PbZryTi1−yO3; SWCNT, single-walled carbon nanotube. |
Artificial vision and tactile sense systems
In the recent years, the use of neuromorphic devices to build artificial neuromorphic sensory systems has attracted much attention257. Eighty percent of the external information received by the human body comes from vision. Vision is an indispensable sense of consciousness for the human body258. With the rise of bionic robots, there is an increasing need to simulate the visual behavior of hardware at the biological level259. Researchers have therefore proposed and designed an artificial vision system based on memristors. The wavelength, intensity, and duration of external light will affect the output current of the memristor, thereby simulating the human body's response to different external light signals260,261. Wen et al. produced a three-terminal memristor device to realize the recognition of red, green, and blue light, as shown in Fig. 17a. The bottom gate of the three-terminal memristive device is n++ Si, the source and drain materials are Al, and the channel layer is a composite of GeOx-modified MXene and zinc tin oxide (ZTO). Under visible-light irradiation of red, green, and blue, the output current value of the three-terminal memristor device increases significantly68. Yan et al. developed a photomemristor based on flexible carbon dot nanoribbon materials to achieve fast memory of Chinese characters. Using carbon dot nanoribbons as the resistive layer, a flexible wearable artificial vision system with TiN/carbon dot nanoribbons/ITO/mica structure was constructed, as shown in Fig. 17b. The optical stimulation of Chinese characters acts directly on the visual system, and the artificial vision system can detect light and store information at the same time. Under the stimulation of 100 light pulses, the artificial vision system can achieve rapid memory of Chinese characters262. Yao et al. prepared an artificial visual perception system composed of RRAM devices and photosensitive electronic components based on PVA and GO hybrid materials, which provides a way for future humanoid robot research and simulation of human visual neural networks, as shown in Fig. 17c. When the photosensitive device is exposed to a light pulse with a power of 175.2 μW cm−2 and a wavelength of 532 nm, the voltage value across the RRAM device is rapidly increased to 2 V, and the current value rises rapidly and returns to the initial state after 100 s, simulating the human eye's perception of light263.
Fig. 17. Artificial vision and tactile sense systems. a, Three-terminal visual memristive device based on GeOx-modified MXene nanosheets68. Reprinted with permission from ref.68. © 2023 Elsevier Ltd. b, Flexible wearable artificial vision system with TiN/carbon-dot nanoribbon/ITO/mica structure262. Reprinted with permission from ref.262. © 2023 The Authors. c, Based on Ag/PVA@GO/ITO artificial vision system with RRAM and photosensitive electronic components263. Reprinted with permission from ref.263. © 2022 The Authors. d, Smart skin based on flexible iron electrets and SWCNT synaptic transistors264. Reprinted with permission from ref.264. © 2020 American Chemical Society. e, Intelligent tactile sensing system based on MWCNTs piezoresistive film arrays and memristor chips265. Reprinted with permission from ref.265. © 2022 American Chemical Society. f, Tactile sensing system based on semivolatile CNT transistors for sensory neurons and perceptual synaptic networks266. Reprinted with permission from ref.266. © 2020 The Author(s). Abbreviations: CNT, carbon nanotube; MWCNT, multi-walled carbon nanotube; RRAM, resistive random access memory; SWCNT, single-walled carbon nanotube. |
The skin is the largest organ in the human body and has complex tactile perception capabilities. As a result, there has been an endless stream of research into electronic skins designed to simulate the functions of the human skin. Research on electronic skin is mostly based on passive pressure-sensitive components, which only convert pressure signals into electrical signals through changes in conductance and capacitance, lacking intelligence. In recent years, scientific researchers have been inspired by the working mechanism of human tactile nerves and have developed a number of tactile sensing systems by combining the functions of neuromorphic devices with the sensing capabilities of an electronic skin267. Wang et al. used flexible iron-electret nanogenerators to convert force signals of different sizes and frequencies into presynaptic action pulses. The presynaptic action pulse acts on the gate of the synaptic transistor, causing changes in the postsynaptic current, as shown in Fig. 17d. The synaptic transistor selects PI as the substrate, Al2O3/SiOx (atomic layer deposition and electron beam evaporation) as the dielectric layer, and MWCNTs as the transistor channel material. Under the force stimulation of 14.2 N, the synaptic weight changed by 56.5%, proving that changes in external pressure can cause changes in synaptic weight, successfully simulating the human tactile nerve function264. Tang et al. integrated a high-performance piezoresistive-film array based on MWCNTs with a memristor-based memory computing chip to achieve intelligent tactile sensing at the hardware level, as shown in Fig. 17e. MWCNTs have a high aspect ratio and good electrical conductivity. By mixing and coating MWCNTs with thermoplastic polyurethane elastomer and using methods such as photolithography and atomic layer deposition, a 4-inch TFT MWCNT array was manufactured. Using a commercial foundry's 130-μm process, the TiN/HfOx/TaOy/TiN memristor, analog-to-digital converter, digital-to-analog converter, driver, and buffer are integrated into a memristor chip. Recognition of handwritten digits is achieved by touching the piezoresistive film array and coordinating on-chip processing of the memristor chip265. Choi et al. constructed a tactile sensing system consisting of a tactile sensor, a voltage-controlled oscillator, a neuron device, and a synaptic network, as shown in Fig. 17f. The tactile sensor is consists of two layers of PDMS sandwiched between a CNT network film, and the neuron device is composed of CNT transistors. The tactile sensor converts pressure changes into resistance changes, and the voltage-controlled oscillator converts the resistance changes into digital signals of different frequencies. The signal output by the voltage-controlled oscillator is introduced into the CNT transistor, and a leakage integral output corresponding to the output frequency of the voltage-controlled oscillator is generated. Finally, the sampled outputs of the CNT transistors are sent to the synaptic CNT transistor network operating in a non-volatile mode, and the learning/recognition process of distinguishing the input stimulus patterns is performed in a supervised learning manner. Test results show that the recognition accuracy of tactile system for different bump situations is higher than 95%266.
Multimodal perception system
Multimodal perception systems usually refer to systems that combine multiple perceptions, such as bionic vision, hearing, touch, and neuromorphic computing devices. This system has multiple sensing capabilities and can better simulate the human body's perception of the outside world. Wang et al. used phototransistors to sense external light and flexible polypropylene-based iron electret nanogenerators to sense external tactile and auditory stimuli, as shown in Fig. 18a-c. The output pulse amplitude of the phototransistor increases with increasing light intensity and produces nearly symmetrical enhancement and suppression pulses under the same light stimulus. For hearing and touch, although it is achieved by using the same flexible polypropylene-based iron electret nanogenerator (FENG), the form factor and operating mode are different. The FENG used for auditory signal detection operates in “stand-alone” mode, with only four edges fixed in a frame to allow free vibration of the membrane. Since the pressure caused by a sound signal is usually much lower, amplification of the pressure signal is achieved using a gain circuit. The FENG used for tactile signal detection operates in “blocking” mode, with one side firmly supported by a rigid substrate. The FENG can generate enhancement and suppression pulses of different amplitudes for different sound signals and pressure signals. The pulses generated by the phototransistor and FENG are input into the gate of the SWCNTs synaptic transistor as a stimulation signal, and the multimodal sensing system is shown in Fig. 18d. Different light intensities, sound decibels, and pressures cause different gate voltage values of the input SWCNTs synaptic transistors, resulting in different output conductance states of the synaptic transistors, as shown in Fig. 18e-g, thus realizing the perception of external information. This multi-modal perception system can not only expand the application scenarios of artificial intelligence by introducing environmental interaction functions, but can also be used in fields such as robotic systems and prosthetics69.
Fig. 18. Multimodal perception system. a, Visual information collection. b, Auditory information collection. c, Tactile information collection. d, Multi-modal perception system integration. e, The repeatability of synaptic transistors to visual stimulation. f, The repeatability of synaptic transistors to auditory stimulation. g, Reproducibility of synaptic transistors in response to tactile stimulation69. Reprinted with permission from ref.69. © 2021 American Chemical Society. |
CONCLUSION AND OUTLOOK
As computational complexity and power consumption increase, the technological development of traditional computers based on the von Neumann architecture has encountered problems such as memory walls and power consumption walls. Based on the basic component of memristors, it is expected to realize a neuromorphic computing system that integrates storage and calculation at the hardware level, which is an effective solution for processing large amounts of complex data. Carbon nanomaterials have the advantages of multiple dimensions, multiple hybridization methods, excellent electronic properties, and good thermal stability. They have broad application prospects in the field of electronic devices and have become strong candidates for realisation of high-quality information-storage RRAM and high-performance neuromorphic computing. In this paper, we start from the dimension of carbon nanomaterials and respectively analyze 0D carbon nanomaterials (CQDs, GQDs), 1D carbon nanomaterials (SWCNTs, MWCNTs, carbon fibers), and the material structural characteristics, physical properties and synthesis methods of 2D carbon nanomaterials (graphene, GO, reduced GO, MXene) are summarized. In addition, the materials and functions of the electrodes and resistive layers of memristors are reviewed. Based on the nanostructures of different carbon nanomaterials, the research progress of memristors based on carbon nanomaterials for RRAM and neuromorphic devices are reviewed. Finally, the application of carbon-based RRAM and neuromorphic devices are summarized in logic operations, neural network construction, and artificial intelligence. Potential applications in visual systems, artificial tactile systems, and multimodal perception systems.
Looking to the future, there is still a long way to go in the research and practical applications of RRAM and neuromorphic applications based on carbon nanomaterial memristors, which is particularly reflected in the following structural aspects (Fig. 19).
Fig. 19. The development direction and challenges of carbon based memristors. |
(1) In terms of the working mechanism of carbon nanomaterial memristor devices, the working mechanisms of existing memristive devices are mainly divided into electrochemical metallization, chemical valence change, thermochemistry, space charge limitation, photogenerated carriers and interfacial barriers, etc. Different mechanisms correspond to different conductance-value change patterns of memristors, and the conductance changes are closely related to the application of memristors. For example, the high repeatability and high stability of the conductance state of memristors are crucial for the construction of high-density ANNs. At the same time, the power consumption of existing carbon nanomaterial memristors is mostly pJ level, whereas the power consumption of human brain synapses is about fJ level. Reducing the power consumption value of memristors is also conducive to better simulating human brain functions. Therefore, optimizing the conductance state of the memristor and reducing the power consumption of the memristor by exploring the new working mechanism of the memristor is crucial to improving the performance of the carbon nanomaterial memristor.
(2) In terms of artificial vision and tactile simulation by carbon nanomaterial memristive devices, most existing memristive devices can only produce memory effects for specific wavelengths of light and cannot process multi-dimensional optical signals. It is far from sufficient for memristors to truly simulate human vision because the human eye can collect light information of multiple wavelengths. In the future, research on memristive devices for simulating artificial vision requires cross-disciplinary integration. In particular, it is necessary to integrate the advantages of electronics, neurology, pattern recognition, and other disciplines to explore the specific relationship between memristive devices and human vision. In addition, it is necessary to explore the application scenarios of memristor-based vision sensors in real life, including but not limited to optical information storage, image classification and recognition, image-edge extraction, event cameras, bionic eyes, etc. In terms of tactile simulation, most existing tactile systems are an assembly of tactile sensors and memristors. In subsequent research, directly realizing the perception of tactile signals on memristive devices can significantly reduce system complexity and improve system reliability.
(3) In terms of simulating various neural networks, most of the existing methods are to extract the weight-update parameters from the test curve of the carbon nanomaterial memristor and input them into the computer and then implement the training of the neural network in the computer and more simulation of neural network functions for memristors. Similarly, memristor-based multi-modal sensing systems also urgently need memristor chips that can directly implement neural network functions on-chip. Therefore, the design of large-scale integration strategies and multi-weighted controllable self-renewal of memristors is the only way to realize “calculation” an important part of the “sense memory calculation” of memristors, and is also the cornerstone of building memristor chips.
(4) In the process of large-scale mass production and application, existing carbon nanomaterial memristors have different processing methods. Many memristors require manual operations during the processing, and there are errors, making it difficult to ensure the consistency of mass production. In the subsequent processing of carbon nanomaterial memristors, it is necessary to explore a standardized carbon-based memristor-processing process similar to the silicon-based micro electromechanical system (MEMS) processing process. Minimizing manual operations during the processing process will help further improve the production yield and consistency of carbon nanomaterial memristors.
In the future, cooperation from multiple disciplines such as optics, microelectronics, materials science, chemistry, artificial intelligence, and neuroscience will be needed to truly realize the potential of carbon-based memristors and apply them to high-performance storage, computing chips, and multimodal perception systems. It can be firmly believed that this review will provide useful guidance for the theoretical research, performance improvement, and large-scale integration of RRAM and neuromorphic applications based on carbon-based memristors.
MISCELLANEA
Funding This work was supported in part by the National Key Research and Development Program of China under Grant 2021YFF0603500; in part by the National Nature Science Foundation of China under Grants 62174068, 62311540155, and U22A2014; in part by the Shandong Provincial Natural Science Foundation of China under Grant (ZR2023ZD03); in part by the Jinan City University Integration Development Strategy Project under Grant (JNSX2023017).
Declaration of competing interest The authors declare no competing interests.
