Research Process of Carbon Dots in Memristors

The explosive growth of digital communication promotes advanced memory and computing devices in Big Data and artificial intelligence era. In particular, memristors hold great promise for in‐memory computing and artificial synapses, expected to break through restrictions on hardware computational power and storage capacity caused by the von Neumann bottleneck and declining Moore's Law. The memristance controllability is vital to memristors, in which functional layer materials are key. As novel functional layer materials, carbon dots (CDs) are expected to overtake silicon‐based materials due to their satisfactory modulation effects on memristance and distinguished memristive performances. The combination of CDs and matrix materials may carry out multimode sensation memristors for interactive intelligent systems. This review first gives a brief introduction to memristors and their functional layer materials, especially different CDs and their relations with memristance. Then the modulation effects of CDs on memristance are highlighted, mainly including the local electric field enhancement effect, the electron trapping and detrapping effect, and the photosensitization effect, accompanied by applications of CDs‐based memristors (CDMs). Lastly, challenges and perspectives of CDMs are pointed out. This work is rewarding to understand the role of CDs in memristors, to guide relevant research about CDMs, and to promote implementation for intelligent memory and computing.


Introduction
The wave in the information epoch has gone beyond the realm of traditional desktops. Mobile interactive systems and industry 4.0 technologies are implanting into human life in various forms, giving rise to numerous interactive data. [1,2] excellent solution processing, high conductivity, biocompatibility, and specific quantum effects, which can provide memristors with distinguished memristive performances. [24][25][26][27][28][29][30] The key performance parameters of a majority of CDs-based memristors (CDMs) have been summarized in Table 1, from which it can be found that nearly half of the reported threshold voltages are lower than 1.0 V, while the threshold voltages of other memristors based on metal oxides, perovskites, and organic materials are above 1.0 V as a whole. [31][32][33] The low threshold voltage is beneficial to modulate memristance and reduce energy consumption. In addition, CDMs also present reasonable switching current ratio, relatively long retention time, and high endurance, suggesting CDs have excellent application potential in memristors.
Currently, the universal explanation about the modulation effects of various materials on memristance, including CDs, is still absent. [31] On the one hand, the research about CDMs is in its infancy. On the other hand, the practical factors interfering the modulation effects are difficult to be controlled, especially the stochastic spatial distribution of CDs and the uncertain migration path of ions. Related controversies limit the development of high-performance CDMs and hinder the establishment of CDMs paradigm. Therefore, it is urgent and necessary to explore the modulation effects of CDs on memristance.
This review will start with the fundamentals to memristors, including their classification, working mechanisms, and functional layer materials, followed by presenting different CDs features related to memristors and discussing integrated memristor array based on CDs. After that, the modulation effects of CDs on memristance are mainly discussed, including local electric field (LEF) enhancement effect, electron trapping and detrapping effect, photosensitization effect, and other effects. Meanwhile, the involved applications of CDMs are also introduced passingly. Finally, challenges and perspectives of CDMs will be presented.

Classification and Mechanism of Memristors
In 1971, Chua proposed and described the fourth fundamental passive circuit element, i.e., memristor, by analyzing the axiomatic definition and logical circuit theory of the capacitor, inductor, and resistor ( Figure 1b). [34] If the memristance M is itself a function of the flowing charge q rather than a constant, memristor will present nonlinear electrical characteristics. [20,35] It was until the year of 2008 that the first recognized memristor (Pt/TiO 2 /Pt) was developed by Strukov et al. [20] After that, the definition of memristor was further broadened to any two-terminal devices that show pinched hysteresis loops in the current-voltage (I-V) curves and periodic response of the same frequency driven by periodic voltage signal. [32,36]

Classification of Memristors
Memristors can be classified into digital and analog types depending on the discrete or continuous characteristic of resistive switching. [37] In digital-type memristors, resistance can be changed between multiple discrete states, showing a digital-type resistive switching (DRS) characteristic (Figure 1c). High resistive state (HRS) and low resistive state (LRS) can be regarded as binary information of "0" and "1." Therefore, digital-type memristors can also be deemed as the resistive random access memory. The performance of the digital memristors is mainly evaluated by threshold voltage (V SET and V RESET ), switching current ratio (I ON /I OFF ), retention time, endurance, power consumption, and switching speed. [38][39][40][41][42] The electron transport modes in memristors need to be introduced here to analyze DRS behavior, which can be judged by fitting the I-V curves. i) Schottky emission (ln(I) ∝ V 1/2 ), which is known as thermionic emission. In this mode, thermally activated electrons overcome the energy barrier and escape from the electrode. [43] ii) Ohmic conduction (I ∝ V), in which current is proportional to electric field. [44] iii) Space charge limited current (SCLC), which is consisted of the Ohmic region (I ∝ V), the Child's region (I ∝ V 2 ), and the trap charge limited current In memristor (dϕ = M · dq), capacitor (dq = C · dv), inductor (dϕ = L · di), and resistor (dv = R · di), the variables q, ϕ, i, and v refer to the charge, magnetic flux, electric current, and voltage, and the physical quantities M, L, R, and C represent the memristance, inductance, resistance, and capacitance. c) I-V curves of the digital-type memristor. d) I-V curves of the analog-type memristor. e) Schematic of the connection between neural synapse and memristor device for artificial synapse. (TCLC) region (I ∝ V n , n ≫ 2). [45,46] In the Ohmic region, current is proportional to electric field; in the Child's region, the injected electrons are trapped to form the space charge effect and drift current; in the TCLC region, the traps are gradually filled by electrons with the increasing electric field, then the concentration of free electrons increases sharply, and the device enters LRS. [45][46][47][48][49] iv) Fowler-Nordheim (F-N) tunneling (ln(I/V 2 ) ∝ 1/V), a quantum effect that electrons directly pass through the triangular barrier at the interface between the metal and functional layer. [50,51] v) Poole-Frenkel (P-F) emission (ln(I/V) ∝ V 1/2 ), in which the electrons in traps are excited into the conduction band. The electric field can decrease the Coulombic potential barrier, and the disorder thermal disturbance increases the probability of electrons escaping from the traps. [45,52] The analysis of these electron transport modes is beneficial to reveal the mechanisms and improve the performance of memristors. Different from digital-type memristors, analog-type memristors feature continuously tunable resistance states, exhibiting an analog-type resistive switching (ARS) characteristic ( Figure 1d). The biological synapse weight, which is used for describing connection strength between neurons, possesses gradual tunability. [21,53,54] Therefore, analog-type memristors can be used to mimic biological synapses. At present, the ANN systems based on transistors face the problems of complicated circuit integration processes and low energy efficiency, which bring tremendous difficulty to large-scale integration. [30,55] Memristors are a better choice for building ANN systems. One reason for this is the three layers of memristors correspond to the pre-synapse, synaptic cleft, and post-synapse of the neural synapse, respectively ( Figure 1e). The migratory metal ions (e.g., Ag + , Cu 2+ ) or oxygen vacancies in memristors are used to emulate neurotransmitters, and the synaptic weight change (ΔG) can be adjusted by the conductance of memristor. [56,57] In addition, the nonlinear memristive behaviors are similar to the nonlinear transmission of neural synapses. Therefore, memristors can realize the perception, thinking, and learning functions by regulating the synaptic weight, and it is termed synaptic plasticity. [58,59] Synaptic plasticity includes short-term plasticity (STP) and long-term plasticity (LTP), which have been recognized as the basis of nerve cells learning and memory. [60] STP dynamically regulates the synaptic transmission efficiency at the conveying transients, and remains the efficiency unchanged during steady-state transmission, corresponding to the short-term memory behavior of the brain. [61] In contrast, LTP can make the synaptic transmission efficiency change stably, which indicates the attenuation barely for synaptic weights after repeated stimulation, matching with a long-term memory behavior of the brain. [60,62,63] The typical paired-pulse facilitation (PPF) belongs to STP. PPF refers to the behavior that the synaptic weight change caused by the second stimulus is higher than that of the first stimulus when the interval time between two adjacent stimulation signals on synapses is very short. [62] Spiking timingdependent plasticity (STDP), as an advanced LTP behavior, is not only an extension of Hebbian learning rules but a foundation for simulating elaborate brain learning. [11] Beyond these behaviors, excitatory post-synaptic current (EPSC), the transition from STP to LTP, and inhibitory post-synaptic current are the common synaptic plasticity in the related research. [37,64] In short, analog-type memristors are naturally suitable for synapse learning and perception, making it possible to perform intelligent computing.

Working Mechanisms of Memristors
Some mechanisms have been proposed to explain the memristive phenomenon, which is contributed to the material design and device optimization, although they are still controversial.

Conductive Filament Mechanism
This mechanism originates from the formation and rupture of the conductive filaments in functional layers, which can be indirectly verified by analyzing the relation between resistance value and temperature, [30] and can also be directly observed using the high-resolution transmission electron microscopy, such as the discovery of Magnéli phase filament and quasi-core-shell conductive filament system consisting of metallic hexagonal-Hf 6 O and its monoclinic or tetragonal HfO x . [65,66] The current increases gradually with the formation of more conductive filaments, and the memristor device enters LRS after performing SET process, while the situation is opposite for the dissolution of conductive filaments. The conductive filament mechanism can be further divided into electrochemical metallization and valence change mechanism according to the migrating ions species. [21] In electrochemical metallization mechanism (Figure 2a), Ag and Al are usually selected as top electrodes. The generated metal ions resulting from the oxidation reaction www.advelectronicmat.de of these active metals migrate to bottom electrode when a positive voltage is applied to top electrode, then these ions are reduced to the conductive metal filaments during migration. In a reverse manner, these filaments are dissolved when applying a negative voltage. [67] In valence change mechanism, inert metal and oxide usually serve as top electrode and functional layer, respectively. The oxygen vacancy electromigration realizes SET or RESET process under different electric fields. [68] In addition, several works also reported the sulfur and nitrogen vacancies mechanism. [42,69]

Electron Trapping and Detrapping Mechanism (Figure 2b)
This mechanism can be understood from a classic model of Cu 2 O-based resistive switching device. [17] The potential well of functional layer materials can trap free electrons. SET process is executed after the defect is filled. When the reverse voltage is exerted, the electrons are gradually released from the defects until RESET is operated. The charge trapping and detrapping tends to occur in the materials with abundant surface defects. [70][71][72]

Phase-Change Mechanism (Figure 2c)
The different crystal phases in materials have different conductivities. The phase change materials can generate the ordered crystal phase with high conductivity under an electric field, inducing the memristor to step into LRS. While the disordered crystal phase with low conductivity can be generated when reversing the electric field, making the device return to HRS subsequently. [33,73] In addition to these working mechanisms, interface modulation mechanism and charge transfer mechanism are also used for explaining memristive results as well. Interface modulation mechanism mainly refers to the regulation of the Schottky barrier layer, which means that electrons trapping/detrapping and vacancies can rebuild the interface barrier to create memristive behaviors. [74] Charge transfer mechanism usually applies to donor-acceptor organic semiconductors, where electrons can be effectively transferred from electron donors to acceptors under an external electric field, facilitating the transition between HRS and LRS. [33] In above-mentioned mechanisms, the static distribution and dynamic migration of ions, traps, and materials are inherently random, resulting in the difficulty of modulating memristance. However, these mechanisms can support a deep cognition to the working mode of memristors. In addition, software simulations have made progress in related materials and devices, such as the first-principles calculations of resistive materials and the simulation of devices in various physical fields, which can provide researchers with a deeper understanding of memristors. [75]

Functional Layer Materials of Memristors
The functional layer, which is also called the insulating layer, active layer, or dielectric layer from the perspective of materials, is the core of the memristor. [76][77][78] It is imperative for memristors to explore the ideal functional layer materials with controllable memristance property and excellent comprehensive performances containing low threshold voltage, long retention time, high endurance, and low power consumption. At present, functional layer materials mainly include metal oxides, perovskites, solid electrolytes, organic materials, 2D materials, and carbon materials. [32,33,60,78]

Common Functional Layer Materials
Metal oxides are prevalent in memristors because of their simple structures, high thermal stability, and easy compatibility with CMOS process. [77] Transition metal oxides predominate metal oxide-based memristors, such as TaO x , TiO x , CuO x , WO x , and HfO x . [79,80] The metal oxides are naturally rigid and require complicated treatment for flexible wearable memristors.
Perovskites have also been studied in functional layers due to their simple synthesis process, low cost, and flexibility. They primarily include organic-inorganic hybrid perovskites, all-inorganic cesium/rubidium lead halide perovskites, leadless and lead-free perovskites, and halide perovskite quantum dots. [32] Because halide ions possess low activation energy of migration, perovskite-based memristors usually show low threshold voltage and working current, suggesting these devices are potential candidates for low-energy artificial synapse. [81][82][83] Solid electrolytes, such as Ag 2 S, CuS, and CuGeTe, usually have lower threshold voltage due to the low cation migration barrier. [84] However, some perovskites and solid electrolytes suffer obstacles to practical applications due to instability in air and moisture and incompatibility with CMOS process reasons. [81,84] Organic materials with low-cost solution processing can be used in flexible wearable applications. [85] Organic materials have the structural adjustability. By replacing different substituents of organic molecular to adjust steric configuration and energy level depth, memristive performances can be regulated subsequently. [86] The inherent inferior thermal stability of organic materials may lead to inferior retention. [84] The emergence of 2D materials boosts the development of memristors. 2D materials used for the functional layer mainly include graphene and its derivatives, transition metal dichalcogenides (TMDs), nitrides, and so on. Graphene and its derivatives can improve memristor performance due to their stable monoatomic structure, high thermal conductivity, and reliable mechanical properties. [78] TMDs (such as MoS 2 , WS 2 , MoTe 2 ) have many features such as tunable bandgap, high carrier mobility, and phase transition, which is beneficial to modulate memristance of TMDs-based memristor. [12,68] The nitrides applied in memristors have also attracted much attention, and the representative hexagonal boron nitrides feature with flat/ uniform surface, superior chemical stability, and high thermal conductivity, which are conducive to reducing the memristance variations, inhibiting the interaction with adjacent layers and dissipating the heat, respectively. [78] However, 2D materials also face some knotty problems such as Van der Waals gap and Fermi pining that can harm memristor performance by inducing barriers and gap states. [78] www.advelectronicmat.de As for carbon materials applied to functional layer of memristors mainly contain CDs, carbon nanotubes, carbon nanofibers, and so on. They hold the comprehensive merits of high electrical conductivity, thermal and mechanical properties, and easy integration with the CMOS technology. [22,87,88] The silicon element is beyond its power to support Moore's law. While the carbon element, the same congener as the silicon, is believed to be a potential candidate to continue Moore's law. [22] Among these carbon materials, CDs show effective memristance optimization effects. [15,89,90]

CDs in Functional Layers of Memristor
CDs are considered as a kind of quasi-0D nanomaterials with the size generally less than 10 nm, abundant surface functional groups (e.g., carboxyl, hydroxyl, amino, and methyl), and sp 2 and/or sp 3 carbon atoms in their interiors. [91,92] Since CDs were discovered in 2004, CDs have been extensively employed in optoelectronics, photovoltaics, biomedicine, and catalysis due to their high stability, simple preparation, and low toxicity. [93][94][95][96][97] In recent years, CDs are increasingly favored by memristors based on their easy compatibility with CMOS process.
The quantum effects of CDs, such as quantum confinement effect and Coulomb blockade effect, are very likely to refine the memristance of CDMs in different ways. [26][27][28][29][30] The quantum confinement effect plays a vital role in quantizing electron energy levels when the size of CDs and their Bohr radius of excitons are in the same order of magnitude, which can act on band gap, energy level alignment, and interfacial potential barrier of memristors. [26,98] In the Coulomb blockade effect, the system energy rises after the injected electrons reaching CDs, which may block other electrons along the conducting paths until the trapped electrons escape from CDs, realizing the storage and release of electrons in memristors. [27][28][29] Compared with semiconductor quantum dots, such as perovskite quantum dots and chalcogenide quantum dots, CDs have better eco-environmental protection characteristic and physicochemical stability, which can avoid the problems of toxicity and moisture exposure. [32,99] CDs are generally considered to have four types, including graphene quantum dots (GQDs), carbon quantum dots (CQDs), carbon nanodots (CNDs), and carbonized polymer dots (CPDs). [92] The schematic of different CD structures that may relate to memristor performances is listed in Figure 3. GQDs consist of π-conjugated single or few graphene sheets with a lateral dimension less than 20 nm and thickness less than 2.5 nm, and fewer chemical groups on the surface. [92] GQDs are mainly obtained by cutting graphene-based materials during top-down synthesis process. While other CDs can be synthesized by pyrolysis of organic molecules or polymers subjected to high temperature using bottom-up methods. Hence, the evidence to distinguish GQDs and other CDs lies in the used precursors. Internal graphene carbon skeleton can improve the stability of GQDs, which is helpful to optimize spatial and temporal variability of memristors. The sp 2 -hybridized graphene domains with large content are conducive to transferring the electrons injected from the bottom electrode, which can reduce power consumption. As verified from Table 1, most GQDsbased memristors show lower threshold voltages than other CDMs. The intrinsic graphene core also endows GQDs with quantum confinement effect and conjugated π-domains. [93] In the quantum confinement effect, the bandgap, energy level alignment, and interfacial potential barrier of the functional layer containing GQDs can be regulated by controlling their particle sizes for optimizing device performance. As for conjugated π-domains, the surface oxygen defects around graphene domains can influence memristance through electron trapping and detrapping effect. Moreover, the zigzag and armchair edges of GQDs generally possess triple and singlet carbenes each, which may affect molecular orbital and bandgap of GQDs. [93] The radicals of triplet carbene may improve the memristance stability owning to the delocalization of radicals, just like the reported molecularly inherent radical memory. [76] In addition, www.advelectronicmat.de the radicals have the magnetic property naturally and correlate to optical performance. In this sense, the research about the edge carbenes of GQDs may bring a new idea for building intelligence devices integrated the light sensing, magnetic memory, and computing functions. The main difference between CQDs and CNDs is not only whether possessing crystalline structure, but also whether presenting quantum confinement, which has become the main strategy to distinguish CQDs and CNDs. Specifically, CQDs possess sp 2 /sp 3 carbon cores and some surface organic groups, which are spherical-like carbon nanoparticles that feature with high crystallinity and carbonization degree. [92] Quantum confinement effect determined by carbon core of CQDs may also dominate the performance of memristors. CNDs are a kind of amorphous spherical carbon nanoparticles that lack quantum confinement and crystal lattice structure. CNDs have high carbonization degree and abundant surface defect sites. [92] The functional groups, which have various energy levels, can be designed and introduced to obtain CNDs with different electron traps. [93] Therefore, memristor performances can be adjusted by controlling the functional groups that trap and detrap electrons. CPDs are a kind of incompletely carbonized CDs with abundant functional groups and polymer chains on their surface. [93] Although the research on CPDsbased memristors was rarely reported, it is speculated that their surface functional groups, conjugated π-domains, molecular states, and crosslink structures may be related to the performance of the memristors based on CPDs.
These CDs doped the second-period elements for memristor have also been explored, such as oxidized carbon quantum dots (OCQDs), graphene oxide quantum dots (GOQDs), N-doped carbon quantum dots (NCQDs), and B-doped quantum dots. These elements of CDs can provide electronic trapping sites, increase oxygen migration energy barrier, or separate photogenerated electron-hole pairs, which endows CDMs with light sensation function or long retention time. [16,38,90,97] The system of CDs doping into some matrix materials was also explored, which can make the distribution of CDs more uniform in functional layer. For instance, the composite of graphene oxide (GO) and CDs can implement the transition from DRS to ARS, [16] the memristors based on GO and CDs have the magnetoresistance phenomenon, [100] and CDs in poly(3,4-ethylenedioxythiophene):polystyrene sulfonate (PEDOT:PSS) layer can enhance the interfacial polarization field of functional layer to promote the electron transport in memristors. [101] The combination of CDs and various matrix materials probably brings the feasibility for future in-sensor computing systems by integrating the sensing, storage, and computing functions.
As the functional layer materials, CDs can also provide a new selection for memristor integration. The array scale determines the capability of intelligent computing and memory. [9] However, some fundamental problems in integrated memristor array seriously limit the application in processing complex computing tasks, such as sneak path current and random variability of memristance. [23] For the problem of sneak path current, the common solution is to integrate CMOS transistors with memristors, [9] such as using one-transistor-one-memristor (1T1R) architecture. In this case, it is a challenge to build high density arrays due to the scaling limitation of silicon-based transistors, and the inherent physical properties of CDs make it possible to solve this problem. On the one hand, the CDs-based integrated array with large scale and nanoscale resolution features is expected to be obtained through the low-cost nanoimprint lithography by virtue of the excellent solution processing of CDs. [33] On the other hand, some CDs with self-assembly property have been reported so far, [94] and such CDs with quantum size can form regular and ultra-high-density bit patterned media through directed and in situ self-assembly, which makes it possible for the integrated application of CDMs. Therefore, CDMs hold promise to detour the expensive extreme ultraviolet lithography technology to obtain the high-density integrated CDMs array. Another problem of integrated memristor manifests as the unstable performance of memristor array, especially the device-to-device variation of memristance can decrease data-processing accuracy during programming. [23] CDs possess an efficient modulation effect on memristance, mainly including local electric field enhancement effect, electron trapping and detrapping effect, and photosensitization effect, which will be discussed in detail later in this review. This means that the integrated array based on CDMs has reliable practicability. In short, as the functional layer materials in memristor, CDs are expected to realize large-scale integration of memristor.

The Modulation Effects of CDs on Memristance
The research about CDMs is still in its infancy because of lacking unanimous consensus about the modulation effects of CDs on memristance. However, some reports have explained related effects from different perspectives. With careful consideration range from CDs features, electron transport modes, and CDMs working mechanisms, the modulation effects of CDs on memristance are mainly divided into three ways: LEF enhancement effect, electron trapping and detrapping effect, and photo sensitization effect, as shown in Figure 4. Moreover, multiple modulation effect may exist in memristor for same CDs under certain conditions.

LEF Enhancement Effect of CDs
In LEF enhancement effect, the electric-field intensity near CDs is higher than other areas in functional layer because the electric-field lines converge around CDs to form a stronger tip electrode, accordingly regulating the growth site and direction of conductive filaments. [15,30] The conductive filaments formed by LEF enhancement effect are inclined to distribute near CDs and tend to grow along the same path during repeated switching cycles, which can reduce the dispersity of conductive filaments for strengthening the modulation effects of CDs on memristance. Up to now, LEF enhancement effect has been used to optimize and modulate memristance performance in memristors by GQDs, CQDs, and CNDs.
As a derivative of GQDs, GOQDs can modulate memristance because they naturally possess the conjugated π-domains, quantum confinement effect, and edge effect. Yan's group reported a memristor with the structure of Ag/Zr 0.5 Hf 0.5 O 2 (ZHO)/GOQDs/ZHO/Pt (Figure 5a,b) in 2017. [102] To explore the modulation effect of GOQDs on memristance, www.advelectronicmat.de a contrastive device (Ag/ZHO/Pt) was fabricated. The results in Figure 5c-e show that SET voltage of the GOQDs-inserted memristor shrinks from (+0.08 to +1.25 V) to (+0.08 to +0.30 V), and from (−0.01 to −0.25 V) to (−0.01 to −0.14 V) for the RESET voltage. Such lower and narrower threshold voltages are conducive to reducing power consumption and improving data reading accuracy. They believe LEF enhancement effect caused by the embedded GOQDs can shorten effective distance of Ag + migration, resulting in the growth of Ag conductive filaments and controllable memristance. Here, Ag conductive filaments are indirectly verified by the relation between the resistance value and temperature at LRS (Figure 5f). The slope of resistance ratio (R/R 0 ) versus temperature (T) is 2.66 × 10 −3 K −1 , close to the Ag nanowires (4 × 10 −3 K −1 ). Thus, Ag filaments become the conductive path of the GOQDs-inserted memristor. This work elucidates memristors can be modulated effectively by virtue of LEF enhancement effect of GOQDs.
The synthetic solution-processed MgO-GOQDs was selected as the functional layer material for researching memristance modulating effect and developing an advanced flexible artificial synapse device (Al/MgO-GOQDs/indium tin oxide (ITO)/ polyethylene terephthalate (PET)) ( Figure 5g). [8] Compared with the contrastive device without GOQDs, the operating voltage and its relative fluctuation of the device containing GOQDs are reduced by 60% and 80%, respectively. Performances were optimized by GOQDs, which is attributed to the enhanced LEF effect and the oxygen reservoir effect (Figure 5h). They also simulated important biological synaptic behaviors successfully, including PPF, STP, LTP, transition from STP to LTP, and STDP, as shown in Figure 5j,k. It is worth noting that the memristor fabricated on PET substrate exhibits stable resistive switching behaviors after multiple bending cycles (Figure 5i). This study shows the solution-processed GOQDs composite can modulate memristance by the enhanced LEF effect and has a potential in flexible artificial neural synapses.
CQDs modulate memristance by LEF effect have also been explored. Yan's group proposed a memristor (Pd/CQDs/ Ga 2 O 3 /Pt) by embedding CQDs in Ga 2 O 3 (Figure 6a) and explored its application in artificial synapse and neuromorphic computing. [30] The excellent resistive switching performance of the device is shown in Figure 6b. SET and RESET voltages are +1.70 and −0.06 V, respectively, and the corresponding powers are as low as 1.20 × 10 −7 and 4.15 × 10 −5 W. In addition, the distribution of SET and RESET voltages are very concentrated, whose variations are only 1.76% and 5%, respectively, which is profited to avoid misreading operations, accurately controlling the programming voltage in circuit design. The cross-sectional transmission electron microscopy (TEM) image in Figure 6c provides clues to the microstructure of the device at ON state, verifying the existence of carbon conductive filaments in functional layer, which probably attributes to LEF enhancement effect that can assist positively charged carbon cluster ions to nucleate and grow. The resulting carbon filaments may have lower dispersity due to the low diffusivity of carbon atoms in Ga 2 O 3 , making memristance easier to be regulated. The memristor was further emulated the synapse performances, such as STP, LTP, and four different STDP behaviors. The behaviors of Hebbian STDP simulation and Pavlovian conditioned response with associative learning and extinction (Figure 6d-f) prove the feasibility of their device in www.advelectronicmat.de simulating synaptic plasticity. Moreover, the digit recognition of the device verifies that the recognition accuracy reaches 92.63% after 250 learning phases by imitating the weight parameters in a single-layer ANN, suggesting that the device has potential in neuromorphic computing. This carbon conductive filaments-based memristor offers a new strategy for neuromorphic applications.
Liu's group demonstrated LEF enhancement effect of CQDs using the Au/CQDs-HfO 2−x /Au memristor. [15] The resistive switching performances (Figure 7a) indicate that SET and RESET voltages decrease by 1 V after inserting CQDs. Besides, the corresponding relative fluctuation declines to 6.4% and 4.9% from 14.6% and 16.2%, respectively, suggesting the optimized effect of CQDs on memristance, and then the memristor device was integrated with a p-GaN/n-ZnO heterojunction to build an intensity-controlled light-emitting diode (LED, Figure 7b), whose electroluminescence intensity can be regulated reversibly by tuning the compliance current (Figure 7c). The integrated device in this work reveals a potential application in low-cost and high-density displays.
CQDs can also undertake an LEF enhancement effect in organic matrix, not only in the previously mentioned oxides. Tang's group introduced CQDs into poly(methyl methacrylate) (PMMA) and fabricated a memristor (Ag/PMMA&CQDs/ fluorine-doped tin oxide (FTO)). [75] The device exhibits better switching repeatability and lower threshold voltages than the device without CQDs. A simulation model about CQDs in internal electric field was built by COMSOL software, discovering that the formation and rupture of conductive filaments were regulated by enhanced LEF effect of CQDs, which is in accord with the above reports about CQDs modulating memristance by LEF effect.
CNDs influence memristance property by LEF enhancement effect was pointed out by Yan's group. They prepared the Ag/HfO 2 /CNDs/Pt memristor, where CNDs were prepared from leaves (Figure 7d). [99] The results show more contractive threshold voltages after adding CNDs. The performances of CNDs in other oxides like MoO 3 and ZrO 2 were also explored, finding that the threshold voltages can be generally optimized due to LEF enhanced effect. The synaptic weight of  [30] Copyright 2020, Royal Society of Chemistry.
www.advelectronicmat.de memristors can be regulated by controlling relative timing (Δt) between pre-and post-pulse. The different neural processes such as Hebbian STDP, symmetrical STDP, and visual STDP can also be realized using different pulse waveforms, as shown in Figure 7e-g. By designing pulse amplitude, withdraw time, and relative timing dexterously, they also simulated supervised learning and interest-based learning activities, as well as mimicked the preview and review learning methods. Their work verifies the positive effects of CNDs in neuromorphic functions of CDMs, facilitating the development of ANN hardware.
In summary, the conductive filaments are mainly formed near CDs by LEF enhancement effect in CDMs, which can reduce the dispersity of filaments. Therefore, these filaments induced by CDs are more concentrated and inclined to repeat on the same path, which is beneficial to modulate memristance and improve memristor stability.

Electron Trapping and Detrapping Effect of CDs
CDs, possessing a large specific surface area, can expose more defect sites to trap and release electrons injected from the electrode. When the electric-field strength gradually increases, electrons are captured until the defects of CDs are filled and the device transfers from HRS to LRS. Conversely, when the polarity of electric field is reversed, the device returns to HRS with the escape of electrons from CDs. Just like LEF enhancement effect, GQDs, CQDs, and CNDs have been reported on modulating memristance by their electron trapping and detrapping effect, respectively.
Kim's group used ZnO-GQDs as electron trapping material to prepare a memristor (Al/polystyrene (PS):ZnO-GQDs/Al), named as 1R. [103] The electron transport in 1R follows SCLC conduction mode (Figure 8a), and a temperature-independent behavior at ON state is presented by the temperature-variable Figure 7. a) I-V curves of the Au/CQDs-HfO 2−x /Au memristor. b) Schematic of the modulated LED with the structure of Au/CQDs-HfO 2−x /Au/p-GaN/n-ZnO/In. c) Electroluminescence spectra of the modulated LED. Reproduced with permission. [15] Copyright 2018, AIP. d) Schematic illustration of the preparation route of CNDs. e-g) Learning regulations of Hebbian STDP (e), symmetrical STDP (f), and visual STDP (g) based on the CDMs. Reproduced with permission. [99] Copyright 2022, Royal Society of Chemistry.

www.advelectronicmat.de
I-V measurement (Figure 8b). The activation energy (E a ) of 1.7 meV can be obtained according to the Arrhenius equation, meaning that the electrons flow away from the capture of GQDs and then transport through filamentary paths. In order to avoid the misreading issue caused by the crosstalk of neighboring cells, poly(3-hexylthiophene-2,5-diyl) (P3HT) was inserted into 1R, building a memristor with a vertically stacked one diode-one resistor (1D-1R) architecture (Figure 8c). At last, the 1D-1R array executed addressing test and the encoded letters "SIMRC" was read accurately (Figure 8d). Their work demonstrates an electronic bistable device using the electron trapping and detrapping effect of GQDs.
GQDs-hBN nanocomposite is endowed with a strong electron capturing effect by combining the large bandgap (5.9 eV) of hexagonal boron nitride (hBN) and the quantum confinement effect of GQDs. Therefore, GQDs-hBN can also modulate memristance by capturing and releasing electrons that can reduce Ag + to Ag conductive filaments. [25] The device was fabricated using the printing technology and packaged with an atomic Al 2 O 3 layer for the actual application, as shown in Figure 8e,f. c) Programmed current path of the 1D-1R device. d) Current histogram after reading process of letters "SIMRC." Reproduced with permission. [103] Copyright 2015, Elsevier. e) Schematic of the flexible Al/PVP-NCQDs/ITO memristor. f) Optical image of the device in bend state. g) I-V curves of the memristor array. h) Resistance of the device over a range of bending diameters. Reproduced with permission. [25] Copyright 2017, IOP.

www.advelectronicmat.de
The results exhibit the critical parameters of switching ratio (10 3 ), write/erase cycling (≈1000), and retention time (10 5 ), taking the edge off crosstalk between adjacent memory cells (Figure 8g,h). This work illustrates that the GQDs-hBN nanocomposite possesses the modulated effect on memristance and potential in flexible and rewritable memristors.
Xu et al. fabricated an analog-type memristor (Ag/Al 2 O 3 / GQDs/Al 2 O 3 /ITO) with a high device yield (>95%) using the atomic layer deposition (ALD) technology. [71] The electron capture process is in charge of GQDs efficiently due to the uniform distribution of GQDs. In this device, GQDs are sandwiched and confined by Al 2 O 3 films, and the sickness of the Al 2 O 3 /GQDs/Al 2 O 3 stacked layer is less than 10 nm. Therefore, F-N tunneling is considered as the main electron transport mode in this device. The gradual changes in conductance endow the device with essential synaptic behaviors of LTP and PPF. Their research utilizes electron trapping and detrapping effect of GQDs, which provides a new method for neuromorphic computing.
As hollow GQDs, graphene nanorings (GNRs) in the Pt/ GNRs-P4VP/ITO/PET memristor modulate memristance by electron trapping and detrapping effect. [46] Typical bipolar resistive switching behavior shows a switching current ratio of 2 × 10 4 in Figure 9a. The electron transport mode of HRS is in keeping with the classic SCLC (Figure 9b). In comparison, LRS is dominated by Ohmic conduction mode (Figure 9c). Carbon conductive filaments consist of the graphitic clusters of sp 2hybridized carbon atoms in GNRs-P4VP, as observed from scanning transmission electron microscopy. The research on the mechanism of the device (Figure 9f-i) illustrates that the electron trapping and detrapping effect coexists with carbon conductive filaments. In addition, small distribution fluctuations The Ag/GQDs:GO/ITO memristor realized bifunctional behaviors including resistive and magnetic switching. [100] The epoxy groups of GQDs can trap and release the injected electrons in different electric fields. This process can trigger reversible oxygen anion migration and transfer the electron orbital hybridization from sp 2 (carbon filaments) to sp 3 (carbon domains), which leads to the resistive and magnetic switching bifunctional effect. Compared with the device without GQDs, SET and RESET voltages of the CDMs with GODs reduce sharply from −1.24 and +1.22 V to −0.43 and +0.43 V, respectively (Figure 10a). Besides, the CDMs have a high device yield of 89% in a 6 × 6 array and small resistance variation in cellto-cell, meaning a potential value in actual applications. X-ray photoemission spectroscopy (XPS) analyses indicate that the proportion of CO bonds decreases and the exposed oxygen vacancies account for 84.0% when the device enters LRS (Figure 10c,d), which also illustrates the carbon atoms surrounded by epoxy groups (COC) have more opportunities to present a stronger saturation magnetization when the device enters HRS (Figure 10b). This magnetically controllable memristor based on GQDs:GO hybrid film is expected to be applied in low-power and high-density spintronics.
The role of CQDs was investigated by Li's group using an electronic bistable memristor (Figure 10e) based on ZnO nanoarrays (NRs) and CQDs composite. [104] The cross-sectional image and energy dispersive spectrometer (EDS) in Figure 10f demonstrate the correct structure of the composite film. The SCLC conduction mode of the device based on composite is deduced from the current density versus voltage (J-V) curves (Figure 10g), suggesting that CQDs can capture the injected electrons. The trapped electron densities were calculated from capacitance-voltage (C-V) curves (Figure 10h), finding the electron densities (5.63 × 10 11 cm −2 ) of the device with CQDs are larger than that of the device without CQDs (4.09 × 10 10 cm −2 ). Likewise, the trapped electron densities of Pd/Al 2 O 3 /ZnO/ GOQDs/ZnO/SiO 2 /p-Si and Pd/SiO 2 /ZHO/GOQDs/SiO 2 /Si are severally 1.21 × 10 14 and 1.48 × 10 12 cm −2 , which are larger than the values of the corresponding devices without GOQDs, as reported by Wang et al. [72] and Jia et al., [26] respectively. These works verify that CQDs and GQDs have a significant electron trapping effect.
Subsequently, Huang's group combined the degradable composite film consisting of CQDs and degradable polyvinylpyrrolidone (PVP) with soluble silver nanowires (AgNWs) to construct a flexible memristor of AgNWs/CQDs-PVP/ AgNWs/PVP (Figure 11a). [105] The device is supplied with reversible resistive switching behavior by CQDs, as shown in Figure 11b. The symmetrical I-V curves reflect a dynamic random access memory storage feature, and electron transport mode is considered to be SCLC caused by the electron trapping and detrapping effect of CQDs. In order to assess the potential feasibility of the device for practical application, the currents of ON and OFF states were measured during the bending tests 100 times. A steady and high switching current ratio (10 5 ) signifies the negligible fluctuations of memristance, suggesting the excellent practicability of the device (Figure 11c). This device can be disintegrated in water (Figure 11d), indicating that it can be used as a transient memory for implanted electronics or data storage security.
To develop analog-type memristive synapses with excellent performance, Zeng et al. used NCQDs to prepare a memristor (Al/PVP-NCQDs/ITO/PET) and researched the modulation effect of NCQDs on memristance (Figure 12a). [37] The DRS behavior is presented at a low NCQDs concentration of 10 wt% (Figure 12b), while resistive switching behavior is turned to ARS at the concentration of 40 wt% (Figure 12c,d), which can be explained by electron trapping and detrapping effect relying on the concentration of NCQDs. Specifically, the trapping centers of NCQDs in high concentration tend to form multiple conductive filaments owing to abundant trapped defects, and the conductance can be modulated continuously by regulating filament number. The transition from STP to LTP by adjusting pulse amplitude and frequency (Figure 12f,g) illustrates that memory retention is enhanced. The STDP emulation (Figure 12h) further affords a potential feasibility of realizing artificial synapses. In addition, the flexible device also possesses excellent conductance stability against mechanical strain (Figure 12e). This work provides an analog-type memristor with high feasibility for ANN and proves NCQDs can modulate memristance by electron trapping and detrapping effect.
Meng et al. fabricated a memristor (Ag/CNDs-PVA/ITO) where CNDs act as electron trapping centers. [70] The abundant surface functional groups trap and release electrons effectively, resulting in their device presenting bistable resistive switching behavior and a large switching current ratio (10 5 ). In addition, the device was explored to emulate the data-programmed behavior during write-read-erase voltage cycles. The stable programmed states reach up to 10 4 cycles under a reading pulse of 0.2 V. In the exploration of CDMs, this work explained resistive switching behaviors using the electron trapping and detrapping effect of CNDs, supporting the modulating effect of CNDs on memristance.
It can be learned from the above discussions that the electron trapping and detrapping effect of CDs depend on surface defects or energy level potential wells, which can realize the memristance modulation. Although the threshold voltages of memristors based on this effect are usually larger than other CDMs attributed to the process of trapping the injected electrons, these values are still within the normal limits (Table 1). Moreover, these slightly larger threshold voltages are also conducive to protecting devices from breakdown caused by an excessive voltage applied by mistake.

Photosensitization Effect of CDs
Some photosensitive CDs, with super light-trapping and high optical absorption ability, may modulate memristance when they are exposed to light, especially high-energy UV light. At present, N-doped CDs are mainly applied to photosensitive memristors, such as NCQDs and NGOQDs. The N element can endow CDs with the photocatalytic property by separating photogenerated electron-hole pairs, resulting in more electrons that can shorten electron tunneling distance and form more controllable conductive filaments to modulate memristance. [106] www.advelectronicmat.de Figure 10. a) I-V curves of the memristor with structure of Ag/GQDs:GO/ITO. Insert: I-V curves of the memristor without GQDs. b) Magnetic hysteresis curves of GQDs:GO hybrid film and GO film. c) C 1s XPS spectra and d) O 1s XPS spectra of GQDs:GO hybrid film at HRS and LRS. Reproduced with permission. [100] Copyright 2021, Elsevier. e) Schematic of the Al/ZnO NRs (CQDs)/ITO device. f) The sectional view of the composite film and line-scan EDS (the inset) of the hybrid film. g) The J-V curves in double logarithmic scale and h) the normalized high-frequency C-V curves of two devices. Reproduced with permission. [104] Copyright 2016, Elsevier.

www.advelectronicmat.de
In 2020, Lin et al. demonstrated a photo-tunable memristor (Figure 13a) using the NCQDs-PVP composite. [106] The SET voltage of only 0.5 V is required for the device treated with UV light, less than the 1.4 V of the device without UV irradiation (Figure 13b), which is beneficial to reduce energy consumption and optimize memristance property. Besides, forming process, an operation of applying a large activation voltage, is not required after UV irradiation, which is helpful to restrain the current overshoot. They proposed a modulation mechanism to explain these phenomena (Figure 13c,d). A photocatalytic redox process is triggered when NCQDs-PVP composite film is irradiated by UV light. The holes react with the surface-absorbed H 2 O, generating oxygen and protons. At the same time, the photogenerated electrons gather near the NCQDs to form the local conductive amorphous carbon region that can shorten the tunneling distance between NCQDs, causing lower SET voltages and eliminating the forming process. The 9 × 9 array composed of 81 individual memristor units was prepared for encrypted image storage, as shown in Figure 13e-g. Taking advantage of the characteristic that different threshold voltages can be obtained under different UV radiation time, the encrypted information can be cracked using different pulses. This research uses the photosensitization effect of NCQDs to connect information storage and encryption functions, exhibiting a promising potential for optoelectronic neuromorphic computing systems.
In their another research, the flexible all-carbon memristor reduced graphene oxide (RGO)/GO-NCQDs/graphene realized the transition from DRS to ARS under UV irradiation. [16] The formation and rupture of conductive filaments (sp 2 RGO) inside GO generate a DRS behavior. Under UV light, GO can be reduced to local RGO domains near the NCQDs. The connections of these domains are equivalent to multiple weak conductive filaments that preferentially grow around the RGO domains, which finally leads to an ARS behavior. Besides, this ARS-type memristor exhibits much higher linearity in synaptic weight changes than that of the DRS-type device. On this basis, the ARS-type memristor was used for the pattern recognition application and presents a high accuracy. Their research proposes a model of weak conductive filament to explain the memristance modulation from DRS to ARS, providing a mechanism for CDMs, and the exploration of the ANN application has also promoted the development of CDMs.
In the same year, Ali et al. researched photosensitivity modulation of memristance and synaptic mimicry using N-doped GOQDs (NGOQDs). [107] The fabricated memristor (Ag/ NGOQDs/Pt) exhibits extremely low SET voltages (0.141 V) and high switching current ratio (10 6 ). It suggests reliable memristive stability of different cycles (Figure 14a). However, the current fluctuates acutely once excited by UV light and returns to its original state gradually after removing the irradiation (Figure 14b), which suggests that the memristor based www.advelectronicmat.de on NGOQDs is light tunable. NGOQDs can separate electronhole pairs when excited by UV light. The photo-excited electrons can trigger Ag filaments formation from Ag ions, which increases the conductivity of the device (Figure 14c). They think UV light is analogous to heterosynaptic neuro-modulatory axon and capable of modulating artificial synaptic plasticity. The device was further used to mimic synaptic plasticity behavior of the transition from STP to LTP (Figure 14d) by assisted UV illumination. This work proposes an optical field-driven memristor with controllable memristance behavior using NGOQDs, which provides an applicable model for hardware-based ANN strategies.
Wang et al. also found that UV light has an essential effect for GQDs on modulating memristance through studying the memristor (Figure 14e) based on the composite consisting of GQDs and multiwalled carbon nanotubes (MWCNTs). [108] The maximum switching current ratio fell off sharply from 4.58 × 10 3 to 10.77 with the illumination time increasing to 30 min (Figure 14f). The reason may be ascribed to the decrease of oxygen-containing groups and the increase of sp 2 carbon atoms, reducing the trap concentration and switching current ratio. Moreover, the switching current ratio of the device increases to 3.3 × 10 3 when mass ratio of GQDs reaches up to 50% ( Figure 14g). The cumulative probability distributions of the resistance with different mass ratios (Figure 14h) show that the device with doped GQDs possesses resistance uniformity.
In short, these works provide ideas and references for the research of photosensitive memristors. It can be known that UV excitation can provide an effective method to modulate memristance for photosensitization CDMs. The combination of light and CDMs will boost the building of the visual cognitive system to realize interactive AI systems.

Other Modulation Effects
Except for the above-mentioned three modulation effects, CDs can also modulate the memristance property by acting as an ionic conductor, [11] increasing the oxygen migration barrier, [38] promoting electron tunneling, [89] acting as the oxygen reservoir, [89] enhancing the interfacial polarization electric field, [101] and limiting ion migration. [109] Sokolov et al. discovered that NGOQDs in the Ag/NGOQDs/Pt memristor act as ionic conductor. [11] The porous stacking of NGOQDs leaves the spaces, and the surface functional groups such as NH, OHCO, CN, and COC cause Ag + The transition from STP to LTP by adjusting pulse amplitude (f) and pulse frequency (g). h) Simulative STDP behavior. Reproduced with permission. [37] Copyright 2021, Royal Society of Chemistry.
precipitation, both of which are favorable for forming Ag conductive filaments (Figure 15a), finally inducing the conduction of memristor.
Using the Al/OCQDs-GO/ITO memristor, Qi et al. researched the effect of oxygen migration energy barrier (E a ) on optimizing the retention time. [38] Compared with the Al/GO/ITO device, the OCQDs-doped device possesses a robust LRS with longer retention times (10 4 s). The sp 2 carbon filaments of the device without OCQDs are fragile when suffering thermal perturbation, which neutralizes the retention time. The Monte Carlo simulation and Meyer-Neldel effect indicate that more COC groups from embedded OCQDs can cause E a rising Figure 13. a) Schematic of the Al/NCQDs-PVP/ITO/PET memristor. b) Forming process and I-V curves of the device without/with UV irradiation. c,d) Schematic of the working mechanism of the device without (c) and with (d) UV irradiation. e) Schematic of UV light irradiation writing encrypted information. f) Different read pulses were applied to the encrypted information. g) The read conductivity value was used for information decryption. Reproduced with permission. [106] Copyright 2020, Royal Society of Chemistry.
( Figure 15b,c), which can decrease the recombination probability of oxygen vacancies and oxygen ions, resulting in prolonged retention time.
Zhao's group introduced GQDs into FeO x -based memristor (Pt/Ti/GQDs/FeO x /SiO 2 /Pt/Ti/SiO 2 /Si) for the further exploration of GQDs. [89] For this device, the threshold voltages decrease by about 40% and the coefficient of variation reduces by nearly 85% compared with the control memristor without GQDs. Two memristance modulation effects were proposed. On the one hand, GQDs can alter the energy band alignment and boost the electrons tunneling, significantly reducing threshold voltage and tightening distribution variation (Figure 15d). On the other The interfacial polarization field enhanced by CDs also influences memristance. Compared with the memristor without CDs, the Al/PMMA/CDs/PEDOT:PSS/ITO device exhibits ferroelectric-like hysteretic I-V curves and a larger memory window. [101] The hysteresis phenomenon is caused by the polarization field generated at the interface of PEDOT:PSS and PMMA under the external electric field. The polarization field still exists even though removing the external electric field (Figure 15f). After inserting CDs, the -OH and -NH groups on the surface of CDs combine with PEDOT + in PEDOT:PSS and -OCH 3 of PMMA, respectively, which increase the polarization field and then enlarge the memory window of CDMs (Figure 15g).
Yan's group achieved the transition from DRS to ARS by regulating the concentrations of GOQDs in Ag/ZHO/GOQDs/ ZHO/Pt memristor. [109] Ag + migration is adjusted by the concentrations of GOQDs, and further the memristance can be modulated by controlling the formation speed of Ag conductive filaments.
The above studies propose some effects of diversified CDs on modulating memristance from different perspectives, such as facilitating ionic migration, increasing oxygen migration barrier, promoting electron tunneling, reserving oxygen, and enhancing the interfacial polarization electric field, which enrich the modulation effects of CDs on memristance and provide references for memristors using other functional layer materials.

Conclusions and Perspectives
CDMs are a potential candidate for improving computational power and storage capacity in the Big Data and AI epoch, and CDs play a pivotal role in CDMs. This review first introduced the classification and mechanisms of memristor. Memristors can be classified into digital-type devices for information storage and analog-type devices for neuromorphic computing. The working mechanisms of memristors were also retrospected, mainly including conductive filament mechanism, electron trapping and detrapping mechanism, phase change mechanism, interface modulation mechanism, and charge transfer mechanism. Second, this review discussed some common functional layer materials, such as metal oxides, perovskites, solid electrolytes, organic materials, and carbon materials. In these materials, CDs, a kind of carbon material, are increasingly concerned due to their physicochemical stability, excellent solution processing, high conductivity, and biocompatibility. More importantly, quantum confinement effect and Coulomb blocking effect of CDs are able to refine memristance property by adjusting energy level alignment and electron movement. The internal carbon skeleton, conjugated Figure 15. a) Schematic of the role of NGOQDs in memristor. Reproduced with permission. [11] Copyright 2019, Wiley-VCH. Schematic of E a for: b) GO and c) OCQDs. Reproduced with permission. [38] Copyright 2018, Royal Society of Chemistry. d) Schematic of the energy band of the memristor after doping GQDs. e) Schematic of GQDs as oxygen reservoirs forming oxygen vacancy conductive filaments. Reproduced with permission. [89] Copyright 2017, Wiley-VCH. f,g) Schematic of polarization field of memristors without (f) and with (g) CDs after removing external electric field. Reproduced with permission. [101] Copyright 2016, Royal Society of Chemistry.
www.advelectronicmat.de π-domain, surface functional group, and edge carbene radical of different CDs (GQDs, CQDs, CNDs, and CPDs) can influence memristor performances by affecting the electron transmission process. Moreover, a large-scale integrated array based on CDMs is expected to be realized when considering some emerging integration and fabrication technologies. Finally, the modulation effects of CDs on memristance were highlighted in detail, mainly including LEF enhancement effect, electron trapping and detrapping effect, photosensitization effect, and others. From the reported literatures, both LEF enhancement effect and electron trapping and detrapping effect were reported in the memristors based on GQDs, CQDs, and CNDs, which may be related to the fact that these CDs have high carbonization degrees than CPDs. LEF enhancement effect facilitates the formation of more concentrated filaments, which is helpful to improve memristance stability. Electron trapping and detrapping effect usually endows CDMs with memristive phenomena and protect the devices from breakdown. Photosensitization effect mainly exists in the N-doped CDs that can shorten electron tunneling distance and form more controllable conductive filaments by a light tunable process. This effect also provides a novel solution for visual cognitive system. Through these effects, CDMs usually exhibit a stable memristance property, reasonable threshold voltage, long retention time, high endurance, and even the transition of memristive switching between DRS and ARS. Accompanying with the memristance modulation effects of CDs and advantages of CDMs, the corresponding applications have also been explored, including information storage, encrypted image storage, artificial synapse, digit recognition, and flexible wearable devices. Looking back to these works on CDMs over the last decade, CDMs will certainly have excellent potential in information computing and memory.
Even so, the research on CDMs is still in its initial stage. Herein, we outline several challenges and perspectives of CDMs for further developing CDMs and provide some suggestions for researchers in related disciplines. i) Although the modulation effects of CDs on memristance have been summarized in this review, the structure-property relationships between the compositions of CDs and memristive performances of CDMs are still obscure. For instance, the consensus on how doped elements (e.g., B, N, O) and surface functional groups of CDs influence the memristive behaviors of CDMs still needs to be explored in depth. Future research should focus on the design and synthesis of CDs with diverse structures and explore their roles in memristors, which will help to establish systematic structure-property relationships. ii) Deep insight into the interactions between CDs and matrix materials is absent. The exploration on these interactions is helpful to realize the in-sensor computing systems integrating the sensing, storage, and computing functions. For example, the memristors fabricated from the composites containing CDs that sense light, pressure, and chemical substances, can be regarded as multimode sensation devices with visual, tactile, olfactory, and auditory. It helps to construct interactive AI systems and processes information from complicated environments efficiently. iii) The research on the interface effects of CDMs is still insufficient. The CDs with the nanoscale structure may cause many interesting phenomena at the interfaces in memristors, e.g., CDs can act as an oxygen reservoir at the interface with the top electrode.
It is helpful for figuring out the role of CDs at the interfaces in CDMs. iv) The coupled electronic and ionic dynamics of memristors, including CDMs, have complex dynamics that urgently need reliable microscopic characterization methods. However, the lack of these characterization methods has caused the severe consequence that different mechanisms were used to explain the devices with the same structure. The appropriate solvent is probably taking full advantage of existing elaborate equipment and developing novel characterization instruments. v) Currently, the research on CDMs, including the amount of data, learning tasks, and synaptic plasticity and stability of memristance, are still far from real intelligent devices. Exploration on CDMs from other aspects needs to be devoted for further development. On the algorithm level, it is necessary to emulate the cognitive model of the brain and create the corresponding learning algorithm for the impulse neural network. On the architecture level, it is indispensable for building a system platform with reliable versatility and reconfigurability in multiple scenarios. On the circuit level, more efficient digital-to-analog converters/analog-to-digital converters are needed to support neuromorphic computing systems with high energy efficiency.
In conclusion, CDs are beneficial to optimize and modulate memristance, and CDMs are a potential device for intelligent memory and computing. The research on CDMs may trigger a new round of circuit revolutions. As the physicist Stanley Williams from Hewlett-Packard labs said, the brutally hard work is just beginning, and now is the most exciting time to be working in the fields of electronics and computing since the days of Shannon, Turing, Bardeen, and von Neumann. [6] www.advelectronicmat.de Haotian Hao is currently pursuing his Ph.D. degree under the supervision of Prof. Yongzhen Yang at the MOE Key Laboratory of Interface Science and Engineering in Advanced Materials, Taiyuan University of Technology, China. His research interests focus on the synthesis of carbon nanomaterials and the fabrication of electronic and spintronic devices in applications for functional memory devices and neuromorphic electronics.
Yongzhen Yang is a full professor at the Taiyuan University of Technology and a part-time doctoral supervisor at the University of Hertfordshire. She received her Ph.D. degree at the Taiyuan University of Technology in 2007. Her work in interdisciplinary research includes carbon nanomaterials for light-emitting diodes, photovoltaics, memristor devices, and other applications. She has authored more than 370 scientific papers in journals and conferences. Now she is leading a research group in exploring carbon materials for advanced electronic devices and promoting related fundamental research toward practical applications.