Intrinsic memristive mechanisms in 2D layered materials for high-performance memory

Modern electronics is developing toward higher intelligence, lighter weight, and better portability but the efficient storage, transmission, and processing of massive data have already become some of the major challenges. With the rapid rise of applications such as big data technologies, cloud computation, and artificial intelligence, the size of data is explosively growing, which has highlighted the requirement for high-performance memory devices with lower power consumption, faster write/read speed, and smaller device size (higher memory density).¹ More importantly, modern computer systems are based on von Neumann architecture that the data storage is separated from the processing unit. The famous von Neumann bottleneck, originated from the enormous power and time consumptions for data transmission, has already become a major bottleneck for the development of advanced data-processing technology.² The non-volatile flash memory in solid state disks, whose areal density is doubling every two years, is based on electrically erasable programmable read-only memory (EEPROM) technology and suffers from slower read/write speed. As a comparison, quick data processing relies on dynamic random access memory that is not only electrically volatile but also power-consuming.³ Therefore, various high-performance non-volatile random access memories (NVRAMs) are under extensive studies, including ferroelectric random access memory (FeRAM), magnetoresistive random access memory, and phase-change random access memory (PCRAM). Another strategy to “bypass” the von Neumann bottleneck is to mimic the neuro-biological architecture and develop artificial neural network and the neuromorphic computing system.⁴ By utilizing the resistive switching characteristics in different functional materials, artificial synaptic devices are fabricated to simulate the behavior of human synapses, where memory and processing are integrated for lower power and faster speed.^1,2

In the past two decades, two-dimensional (2D) layered materials have attracted a wide range of interests owing to their unique electronic and opto-electronic properties.^5–7 Inorganic 2D layered materials can be classified into four major species, including MXenes (ultrathin transition metal carbides and nitrides), Xene (ultrathin monoelemental materials), TMDCs (transition metal dichalcogenides), and nitrides. Various 2D materials exhibit ample performance diversity,⁸ such as graphene and black phosphorus with high carrier mobility,⁹ transition metal dichalcogenides (TMDCs) with tunable semiconducting band structures,¹⁰ and hexagonal boron nitride (h-BN) with a large bandgap (6 eV).¹¹ Moreover, considered as a promising successor for silicon electronic to maintain Moore's law,¹² 2D materials are prospective for futuristic data storage and high-speed computing because of their potential in miniaturization, flexible and wearable electronics, high-density storage, ultra-fast response, and high on/off ratio.¹³ More importantly, the unique van der Waals interface allows the integration of different 2D materials without lattice mismatch issues, which provides a new material platform for the design of high-performance memory devices.

First, extensive efforts have been made to the integration of 2D materials with traditional functional materials for high-performance memory devices, especially for FeRAM and synaptic memristors. For example, Lee et al. used 200-nm thick poly[(vinylidenefluoride-co-trifluoroethylene) (P(VDF-TrFE)] as ferroelectric gate dielectric to modify the conductivity of a single-layered MoS₂ transistor. The proposed ferroelectric field effect transistor (FeFET) demonstrated high on/off ratio of 10⁵ and excellent retention performance.¹⁸ Zhang et al. has dispersed MoS₂ strips into polyvinylpyrrolidone (PVP), and the MoS₂-PVP membrane demonstrated excellent resistive switching performance thanks to the charge capture and de-capture response of MoS₂. However, the integration of 2D materials with other functional materials typically sacrifices the uniqueness of 2D materials, especially their flexibility and atomically thin nature. To pursue the next-generation high-density storage based on 2D materials, the intrinsic memristive mechanisms of 2D materials should be studied so that they can replace the current function layer materials. Recently, intrinsic ferroelectric polarization switching and resistive switching are reported for various 2D materials, which follow the mechanisms for FeFET and memristors, respectively. For example, by replacing P(VDF-TrFE) with 2D CuInP₂S₆ thin film, the MoS₂ FeFET demonstrated an on/off ratio over 10⁴.¹⁹ Herein, a comprehensive review of the physical mechanisms (as shown in Fig. 1) for the polarization switching and resistive switching in 2D materials is reported. To provide a futuristic insight into this new research field, the results will be discussed in view of their potentials and challenges.

Ferroelectricity (FE) is a property of certain materials that have a spontaneous electric polarization that can be reversed by applying an external electric field, which can be memorized for years.^20,21 Ferroelectric materials are, therefore, widely adapted for non-volatile memory devices, especially for FeRAM.²² FeRAMs usually have lower power consumption, higher endurance, and faster writing speed in contrast with other non-volatile memory devices. However, a major challenge for FeRAMs is the low storage density and, thus, the high cost. The past decade has witnessed the explosive development of big data science and technology, which further raises the requirement for low-cost high-density storage. A critical issue for reducing the size of FeRAM is the size effect that when the thickness of a traditional ferroelectric thin film is reduced to nanometer-scale, the spontaneous polarization of the material will disappear because of the depolarization electrostatic field arising from imperfect charge screening and the dead layer effect.^23–27 The dead layer effect results from the non-ideal surface structure and the consequent interfacial strain and reduction of dielectric constant.^28–31 However, the unique van der Waals surface, which has in-plane integrity and chemical stability (free of dangling-bonds), allows 2D layered materials to overcome the dead layer effect. By considering the variety of 2D materials and 2D ferroelectric mechanisms, more operational mechanisms and device configurations are to be envisaged. For example, it is reported that the unique non-uniaxial polarization and the consequent dipole-locking in In₂Se₃ can yield resistance to the out-of-plane (OOP) depolarization field.³² More inspiringly, 2D materials usually have the above-RT Curie temperatures (as shown in Table I), which might be benefited from their in-plane integrity and the consequent resistance to the structural distortion of a single ferroelectric unit cell, enabling themselves to be adopted for electronic applications under various scenarios. Therefore, research efforts are turning to 2D materials for futuristic ferroelectric thin film materials.

As FeRAMs typically rely on the OOP polarization of the ferroelectric functional layer, only the OOP ferroelectricity in 2D materials is discussed in this section. OOP ferroelectricity in 2D materials can also provide non-linear dielectric response for better electrical switching capability,^28,33 which makes 2D ferroelectric materials excellent passive gate dielectric materials, whose spontaneous polarization induced electric field can effectively modulate the channel current. Moreover, several 2D ferroelectric materials have demonstrated a correlation between its OOP polarization and in-plane (IP) electrical conductivity,^32,34,35 which, therefore, can be independent of converting its polarization signal into electrical signal without the need of an additional channel material. The resultant ferroelectric semiconductor FeFET typically has simpler device structure, higher current on-off ratio, and lower operation voltage.³⁶ 

Intrinsic room-temperature (RT) ferroelectricity was theoretically predicted in several 2D layered materials, including 1T-MoS₂,^37–39 phosphorene,⁴⁰ In₂Se₃,⁴¹ and bilayer h-BN.⁴² Room-temperature 2D FE was first discovered in 2D CuInP₂S₆ (CIPS) with a transition temperature around 320 K,³³ although the 2D ferroelectricity was experimentally observed in atomically thick SnTe at low temperature in 2015.⁴³ Later, vertical piezoelectricity and FE in indium selenide (In₂Se₃) were verified in 2017.⁴⁴ Recently, two- or three-layer tungsten ditelluride (WTe₂) is reported to exhibit spontaneous OOP electric polarization that can be switched by electrostatic gating.⁴⁵ In this part, we present a review on the experimentally proven intrinsic mechanisms for 2D polarization switching so as to provide an insight and guidance for the futuristic development of 2D ferroelectric materials.

CIPS is the first reported experimentally proven 2D layered material with OOP ferroelectricity. CIPS is a member of the transition metal thiophosphate (TMTP) family, which typically has the form of M₁M₂P₂X₆, where M₁ and M₂ are two transition metal ions with total valence of +4, P is phosphorene, and X is a chalcogen (S, Se, Te, etc.). The room-temperature OOP ferroelectricity of the CIPS thin film with a thickness of 50–100 nm was first observed with PFM by Belianinov et al.⁴⁶ Later in 2016, Liu et al. reported the observation of room-temperature ferroelectricity in a 2D CIPS with a thickness down to 4 nm.⁴⁷ Odd–even effect was not observed in CIPS typically because CIPS has a non-centrosymmetric lattice structure.

In 2D layered materials, the unique van der Waals layer structure divides the crystal space into multiple interfaces and spaces, which, therefore, could retain more stable ferroelectric states for ionic displacement in the OOP direction,^51,52 compared to traditional 3D ferroelectric materials. A detailed DFT calculation of the energy minima in CIPS was reported by John et al., indicating that the copper ions can not only have intralayer OOP displacement but also move into the van der Waals gaps of the materials. The switching of the four ferroelectric (two intralayer and two interlayer) states is achieved by applying external field and in-plane strain while the four quantum wells could exist simultaneously when the lattice constant is c = 13.09 A, as shown in Figs. 2(a)–2(c). As the interlayer states require higher energy and larger displacement compared to the intralayer states, they typically have much higher polarization and the Curie temperature.⁴⁸ 

Interestingly, although the vertical polarization switching usually originates from the vertical inversion of the local dipoles (the vertical offset of copper ions in CIPS), they can also have in-plane movement under an in-plane electric field, which leads to the local reversal/retention of polarization and a macroscopic motion of ferroelectric domains [Figs. 2(d) and 2(e)]. Despite of the consequent instability of domains, these movable ions/dipoles have also introduced intriguing editability to the domains' location, size, and quantity of dipoles. Notably, Xu et al. replaced the traditional Au/Cr electrodes with copper electrodes, which injected a large amount of copper ions into the channel to enhance its ionic conductivity. On the contrary, copper ions will decrease under an in-plane electric field because they were reduced to copper atoms at the cathode.⁴⁹ In a word, the intrinsic ionic conductivity in CIPS could possibly lead to the effective modulation of its OOP ferroelectricity.

Recently, a thorough study of the ferroelectric ionic behavior in CIPS was conducted by Zhou et al.⁵⁰ Like traditional ferroelectric materials, polarization switching in CIPS cannot be achieved for temperature above the Curie temperature because of the deteriorated ferroelectricity. Neither is it achievable for temperature below a certain temperature necessary for the conduction of Cu ions. Also, the polarization of the local dipoles is pinned by the defect-dipoles (a small number of Cu ions are trapped in deep-level non-equilibrium states), which typically coupled to the matrix polarization by the symmetry-conforming property. As a result, the ferroelectric switching of CIPS typically has a slow switching speed and ultra-long retention times compared to other traditional ferroelectric materials [Fig. 2(f)].⁵⁰ 

A unique polarization switching mechanism in 2D layered materials is the interlayer charge redistribution driven by the interlayer potential variation when an in-plane interlayer translation happens. With first principle calculation, Li and Wu predicted in 2018 that 2D OOP ferroelectricity widely exists in AB stacking (hexagonal lattice) of both bilayer 2D crystals and binary compounds with similar lattice constant.⁴² The authors took graphene/h-BN heterobilayer as an example [Fig. 3(a)]. The calculated potential difference between the AB1 stacking (B under C, N under hexagon center) and AB2 stacking (N under C, B under hexagon center) is 0.21 V, while the corresponding interlayer translation distance is only one-bond length [Fig. 3(b)].⁴² Although Li and Wu have predicted the OOP ferroelectricity for more than 10 hexagonal materials (h-BN, graphene/h-BN, AlN, ZnO, MoS₂, GaSe, MXene, Cr₂NO₂, VS₂, MoN₂, etc.), experimentally proven OOP ferroelectricity from interlayer translation was first reported in bilayer and trilayer of 1T′ WTe₂, a metal phase TMDC with orthorhombic lattice, by using graphene as an electric-field sensor [Fig. 3(d)].⁴⁵ The strong ferroelectricity in WTe₂ typically benefits from its low crystal symmetry. On one hand, WTe₂ has only one symmetric plane, which, therefore, only allows translation in one direction, resulting in high consistency over the whole domain [Fig. 3(c)]. On the other hand, the non-centrosymmetric crystal structure of few-layer WTe₂ also gives rise to its ferroelectric polarization. Notably, different from the traditional odd–even effect, monolayer WTe₂ demonstrates no ferroelectricity, which further proves the causality between its ferroelectricity and interlayer correlation.

Systematic theoretical studies of the relation between the interlayer sliding and the OOP ferroelectricity in bilayer WTe₂ were reported by several research groups.^14,53,54 Interestingly, the vertical mirror reflection of its crystal structure can be realized by the horizontal movement of the top layer by 0.72 Å and the native polarization can, therefore, be switched [Fig. 3(e)]. Although the ferroelectric switching barrier is only ∼0.6 meV per unit cell, the material demonstrated the above-RT Curie temperature because of its in-plane rigidity.¹⁴ However, because of the screening effect from WTe₂’s native semi-metallic nature, only the outmost two layers mostly contribute to its ferroelectricity.⁵⁴ 

A major challenge for the observation and application of these ferroelectric materials is the precise control of the crystalline orientation for both layers. A small angle twist within the bilayer will give rise to the formation of a Moiré superlattice pattern,¹⁴ where the relative interlayer position varies from region to region, inducing different polarization domains that will counteract each other. Also, the monocrystalline ferroelectric flakes are usually preferred because the shear forces from the non-uniaxial motions of different crystalline domains tend to counteract each other at the grain boundaries of polycrystalline samples.^55,56 A feasible tactic to deal with these two issues is to develop robust and facile growth techniques that produce high-crystalline samples.

The low energy barrier for the phase transition in 2D layered materials allows the effective electrical modulation of their fundamental electrical and chemical properties.⁵⁷ For example, due to the different atomic arrangements, TMDCs have both the semiconducting phase (the 2H-phase) and the metallic phase (the 1T-phase and the 1T′ phase) and the transition between these phases can be achieved via electrostatic doping.⁵⁸ However, the projections of the atomic positions in these phases to the OOP direction are centrosymmetric, which, therefore, present no OOP ferroelectricity.

Ding et al. predicted the room-temperature ferroelectricity in In₂Se₃ and other III–VI van der Waals materials.⁴¹ The OOP polarization switching typically originates from the phase transition of the materials between two non-centrosymmetric and energetically degenerative ferroelectric-zincblende (FE-ZB′) structures with opposite polarization, having a displacement of the central Se atom and the reconstruction of its chemical bonds with the neighboring In atoms. Notably, the activation barrier for the direct transition of the central Se atoms is as high as 0.85 eV per unit cell. However, the energy barrier for an alternative process is only 0.066 eV, which is a three-step motion involving two intermedium phases. As depicted in Fig. 4(a), the In₂Se₃ will be first transformed from the α-phase FE-ZB′ structure to a metastable β-phase fcc′ structure, and then the central Se atom rotates and degenerates into a fcc′ structure, finally reaching a stable FE-ZB′ structure with reverse polarization.⁴¹ 

In 2017, the experimental observation of the OOP ferroelectricity in the In₂Se₃ thin film (thickness ∼10 nm) was reported by Zhou et al.⁵⁹ The polarization switching is observed with PFM [Figs. 4(b) and 4(c)], and the ferroelectric FE-ZB′ crystal structure was verified by scanning transmission electron microscopy, second-harmonic generation, and Raman spectroscopies. It is worth mentioning that because the phase transition is typically an intralayer change, the ferroelectricity of In₂Se₃ declares no appreciable dependence on its thickness, which is different from the previously mentioned interlayer mechanisms. Furthermore, it is later reported that as the atomic motion of the central Se atoms is not parallel to the OOP direction, an IP ferroelectricity can be observed along with the OOP ferroelectricity while the OOP polarization is several tens-of-time weaker than the IP polarization. The IP polarization switching has an impact on the Schottky barrier height at the In₂Se₃/electrode interface and, therefore, will modulate the in-plane current. This unique intercorrelated IP and OOP ferroelectricity in In₂Se₃ is, therefore, highly promising for novel ferroelectric semiconductor field effect transistor (FeS-FET).³⁴ 

Since the discovery of ferroelectricity in In₂Se₃, many follow-up studies have been reported.^{32,34,59–64} Although room-temperature ferroelectricity was predicted in other III–VI van der Waals materials,⁴¹ related experimental research is barely reported, while it is also extensively expected. More interestingly, structural distortion induced OOP ferroelectricity was also reported in graphene oxide^65–68 and 1T-MoS₂.³⁹ 

To present the status of the development of OOP ferroelectric 2D materials, we have collected the properties (thickness, polarization, retention time, the Curie temperature, etc.) of the experimentally proven 2D ferroelectric materials in Table I. Encouragingly, they generally do possess the above-RT Curie temperatures and long retention times. However, the development is still at the preliminary stage since many other potential 2D ferroelectric materials have not been demonstrated. According to the summary by Guan et al., more than 60 articles were reported for theoretically predicting the ferroelectricity in more than 40 different 2D materials. Notably, the materials predicted by the three proven mechanisms are the most promising candidates. But there still exist several critical challenges ahead for the further development of 2D ferroelectric materials.

As shown in Table I, the precise measurement of remanent polarization for ultrathin materials is still a problematic topic. Currently, several techniques are involved for the observation of the polarization switching, including piezo-response force microscopy (to detect the piezoelectric effect) and second-harmonic generation measurement (to detect the non-linear optical process). However, the direct measurement of remanent polarization still relies on the traditional parallel plate capacitor method which is generally not applicable to ultrathin materials because of the large leakage and tunneling current. Fan et al. has adopted an indirect method that the potential drop in WTe₂ within the dielectric layer was evaluated by the current variation in the graphene channel in a graphene-FET.⁴⁵ Notably, the analysis of quantitative annular bright field imaging was recently adopted to precisely measure the polarization in ultrathin (∼0.6 nm) perovskite ferroelectric films.⁶⁹ Rather than the macroscopic measurement of total polarization, the polarization of a single ferroelectric domain can be evaluated by the precise measurement of the confined crystal structure and the variation of the distance between ions.⁷⁰ Moreover, attention should also be devoted to the interface quality and interfacial defects in 2D ferroelectric devices because the charge trapping at interfaces could counteract the capacitive modulation on the channel conductivity, especially at room temperature.⁷¹ 

Despite the explosive increase in transistor density, computing technology still relies on the traditional von Neumann architecture, where the computing and storage units are physically separated, and the repeating back-and-forth transmission of data has consumed a huge amount of time and power. The processing speed of a CPU is severely restricted by the low-efficient data transfer process, which is well known as von Neumann bottleneck and has become a critical challenge for the development of big data science and technology. In comparison, the human brain typically has moderate speed and low power consumption than computers when complex signals processed.^72,73 Therefore, extensive efforts have been dedicated to imitating the human–brain computation and developing new computing technologies based on neuromorphic engineering.

The key for mimicking brain-like computation is the development of high-performance artificial synapses. The transmission of an electrical signal between two adjacent neurons is controlled by a synapse, while the receptivity of the synapse determines the response of the post-synaptic neuron.⁷⁴ The sensitivity of a synapse typically increases with the increase in the historical sum of all the signals going through the synapse. Similarly, a memristor, which is a typical artificial synaptic device, has a variable resistance depending on the historical sum of the current through the device. Note that as the two-terminal memristors usually suffer from poor stability, reproducibility, controllability, and dynamic linearity, one also uses three-terminal synaptic transistors to imitate synapses.⁷⁵ However, the scaling of a synaptic transistor is restricted by the existence of gate dielectric/electrode, which is, therefore, not promising for high-density computing.

In this section, we discuss the four intrinsic mechanisms for the resistive switching in experimentally reported 2D synaptic materials. We also discuss the potentials of these mechanisms for artificial synapses in terms of different performance metrics, including dynamic linearity, dynamic range, number of distinguishable states, retention, stability, reproducibility, and endurance.

The formation of conductive filament (CF) by the aggregation of ions under the biased-electric field is a well-recognized mechanism for previously reported memristors while the functional layer is usually metal oxides such as HfO₂,^76,77 ZnO₂,⁷⁸ and ZrO₂.⁷⁹ Because of the widely observed excellent IP and OOP ionic conductivity in 2D layered materials,^80–82 various CF-type 2D memristors have been precedingly reported.^15,81–90 Depending on the source of ions, CFs are divided into two major categories, i.e., metal-ion CF and vacancy CF. The formation of metal-ion CF relies on the contribution of metal ions from an active metal electrode through redox reaction, while the metal ions are reduced to metal filaments when they reach the other inert metal electrode. Ag, Al, and Cu are usually adopted for the active electrode, while Au, Pt, and ITO can be adopted for the inert electrode. As a comparison, the formation of vacancy filament relies on the local vacancy defects and, therefore, has no specific requirements for electrode material. Zhao et al. reported that the two mechanisms can coexist in an Ag/MoS₂/Au memristor [Fig. 5(a)], the device, therefore, has two low-resistance states.⁸⁸ However, because of the higher migration barrier energy of the sulfur ions (1.13 eV) in MoS₂ compared to Ag (0.58 eV), the vacancy filament is dominating only when a negative voltage is applied to the active electrode [Figs. 5(b) and 5(c)].⁸⁸ 

The control of the formation and rupture of the CFs is crucial to improve the performance of CF-based synaptic devices. However, the previous CF memristor based on oxides typically suffers from the randomness of the ionic motion under an electric field and the consequent large device-to-device and cycle-to-cycle performance variation and low linearity.⁹⁴ In contrast, because of the ultrathin nature of the 2D materials, the top and bottom electrodes can be easily connect by the CF while the change of resistance is attributed to the change of cumulative area of the CF rather than the length [Fig. 5(d)]. Therefore, the corresponding stability and linearity are substantially improved, compared with the previous CF memristors [Figs. 5(e) and 5(f)].⁹² Moreover, the corresponding switching voltage is also substantially decreased, which is important for the simulation of low-power neuromorphic circuits.⁹² However, the ultrathin nature of 2D materials also induces large off-state current. The dynamic range of the reported CF-based 2D memristors are usually only around ∼100. Proper thickness engineering is necessary to balance the pros and cons of CF-based 2D memristors.

Notably, since the intrinsic “pore”-size of graphene is smaller than the van der Waals radii of the smallest atoms, graphene has excellent impermeability for the migration of molecules, atoms, and ions,^95,96 which makes graphene a perfect barrier material to modify the CFs.⁹⁷ For example, with the graphene blockade, a tunneling barrier could be inserted between the CF and the inert electrode. The change of conduction mechanism provides a new route for the enhancement of the on/off ratio of current.

Controlling grain boundaries (GBs) is critical for the optimization of 2D synaptic devices because the defect-abundant GBs can serve as predefined paths for the migration of electrons, atoms, and ions. Therefore, the manipulation of GB has potential in the guidance, restriction, and even filtration of particle motion. A GB memristor was first reported by Sangwan et al.⁹⁸ Whether a GB is connected to one electrode (intersecting-GB), connected to both electrodes (bridge-GB), or isolated from both electrodes (bisecting-GB), it can produce a resistive switching of the channel by its conduction for dopants. As shown in Fig. 6(a), the intersecting-GB device has the best synaptic performance (dynamic range over 10³).⁹⁸ The resistive switching in an intersecting-GB MoS₂ memristor originates from the electric field-driven migration of dopants between the GB and the depletion region around the electrode, while the dopants can substantially increase the channel conductivity when transported to the GBs. Similarly, a CF will form in a bridge-GB device. However, the highly conductive filament produces excessive heating and cause the consequent thermal rupture of itself. Therefore, a negative differential resistance (NDR) effect was observed after the current spike. What is more, the resistive switching in a bisecting-GB device originates from the dislocation motion of the sulfur atoms and the increase of sulfur vacancies under high electric field. The corresponding motion of the bisecting-GB was observed by atomic force microscopy (AFM) [Figs. 6(b) and 6(c)] and electrostatic force microscopy (EFM) measurement. Later, Sangwan et al. reported that the migration of vacancies within the GB could also induce resistive switching because the electric field induced aggregation of charged defects has a localized modulation to the adjacent channel–electrode Schottky junction.⁹⁹ 

Recently, a vertical 2D memristor based on the GBs in multilayer h-BN was reported. With a vertical electric field, boron ions are driven to the electrodes through the GBs [Fig. 6(c)]. The generation of boron vacancies boosted the migration of Ti ions into the GBs and the formation of a CF, the resistive switching of a 100 × 100 μm² device is shown in Fig. 6(e). Notably, the selective conduction of Ti ions is relied on the Ti/h-BN interaction while the conduction of noble metal (e.g., Au and Pt) ions is not allowed. On the contrary, the large diffusivity of Cu will cause severe contaminant to h-BN and cause irreversible breakdown.¹⁷ 

The electrical properties of 2D TMDCs can be effectively tuned by phase engineering.¹⁰⁰ Monolayer TMDCs typically has two common phases: the semiconducting 1H-phase and the metallic 1T-phase. The conductivity variation between the two phases is as high as 107,¹⁰¹ guaranteeing its potential in synaptic devices in terms of dynamic range. In addition, TMDCs have multiple distorted phases based on the 1H-phase and the 1T-phase, such as the 1T′-phase and the T_d-phase, allowing TMDC synaptic devices to have multiple conduction states.¹⁰² 

Extensive efforts have been devoted to the phase-selective synthesis of TMDCs while the post-synthesis phase transition is still very challenging. To achieve the post-synthesis phase transition in TMDCs, harsh conditions such as thermal treatment,^104,105 in-plane strain,¹⁰⁶ and metal-atom intercalation^107,108 are usually required. These conditions are unfavorable in electronic devices, while a reversible, electric field-induced localized phase transition is necessary for synaptic applications. Zhu et al. has reported a phase-transition 2D memristor by converting the electric field into the in-plane motion of the pre-intercalated Li⁺ ions in MoS₂, which benefits from the excellent ionic conductivity of 2D materials. As shown in Fig. 7(a), a positive voltage applied to electrode A drives the Li⁺ ions to electrode B, transforming the pristine 2H-phase MoS₂ into the 1T′-phase MoS₂, thereby increasing the total conductivity of the whole channel and vice versa.¹⁶ Although the reported prototype device only demonstrates a dynamic range of ∼5 and a high spike voltage (4 V), the freedom on the control of intercalated ions has provided various possibilities for the further optimization of the device performance.

A pure electric field-induced phase-transition memristor was reported by Zhang et al. as shown in Fig. 7(b). By applying a 2.3 V voltage on a 24 nm 2H MoTe₂ thin film, resistive switching with a current ratio of ∼50 was observed [Fig. 7(e)]. From high-angle annular dark field (HAADF) STEM images of the device cross section, the authors unveiled a new phase transition from the 2H-phase [Fig. 7(c)] to a distorted 2H_d-phase [Fig. 7(d)], which is responsible for the resistive switching. However, the 2H_d-phase is eventually unstable and will recur to 2H-phase according to the DFT calculations.¹⁰³ Similar pure electric field induced phase transition in MoTe₂ was reported by Kim et al. By applying a high electric field on MoTe₂, the 2H-phase MoTe₂ can be transformed into the 1T′-phase.¹⁰⁹ Notably, the required switching of electric field can be substantially lowered by introducing W to form Mo_1−xW_xTe₂ to destabilize the 2H-phase.¹⁰⁰ 

As discussed in Sec. II, ferroelectricity was theoretically predicted in various 2D layered materials with different polarization switching mechanisms. The polarization switching is generally derived from a transition to a plane-/centrosymmetric structure, which, therefore, is not necessarily accompanied with conductivity alteration. However, the remanent polarization in 2D channel will modulate the channel/electrode Schottky barrier^34,110 and/or the channel/substrate interface,¹¹¹ thereby modulating the channel current. Ferroelectric memristors are promising for large dynamic range because the tunneling current at the channel–electrode junction is very sensitive to the interfacial potential (the sensitivity could be further enhanced by using MIS tunneling junction).^112,113 Because of the intrinsic non-volatile multilevel memory effect, 2D ferroelectric synaptic devices are also promising in producing multiple conduction states and high linearity.¹¹⁴ However, stability is a critical issue for ferroelectric synaptic devices, especially for high temperature scenarios.¹¹⁵ 

As the study of 2D ferroelectricity is still in its infancy, 2D ferroelectric switching has been barely reported until recently. Hou et al. reported a α-In₂Se₃ ferroelectric memristor. Benefited from the unique intercorrelated IP and OOP ferroelectricity, an OOP polarization induced by the in-plane electric field pulls/pushes the electrons to/from the oxygen vacancies at the channel substrate interface, thereby resulting in the resistive switching in the channel.¹¹¹ The dynamic range is ∼10, and the retention is more than 300 s. A Pt/SnS/Pt ferroelectric analog synapse was later reported to have high linearity, long retention, and excellent reproducibility but a low dynamic range of ∼20.¹¹⁴ Because of the high work function of Pt, an Ohmic contact rather than Schottky contact is formed at the SnS/Pt before poling. As a comparison, Higashitarumizu et al. adopted Ag electrode and the Schottky junction at SnS/Ag interface is well established, leading to a high dynamic range in the Ag/SnS/Ag 2D memristor.¹¹⁰ Notably, SnS is stable under the ambient environment and no passivation layer is needed. As a comparison, an Al₂O₃ passivation layer is needed in a recently reported In₂Se₃ FeFET.³⁶ 

This Perspective summarizes the polarization switching and resistive switching properties in 2D layered materials with various mechanisms. 2D layered materials show significant potential in high-performance, high-density, low-power, and wearable non-von Neumann memory devices. Nevertheless, as the development of 2D memristive devices and applications is moving forward, a large number of new materials and/or new device configurations are to be explored, and several challenges remain to be solved. First, the fabrication of large-area monocrystalline 2D layered materials is crucial for the enhancement of remanent polarization because the difference of the direction of spontaneous polarization from different domains can cause the counteraction of each other and the vanishment of macroscopic polarization. Moreover, the asynchronism of the crystalline motion between domains could produce in-plane tension against itself. In addition, controlling the GBs in the as-grown materials is significant for achieving reliable resistive switching since the highly dense GBs in polycrystalline 2D layered materials would greatly affect the memristive properties. By incorporating salts as the growth promoters, a wide variety of 2D chalcogenides can be synthesized with large domain sizes.¹¹⁶ Second, a passivation layer should be adopted to encapsulate the material and isolate it from the ambient environment. The van der Waals interface has strong interaction with the environmental molecules,¹¹⁷ which generates significant impact on the ultrathin-film ferroelectricity and the resistive switching properties. Moreover, WTe₂ and some other 2D materials suffer from poor air-stability if encapsulation is absent.^118,119 Third, owing to the all-surface nature of 2D layered materials, the interface quality is a major concern when memristive devices with 2D materials designed and fabricated. The interfacial trap states may induce unwanted relaxation of the memristive states.¹²⁰ Hence, interface engineering is crucial for achieving high-performance non-volatile 2D memristive devices.

The work is in part supported by Research Grants Council of Hong Kong, particularly, via Grant Nos. 14203018 and N_CUHK438/18, and CUHK Group Research Scheme, CUHK Postdoctoral Fellowship, and No. ITS/390/18 by Innovation and Technology Commission, Hong Kong SAR Government.

Data sharing is not applicable to this article as no new data were created or analyzed in this study.