Memory and Synaptic Devices Based on Emerging 2D Ferroelectricity

Memory devices are an essential part of modern electronics. Efforts to move beyond the traditional “read” and “write” of digital information in volatile and non‐volatile memory devices are leading to the rapid growth of neuromorphic technology. There is a growing demand for memory devices with continuous memory states with various retention times and greater integration density with more energy‐efficient mechanisms. Two types of memory devices (i.e., non‐volatile digital memory and neuro‐synaptic devices) have been extensively investigated with emerging materials. Among numerous materials for such memory devices, in this review, the authors focus on 2D ferroelectric materials for promising memory and synaptic devices. Three types of memory devices based on 2D ferroelectric materials are classified and discussed here: 1) ferroelectric gating of semiconducting channels, 2) active ferroelectric channels, and 3) ferroelectric tunnel junction devices. It is known that atomically thin geometry competes with ferroelectricity, which can degrade the quality of the devices based on atomically thin ferroelectric materials. Various efforts to resolve the fundamental issue with emerging 2D ferroelectric materials and how they can be used as a critical element for memory and synaptic devices are surveyed.


Introduction
Moore predicted that the density of integrated circuits with silicon-based transistors would double every two years, based on the rapid development of fabrication techniques. [1][2][3] Now we are facing extreme fabrication challenges to continue meeting the DOI: 10.1002/aelm.202300211 prediction. In particular, the fabrication cost of integrating more circuits is increasing much faster than the semiconductor industry can afford. [4][5][6] Accordingly, new materials and memory device architectures are in demand, and extensive studies have been conducted for that purpose. [7][8][9][10][11] The primary candidates for nextgeneration memory devices are based on the type of switching mechanism, considering the materials' magnetic, resistive and phase-change, or ferroelectric features. [12][13][14][15][16][17][18][19] The switching characteristics of conventional 3D materials have been actively investigated, but the fabrication cost and fine engineering required for the device performance still remain critical issues with such 3D materials. Since the discovery of graphene, atomically thin 2D [20][21][22] and 1D [23] materials have emerged as new candidates for future memory devices. Moreover, a new research trend toward neuromorphic technology, mimicking biological synapses, which promises unprecedented memory operation has accelerated the exploration of new materials. [24][25][26][27][28] In this review, we survey one of the most important candidate materials' groups and their applications to memory and synaptic devices, 2D ferroelectric materials-based memory and synaptic devices. [29][30][31][32] For a systematic review, the specific materials and device structures are separated into three different device categories for 2D ferroelectric materials: ferroelectric field-effect transistors (FeFETs), active ferroelectric channel devices (such as ferroelectric semiconductor-based FET, FeSFET), and ferroelectric tunnel junction (FTJ) devices.

Memory Device Applications of Ferroelectric Materials
The memory device has been a critical component in semiconductor technology. Among the many types of memory devices, RAM is based on "read" and "write" operations, performed in any order, [68] which enables fast and wide computing applications. In the initial RAM, the presence of "volatility" in the device was an important aspect. With the development of nonvolatile flash memory devices, there have been extensive studies on next-generation, non-volatile RAM devices for faster computing and memory applications. Such non-volatile memory devices include magneto-resistive RAM (MRAM), [12,13] resistive RAM (RRAM), [14,15] phase-change RAM (PRAM), [16,17] and ferroelectric RAM (FeRAM). [18,19] In this review, we will mainly discuss memory and synaptic devices based on emerging 2D layered ferroelectric materials, beyond silicon-based volatile RAM devices. To overcome the technical limits of silicon-based memory devices, numerous studies have been conducted on FeRAM devices with 2D materials, targeting high integration density, long retention time, and fast read/write speed. [69] Besides such traditional memory applications, we will also discuss a new and recent application of FeRAM: synaptic devices for neuromorphic computing. Synaptic devices emulate biological synaptic operations and are based on the hysteretic features of memory devices. Multiple studies have attempted to achieve low-energy operation and continuous memory states in synaptic devices, using 2D ferroelectric materials. [70] We first present the basic features of FeRAM for memory and synaptic devices, including their key working principles, device structures, expected performances, and technical issues.
Conventional ferroelectric memory devices can be classified as either FeFET [33,35,36] or FTJ [41,42,71] devices. The two types of memory devices operate using electrically and optically switchable non-volatile polarization in ferroelectric materials. While Fe-FETs use semiconducting channels modulated by an underlying ferroelectric insulator (Figure 2a), in FTJ devices a ferroelectric insulator is used as a tunnel barrier between two metallic electrodes ( Figure 2b). Both FeFET and FTJ devices exhibit nonvolatile memory operation.
A critical issue in ferroelectric devices, involving semiconducting channels, is the presence of a substantial depolarization field which decreases the amount of time stored information is retained. In FeFETs, the semiconducting channels are gated by the electric dipole field of the ferroelectric insulators. Since the dielectric constant of the semiconductor in the channel is small (compared to that of metal), the screening length in the semiconducting channel becomes long with respect to the channel dimension and incomplete compensation of the dipole field occurs at the interface (Figure 3). In this way, a depolarization field is generated at the interface, which degrades the polarization field from the ferroelectric material. [72] As memory devices have been further scaled down, large depolarization fields and large leakage current issues have created problems for reliable and practical memory operation. [73] Accordingly, a new type of device, FeSFET, was developed with a ferroelectric semiconducting channel (Figure 2c). The direct control of the current channel partly resolves the issues of unfavorable depolarization fields and leakage currents. [50] We note that the use of atomically thin ferroelectric layers and high dielectric constant materials decreases the depolarization fields and leakage currents in the FeSFET. [74]

Synaptic Device Applications of Ferroelectric Materials
Synaptic devices based on emerging 2D ferroelectric materials are being developed for the energy-efficient neuromorphic computing of massive data, emulating biological synaptic operations.

Figure 4a
depicts synaptic data processing; neurotransmitters are sent from a pre-synapse to a post-synapse, which changes the synaptic connectivity (weight) between two neurons. In synaptic signal processing, the short-and long-term synaptic weight changes are the key operations, and are known as "synaptic plasticity". [75] Synaptic devices are now being designed to mimic biological synaptic plasticity. [76][77][78][79] Input spikes (e.g., stimuli introduced as electrical or optical pulses from the pre-synapse) produce a postsynapse current (PSC) in the post-neuron (Figure 4a). [80] Two kinds of PSC are present: excitatory and inhibitory PSC (EPSC and IPSC), which increase and decrease PSC by input spikes, respectively. Short-term plasticity (STP) and long-term plasticity (LTP), described in Figure 4b, are determined by the retention time of the synaptic weight of the memory device. [81] In general, STP conducts synaptic signal processing and LTP enables learning.
We note another important learning mechanism, spike-timing dependent plasticity (STDP). When two input spikes from the pre-and post-synaptic neurons arrive at a synapse, the relative time difference between the two signals determines the sign and magnitude of the change of the synaptic weight ( Figure 4c). [82][83][84] A similar plasticity, paired-pulse facilitation (PPF) or depression (PPD) is illustrated in Figure 4d with a time interval of 50 ms. In PPF (PPD), the PSCs (A1 and A2 in Figure 4d) are further increased (decreased) when the time interval between two input spikes is shorter. [83,85,86] The aforementioned plasticity has recently been achieved in synaptic devices based on 2D materials. A key advantage of 2D materials in this application is the extremely low energy consumption (down to 10 fJ per one synaptic event). [87] Ferroelectric 2D materials in particular are considered promising because of their voltage-controllable polarization with multiple polarization levels (synaptic weights), low operating power, and fast switching speed. [88][89][90][91][92][93][94] In the next chapter, we will introduce emerging ferroelectric 2D materials for memory and synaptic devices.

Ferroelectricity in Atomically Thin Geometry
Extensive studies have been conducted on the critical thickness of ferroelectricity in various 3D oxide materials in an effort to reduce their thicknesses. The 3D ferroelectric materials need to possess a broken centrosymmetry or inversion symmetry in their crystal structures, which generates intrinsic ferroelectric properties with net polarization. Representative examples of such materials include Nb-doped hafnium oxides (HfO 2 ), perovskite oxides, and their derivatives. Among the materials, we note that HfO 2 is a 3D material, but its critical thickness for ferroelectricity reaches 1-2 nm. [30,107] Various 2D layered ferroelectric materials have been found in van der Waals (vdW) crystals. For example, the inversion symmetry of CuInP 2 S 6 (CIPS) can be broken under an electric field because the positions of P-P dimers or Cu ions are slightly changed by the electric field; as a result, electric polarization (P) and its switching can be achieved in ultrathin CIPS (Figure 5a). [108] In 2 Se 3 has quintuple layers (QLs) of Se-In-Se-In-Se, and the QLs are weakly bounded by vdW interactions, as shown in Figure 5b. The displacement of Se atoms in the central Se layer in the QL produces a spontaneous out-of-plane and in-plane electric polarization, as shown in the side and top views of In 2 Se 3 in Figure 5c. [109] Ferroelectricity can be obtained in certain non-ferroelectric materials (i.e., without inversion symmetry breaking) by decreasing their thickness down to an atomically thin limit. For example, certain 3D bulk vdW crystals, including SnTe, SnS, SnSe, and GeS, have no ferroelectricity in their pristine bulk form. However, switchable electric polarization can be observed in their atomically thin geometry. Theoretical calculations have explained that the ferroelectricity originates from lattice distortion and the resulting atomic displacement in the atomically thin geometry. [110][111][112][113] Kwon et al. demonstrated that SnS becomes ferroelectric when its thickness is below 6 nm. Thus, thin SnS could be used as a semiconducting channel with ferroelectricity for synaptic device applications. [52] The authors further showed that a screening effect results in poor ferroelectricity in thick SnS due to the increased charge density in the channel produced by the electric polarization, which degrades the performance of synaptic devices. . Biological synaptic signal processing. a) A schematic of biological neurons, synapses, and artificial neuronal systems. Reproduced with permission. [80] Copyright 2020, The Author(s). b) Short-term potentiation and depression (top) and long-term potentiation and depression (bottom) by post-synapse current/potential. Reproduced with permission. [81] Copyright 2020, The Authors. Published by Wiley-VCH GmbH. c) STDP implementation. Two graphs of time versus voltage show the relative timing of the pre-and post-synaptic spikes. The upper panel shows the potentiation of the synaptic weight (matched with the SET process), and the lower panel shows the depression of the synaptic weight (matched with the RESET process). Reproduced with permission. [84] Copyright 2018, The Author(s), published by Springer Nature. d) PPF and PPD operations by two pulses with a time constant of 50 ms. Reproduced with permission. [86] Copyright 2015, WILEY-VCH GmbH.

Layer-Stacking Driven Ferroelectricity
The artificial stacking of vdW layered materials allows unprecedented ways to produce ferroelectricity in the atomically thin geometry. The stacking of layers has several degrees of freedom: the choice of materials, their relative lattice orientations (i.e., twist angles), and displacement. A representative physical phenomenon from the materials' engineering is the formation of moiré lattice (potential) where strongly-correlated electronic band structures can be generated with much longer lattice constants. Based on the strongly-correlated electrons and feasible lattice distortion, non-polar materials (e.g., bilayer graphene) can demonstrate ferroelectricity.
The rise of ferroelectricity in bilayer graphene with top and bottom hexagonal boron nitride (h-BN) layers is schematically shown in Figure 6a. [59,114] We note that the Bernal stacking (ABstacking) of bilayer graphene does not induce any ferroelectric features by itself. Two kinds of electronic band structures are involved in the ferroelectricity in the system: AB lattice symme-try breaking resulting in bandgap (Δ) and Hubbard bands from strongly-correlated electrons.
A vertical electric field in the "+z" direction breaks the AB lattice symmetry and generates a bandgap (Δ) in the bilayer graphene in Figure 6a. Due to the electric field, electrons are accumulated in the bottom graphene, which could be considered a trivial polarization. In the presence of moiré potential, however, strongly-correlated electrons in the bottom graphene split the moiré valence band into upper and lower Hubbard bands under the electric field Figure 6b. When the correlation energy (i.e., onsite repulsion) U is larger than Δ, certain electrons should move to the top graphene layer with lower energy states (Figure 6b). The unique interlayer charge transfer is opposite to the electric field direction, and thus, becomes hysteretic, which is a novel origin of ferroelectricity by layer stacking. [59,114] In addition to the electronic origin, lattice sliding partly contributes to the ferroelectricity, where carbon atoms overlapping with either boron or nitrogen atoms in the h-BN layer produce two polar states ( Figure 6c). The polarization direction is indicated by the black arrow (out-of-plane direction). Reproduced with permission. [108] Copyright 2016, The Author(s), published by Springer Nature. b) 3D crystal structure of In 2 Se 3 . c) A side view of In 2 Se 3 . When a vertical electric field is applied, Se atoms move and form a polarization in the opposite direction. Reproduced with permission. [109] Copyright 2017, The Author(s), published by Springer Nature.
Orbital engineering via lattice sliding has been used to generate polar states in other layered materials. Bulk h-BN has no ferroelectricity with its AA' stacking order ( Figure 6d). However, bilayer h-BN with AB (BA) stacking orders has significantly deformed orbitals of nitrogen atoms, as shown in Figure 6e,f. The two polar states can be converted by lattice sliding, which is reversibly controlled by electric fields. The polarization switching resulting from lattice sliding can be observed by Kelvin probe microscope (KPFM), piezoelectric force microscope (PFM), and transport measurements. [40,[60][61][62] In semiconducting transition metal dichalcogenides (TMDs), orbital engineering can be achieved via lattice sliding, for H (hexagonal) and R (rhombohedral) stacking orders, as shown in Figure 6g. For example, Marmolejo-Tejada et al. demonstrated lattice sliding-induced polarization switching in several TMDs (WSe 2 , MoSe 2 , WS 2 , and MoS 2 ) with R-stacking, by PFM and transport measurements; all the R-stacking materials have inversion symmetry breaking. [115] Hetero-structured TMDs, such as MoS 2 /WS 2 , have exhibited ferroelectricity. [62] The relative orientation between the two TMDs can be confirmed by the second harmonic generator (SHG) (Figure 6h), and their ferroelectricity can be demonstrated by PFM and FTJ device measurements.

2D Ferroelectric Metals
It is generally considered that ferroelectric (polar) metallic materials cannot be practically used in device applications due to their inevitable screening effect. In the atomically thin geometry, however, ferroelectric and metallic 2D materials can be used as active channels in devices based on ferroelectric operation, because the screening effect can be minimized in the 2D materials; external environments which possess small dielectric constants can pro-vide the lower screening effect, with sufficient screening of the electric field.
Despite their metallic features, certain 2D polar metals (e.g., WTe 2 , MoTe 2 , Mo 1−x W x Te 2 , and ZrI 2 ) have been reported to possess ferroelectricity with switchable polarization. Figure 7a shows three common crystal structures of the polar materials: non-polar monoclinic 1T', polar orthorhombic T d with upward spontaneous polarization (T d,↑ ) and downward spontaneous polarization (T d,↓ ). The three crystal structures can be switched by external electric fields. We note that the single layer of all three structures has the same lattice structure. In polar metals, relative interlayer sliding enables distinct structures for polarization. [116] Unlike the polarization switching mechanism of conventional ferroelectric materials (i.e., atomic or ionic displacement), the switchable polarization in the above 2D ferroelectric metals arises due to uncompensated interlayer charge transfer, originating from weak interlayer sliding. [105,106] Ferroelectricity also emerges in certain non-polar monoclinic 1T' phases of TMDs, which are maintained down to atomically thin geometry. For example, it is known that monolayers of MoTe 2 , WTe 2 , and Mo 1−x W x Te 2 exhibit no ferroelectricity due to their non-polar crystal structure (1T'), but their bilayer and trilayer geometries with polar crystal structure T d show spontaneous polarization. In WTe 2 and Mo 0.5 W 0.5 Te 2 , polarization switching is achieved by interlayer sliding. [105,106] Their conductivity can be changed abruptly during operation in FeFETs when the electric polarization is switched. [55,56,117] Such spontaneous polarization results from the phase transition between their two polar T d structures. Moreover, ferroelectric Mo 1−x W x Te 2 can be optimized by adjusting the atomic substitution x and electric gating voltage V g , enabling tunable polar structures and robust ferroelectricity up to T = 400 K. [56] The laser-driven phase transition to a ferroelectric state has been observed in a MoTe 2 monolayer. [118] As shown in the transmission electron microscopy (TEM) image in Figure 7b, the Figure 6. Layer-stacking driven ferroelectricity. a) A schematic of h-BN/bilayer graphene/h-BN system structure. Reproduced with permission. [114] Copyright 2022, The Author(s), published by Springer Nature. b) The electronic structure of the h-BN/bilayer graphene/h-BN system under an external electric field. Reproduced with permission. [59] Copyright 2020, The Author(s), published by Springer Nature. c) AB (-P polarization) and BA (+P polarization) stacking orders, separated by a domain wall. d-f) No polarization is shown in the AA' stacking order, while downward (upward) polarization occurs in the AB (BA) stacking orders. Reproduced with permission. [40] Copyright 2021, The American Association for the Advancement of Science. g) Polarization in different stacking orders in TMDs. Reproduced with permission. [60] Copyright 2022, The Author(s), under exclusive license to Springer Nature Limited. h) Polarization in CVD-grown MoS 2 /WS 2 . Reproduced with permission. [62] Copyright 2022, The American Association for the Advancement of Science.
Te atoms have been slightly shifted by the laser irradiation, generating spontaneous electric polarization in the out-of-plane direction. [118] Figure 7c shows that in-plane polarization is completely compensated, and thus, out-of-plane polarization is maintained, producing a net polarization in the 1T' phase.
Certain ferroelectric metals exhibit a superconducting state at low temperatures. Since ferroelectric polarization induces an electric field, the superconducting state can be manipulated with a unique hysteretic switching behavior. Jindal et al. demonstrated control of the transition between ferroelectric superconductive states in bilayer T d -MoTe 2 . [58] In addition, a FeFET with a superconducting Nb-doped SrTiO 3 (Nb-STO) channel and ferroelectric Pb(Zr, Ti)O 3 (PZT) gate oxide has been fabricated. With positive and negative polarization (+P and -P) of the PZT gate, the electrons in the Nb-STO channel are depleted or accumulated, which modulates the critical temperatures of the superconducting state. At low temperatures, the switching of the ferroelectric polarization induces a phase transition in the channel between the normal state (with +P polarization) and the superconducting state (with -P polarization). . Spontaneous polarization is shown as P. The yellow spheres refer to W atoms and the black spheres represent Te atoms. Reproduced with permission. [116] Copyright 2020, The Author(s), published by Springer Nature. b) An atomic resolution TEM image of monolayer d1T-MoTe 2 . Cyan and orange colors represent Mo and Te atoms, respectively. Scale-bar = 2 Å. c) A top-and side-view of the charge density difference between ferroelectric 1T' and paraelectric 1T phases, obtained by DFT calculations. Reproduced with permission. [118] Copyright 2019, The Author(s), published by Springer Nature.

Ferroelectric Gating of 2D Semiconducting Channels
Ferroelectric materials have been investigated for use in promising non-volatile memory and synaptic devices. The first approach controlled the carrier density of a semiconducting channel using the polarization field from a ferroelectric material. [33,34,[36][37][38][39][119][120][121][122][123][124] We note that ferroelectric materials have certain domains that have different polarization directions along their easy axis. The continuous change in the polarization field from the ferroelectric domains enables multiple conductivity states in the semiconducting channel, which can be used as a synaptic device and learning process. [125] Here, we review recent studies on the ferroelectric gating of low-dimensional semiconducting channels, which includes gating via electrical or optical stimuli, mimicking neuromorphic synaptic devices. We also summarize strategies to improve device performance and the integration of FeFETs based on 2D semiconducting materials.

Low-Dimensional FeFETs
A thin ferroelectric material Hf 0.5 Zr 0.5 O 2 (HZO) and a semiconducting channel of MoTe 2 have been used for a non-volatile memory device with a high ON/OFF ratio of 9 × 10 3 over 100 cycles. [33] The HZO/MoTe 2 FeFET showed polarization induced by charge trapping at the interface between the HZO and MoTe 2 , and the polarization direction could be controlled by the number of trapped charges at the interface, as shown in Figure 8a. The HZO/MoTe 2 FeFET exhibited both short-and long-term plasticity with sequential electric pulses, which can be applicable in a synaptic device, as shown in Figure 8b. The asymmetric ratio (AR), which is the quantitative value of the nonlinearity of the long-term potentiation-depression, was ≈0.16 in the HZO/MoTe 2 FeFET. The multiple conductance states allow an effective learning process, as shown in Figure 8c. [33] A FeFET based on 2D ferroelectric and 2D semiconducting materials exhibited good reliability and tunability. Figure 9a shows a FeFET with a ferroelectric CIPS and semiconducting MoS 2 . [34] Depending on the polarization direction of the 2D ferroelectric, the 2D semiconducting channel exhibits high or low resistance states (HRS or LRS) as shown in Figure 9b. The CIPS/MoS 2 FeFET has synaptic behaviors such as EPSC, PPF, and spike-amplitude-dependent plasticity (SADP) (Figure 9c). Figure 10a shows the architecture of a FeFET with dual gating ferroelectrics (top and bottom gating operation). A ferroelectric polymer P(VDF-TrFE) and semiconducting TMDs have been used as the bottom gate dielectric and 2D channels, respectively. Pulse timing-dependent current changes with excitatory (left) and inhibitory (right) features. c) Schematics of the pattern learning process based on the FeFET. Reproduced with permission. [33] Copyright 2022, Wiley-VCH GmbH.  Reproduced with permission. [35] Copyright 2022, American Chemical Society.
At the interface between the ferroelectric polymer layer and the semiconducting 2D channel, the electronic band structure can be modulated, leading to significant band bending at the interface. Thus, the dual-gating FeFET exhibits four different conductance states by changing the polarization of the ferroelectric layer, which can be controlled by the top and bottom gate voltages ( Figure 10b).
The four conductance states in the FeFET in Figure 10a could be used as logic gates, as shown in Figure 10c. In addition, the FeFET showed synaptic device performance with LTP and longterm depression (LTD). The LTP and LTD were demonstrated by voltage pulses (Figure 10d), and an accurate learning process could be achieved with the FeFET (i.e., an accuracy of 86-88% by 40 training processes, as shown in Figure 10e). [35] A similar FeFET, a ferroelectric-based synaptic device with a ferroelectric polymer P(VDF-TrFE) layer and 2D semiconducting MoS 2 channel, is shown in Figure 11a. [36] Figure 11a shows the polarization domains of the P(VDF-TrFE) layer, which leads to various polarization hysteresis curves and multiple conductance states over an ON/OFF ratio of 10 4 . In Figure 11b, the potentiation and depression of the conductance (G) show LTP, LTD, and STDP, which are the elemental operations in synaptic devices. The authors demonstrated the basic performance of a spiking neural network (SNN) using the FeFET. [36]

Optically-Operating FeFETs
It is known that the polarization hysteresis curves of PZT and BaTiO 3 (BTO) can be modulated by light illumination (see Figure 12a,b). [126,127] We explain the phenomenon using optically excited and trapped charges in certain ferroelectric domains. Accordingly, FeFETs with semiconducting 2D materials enable optically-controllable polarization and conductance. Under light illumination, free charge carriers are generated in the conduction bands of the semiconducting 2D channels in the FeFETs, which induces an additional electric field; the polarization of the ferroelectric layer underneath the 2D channel can be modulated by light illumination. In contrast to reversible polarization switching using electric pulses (fields), polarization switching by optical pulses occurs in an irreversible way, meaning that the "write" process (from the OFF to ON state) can only be achieved by light illumination without an erasing operation.
Recently, a FeFET has been reported with ferroelectric PZT and BTO layers and 2D semiconducting channels of WS 2 and MoS 2 , which exhibits a light response that flips the polarization direction. In the FeFET, excited charges are accumulated in the 2D TMDs (semiconducting) channels, which enhances the electric field high enough to flip the polarization direction of the ferroelectric domains. Figure 12c,d shows that the polarization of PZT/MoS 2 and BTO/MoS 2 has been switched by optical pulses. [119,120] The authors reported that the phase and the amplitude signals in PFM were only changed in the PZT/MoS 2 or BTO/MoS 2 area, not in the bare PZT or BTO area upon the light illumination. Li et al. suggested a detailed mechanism, as described in Figure 12e. In the dark condition and upward polarization direction of the BTO, electrons are accumulated at the interface between the BTO and MoS 2 . Due to the charge distribution, the heterostructure exhibits an electric field, as shown in the left panel in Figure 12e.
Once the light is on, excitons are formed over a long distance. If there is no external electric field, intralayer exciton formation should be dominant. However, in the presence of an external electric field, interlayer exciton formation becomes dominant, and the electron-hole pairs are separated by the external electric field. As shown in Figure 12e, hole carriers are accumulated at the interface between the BTO and MoS 2 . The hole carriers make the E and E' weaker and can flip the polarization direction, although the E bi does not change. If enough hole carriers are accumulated, the polarization is switched downward.
FeFETs with a structure of PbZr 0.2 Ti 0.8 O 3 /WS 2 show mixed polarization domains, which can be controlled by electric and optical pulses, [37] as shown in  Optically-controlled ferroelectric materials. a) Polarization hysteresis curves with polarization configuration in ferroelectric BTO with and without light illumination. Reproduced with permission. [126,127] Copyright 1994, American Institute of Physics. b) Locked polarization because of charges trapped at the domain boundary. Reproduced with permission. [126] Copyright 1994, American Institute of Physics. c) PFM results of PZT/MoS 2 FeFET. Reproduced with permission. [119] Copyright 2015, American Chemical Society. d) PFM results of a BTO/MoS 2 system. e) A schematic diagram of how the polarization can be changed with a combination of perovskite ferroelectric materials and MoS 2 (2D TMDs). E is the electric field in MoS 2 due to the accumulated charges at the interface. E is the electric field in BTO due to the accumulated charges at the interface. E bi is the built-in electric field due to the charge distribution in BTO. P updown is the polarization direction in BTO which is opposite the direction of the E bi . The charges accumulated at the 2D TMDs when light enhanced the external electric field, which can flip the polarization of the perovskite materials. Reproduced with permission. [120] Copyright 2018, The Author(s), published by Springer Nature.
In the PbZr 0.2 Ti 0.8 O 3 /WS 2 FeFET, the LTP was separately obtained using electric (both depression and potentiation) and optical (only potentiation) pulses, respectively (Figure 13b,c). Nonvolatile memory with multiple states and frequency dependency is demonstrated in Figure 13b,c, which are key operations for the synaptic device (i.e., a memtransistor).
Because the polarization switching introduced by weak light illumination can be effectively screened by the WS 2 layer, the light intensity can determine the LTP and STP features of the PbZr 0.2 Ti 0.8 O 3 /WS 2 -based FeFET. With low light intensity, the retention time becomes shorter (i.e., STP), while a long retention time (i.e., LTP) is observed under strong light illumination (Figure 13d).
In BTO/MoS 2 -based FeFETs, both electrical and optical pulses induce long-term erasing or writing processes [38] (Figure 13e). The conductance of MoS 2 is changed by the wavelength of the light. In Figure 13f, for the same light pulse width, the conductance change is larger when the wavelength is shorter. Using this  [37] Copyright 2020, American Chemical Society. e) A schematic of the polarization flip induced by the light illumination. f) Wavelength dependency of the conductance when the light pulse is applied at the yellow region. g) Conceptual graphic encoding of the color information into the corresponding photocurrent intensity. Reproduced with permission. [38] Copyright 2021, Elsevier Ltd. feature, the authors conducted in-sensor neuromorphic visual preprocessing for image recognition (with 90% accuracy) and a color mapping test (Figure 13g). When the pulse number was increased to 100, the conductance increased and was able to clearly identify images and perform the color mapping test.
Optically-controllable ferroelectric polymer-based FeFETs have been investigated with one of the most frequently used ferroelectric polymers, P(VDF-TrFE) with the phase. An asymmetric Fe-FET (AFeFET) with ferroelectric P(VDF-TrFE) and 2D semiconducting TMDs channels was suggested. Figure 14a shows that P(VDF-TrFE) partially covers the semiconducting MoS 2 channel, where a potential barrier is made between the gated and nongated channel areas. [39] This asymmetric interface results in a highly enhanced ON/OFF ratio compared to common FeFETs.
As shown in Figure 14b, AFeFET exhibits a large switching window compared with normal MoS 2 FETs, which is a major characteristic of memory devices. The authors confirmed that the clockwise hysteresis behavior in Figure 14c is strong evidence of the ferroelectricity of the device. Furthermore, in Figure 14d, self-rectifying behavior was observed in the AFeFET with electrical pulses. When the authors applied gate pulses with different amplitudes, the polarization of the P(VDF-TrFE) domains flipped gradually, leading to a continuous increase in channel current.
The polarization of the P(VDF-TrFE) was reliable over numerous operations (up to 1000 s), which indicates a practical retention time (Figure 14e). Besides electrical pulses, optical pulses can also modulate the conductance of the system. In Figure 14f, when the authors increased the light intensity of the optical www.advancedsciencenews.com www.advelectronicmat.de pulses, the current increased step-by-step simultaneously with the optical pulse sequence. Along with the light intensity increase, the ON/OFF ratio also increased (Figure 14g).

FeFETs with Moiré Superlattices
Moiré superlattices have been extensively studied because of their unique long-range periodicity, which enables fascinating physics and chemistry in artificial atomic crystals. However, the transport properties observed in the systems have limited the use of the unique electronic structure for novel ferroelectricity, and Fe-FETs from moiré superlattices. A few FeFETs with moiré lattice have been reported so far: FeFETs based on bilayer graphene, [59] bilayer h-BN, [40] and bilayer TMDs. [60,61] Parallel stacking of bilayer h-BN (P-BBN) is one of the famous insulating ferroelectric materials with a moiré superlattice. Yasuda et al. studied P-BBN-based devices and showed ferroelectric hysteresis in the devices. In their device schematic (Figure 15a), the P-BBN acts as the bottom gate (V B ). Thus, the hysteresis (i.e., resistance peak shift) is generated from the bottom gate sweep (Figure 15b) not by the top gate sweep (Figure 15a). The peak shift is larger than in other moiré-based ferroelectric systems, and the coercive field is smaller than in other systems, which means that the P-BBN system uses less energy than other systems.
In Figure 15c, the non-twisted bilayer h-BN system shows an abrupt polarization change, while the twisted bilayer h-BN system exhibits a gradual polarization change. The abrupt change is an advantage for ordinary memory devices, and the gradual change is critical for synaptic devices. We note that the operating temperature ranged from T = 4 to 300 K (Figure 15d). Electric pulses modify the ferroelectric polarization (Figure 15e). The moiré system is reliable over a long time, as shown in Figure 15f, which is promising for memory and synaptic devices.

Active Ferroelectric Channel Devices
Highly-integrated and nanometer-scale FeFETs have suffered from short retention times and large leakage currents, which will require further breakthroughs before 2D ferroelectric materials can be applied for practical memory and synaptic devices. In Section 2, we introduced ferroelectric semiconductors to resolve these issues. Mobile charges in ferroelectric semiconductors can effectively screen the depolarization field across the semiconducting channel, because of their abundant carriers. As a result, the geometry of ferroelectric semiconductor-based synaptic devices can be reduced down to the 2D limit (i.e., atomically thin geometry). In this section, we survey recent progress on synaptic devices with active ferroelectric channel materials, which exhibit typical memory functions as well as synaptic behaviors such as STP, LTP, PPF, and STDP.

Ferroelectric Semiconducting Channels
The two device structures of conventional FeFETs and FeSFETs are compared in Figure 16a. [50] A pioneering ferroelectric semiconducting material, a layered semiconductor, -In 2 Se 3 , was used for FeSFETs. The material exhibits inherent out-of-plane and in-plane polarization directions; its ferroelectricity is induced by a slight shift in the Se atoms in the layered structure. FeS-FETs with an -In 2 Se 3 channel have demonstrated ultrafast nonvolatile memory switching with large hysteresis windows and long-term robust retention of information. [29,51,70,101,[128][129][130] It has been reported that such FeSFETs can emulate biological synapses with low operating energy (Figure 16b,c). [29] Figure 17a shows another synaptic device using -In 2 Se 3 in the semiconducting channel. [51] The authors demonstrated both non-volatile memory and synaptic device operations based on the -In 2 Se 3 . A unique performance in the device was the ultrafast transition between LRS to HRS by electric pulses with a width of 40 ns, as shown in Figure 17b.
For neural computing with the device, the top gate voltage simulated the presynaptic input and the ferroelectric channel current was monitored as the PSC. When a negative gate voltage spike pulse is applied, the PSC responds to the signal but quickly returns back to its initial state, which mimics a typical biological STP. With a low spike voltage (± 0.5 V, 30 ms), progressive EPSC and IPSC modulation could be achieved, which mimics LTP and LTD. Furthermore, the authors estimated the energy consumption per spike to be 234 fJ, which is much lower than biological energy consumption. Thus, using -In 2 Se 3 as an active channel provides a promising perspective for high-density and low-energy applications in memory and synaptic devices. Figure 17c shows a schematic of a ferroelectric synapse device with a structure of Pt/SnS/Pt. The two-terminal synaptic device with Pt/SnS/Pt exhibits synaptic behaviors at room temperature. Conductance change in the potentiation and depression operation has been observed in the Pt/SnS/Pt with a high ratio of maximum and minimum conductance (G max /G min = 20.5) using electric pulses with various amplitudes (Figure 17d). The multilevel conductance states and high conductance ratio could be achieved by partially switching the ferroelectric SnS domains, which led to high pattern recognition accuracy in an artificial neural network (ANN) simulation. [52] The perfect linearity (Figure 17d), long retention time, and reliable operation with low cycle-tocycle/device-to-device variations make the device promising for neuromorphic synaptic devices. [52]

Ferroelectric Heterostructure Channels
Numerous synaptic devices with 2D ferroelectric active channels have been reported so far. Among them, certain studies have used complex architectures with, for example, a floating gate and charge trapping layer, and photosensors. To address the challenges of producing 2D ferroelectric memory and synaptic devices with simple architectures, 2D vdW heterostructures and more complex device structures with various functional materials have been suggested.
For multifunctional optoelectronic synaptic devices, a 2D vdW heterostructure of -In 2 Se 3 /GaSe has been proposed. [53] The -In 2 Se 3 /GaSe heterostructure device structure is shown in Figure 18a. The -In 2 Se 3 and GaSe have light absorption capacity with light wavelengths shorter than near-infrared 870 nm and visible 620 nm, respectively, which makes them useful as proper optoelectronic synaptic devices. In the -In 2 Se 3 /GaSe heterostructure device, the channel conductance between the source and drain electrodes is defined as synaptic weight, which can be tuned by electrical and light stimuli. Synaptic device features, such as PPF, LTP, and LTD, were demonstrated using consecutive electrical pulses in the device. Ferroelectric polarization switching is the key mechanism of device operation; the conductance of the chan-nel is modulated by local ferroelectric domain switching in -In 2 Se 3 . Applying light pulses increases the conductance, which mimics the biological retina. If the periodic light pulses are stopped, the conductance decreases back and is saturated. In addition, different wavelength light pulses produce distinct responses in the device, as shown in Figure 18b. The distinct behaviors can be  After applying a spike with +8 V for 40 ns, the channel current is changed from the initial LRS "1" to HRS "0" with non-volatile features. The inset shows the actual waveform: a 40 ns ultrafast write spike. Reproduced with permission. [51] Copyright 2021, The Author(s), published by Springer Nature. c) Schematic of a FeSFET based on SnS. d) Conductance changes in potentiation and depression with incremental pulses (+1 to 3.5 V, 25 mV steps, 100 states). Reproduced with permission. [52] Copyright 2020, American Chemical Society. understood by the absorption of the two materials. The highest PSC is achieved with 450 nm wavelength light because both -In 2 Se 3 and GaSe absorb that wavelength light, which produces more photoexcited carriers, increasing the current up to 52 nA (left panel in Figure 18b). -In 2 Se 3 only absorbs 808 nm wavelength light (the middle panel in Figure 18b), and 980 nm wavelength light is not absorbed by either material (right panel in Figure 18b). Accordingly, the -In 2 Se 3 /GaSe heterostructure device has three photocurrent levels, which can distinguish the color of light.
Besides converting light signals into electric signals, the -In 2 Se 3 /GaSe heterostructure device can distinguish the color of light, like the retina. As shown in Figure 18c, Pavlov's dog experiment was conducted to mimic the classical conditional reflex. The light pulse stimuli and electrical pulse were considered the food and bell ringing in the experiment, where the threshold for response (dog's salivation) was measured to be 7 nA. Before training (only electrical pulses were used), the PSC cannot overcome the threshold, and no response is observed (violet curve in Figure 18c). When light pulses are introduced, a PSC of 10 nA occurs and the dog's response to food takes place (green curve in Figure 18c). The training process is shown by the blue curves in Figure 18c when both electrical and optical pulses were applied simultaneously. After the training, the PSC can reach up to 8.9 nA with only electrical pulses, which overcomes the threshold (7 nA). When an electrical pulse train was repeated for a period of 120 s, the PSC gradually decreased, which is similar to the forgetting process in biological nervous systems. Thus, the -In 2 Se 3 /GaSe heterostructure device can act as a retina, which processes electrical and optical signals in a complex way.
For low power consumption synaptic device operation, a WSe 2 /In 2 Se 3 heterostructure-based device was developed by Li et al., as shown in Figure 19a. [54] We note that, prior to this study, the range of input signals had not covered the short-wavelength infrared (SWIR) range in synaptic devices. SWIR light is known to be less affected by atmospheric scatterings, compared with visible light, and SWIR is used in various fields for this reason. The authors demonstrated essential light tunable synaptic functions, such as STP, LTP, and PPF, in the WSe 2 /In 2 Se 3 heterostructurebased device with, for the first time, short-wavelength infrared range light (up to 1800 nm).
The current responses to 1800 nm wavelength light with an intensity of 5.88 Wcm −2 and duration times of 1 and 5 s are shown in Figure 19b. An optical pulse train of SWIR light generates reliable potentiation (Figure 19c). The operating energies of two devices, a device with only an In 2 Se 3 device and a WSe 2 /In 2 Se 3 heterostructure device, were compared using the same intensity light signals (wavelength = 750 nm). The operating energy consumption of the In 2 Se 3 device was measured to be 12 pJ per event (Figure 19d), which is 10 times larger than the energy consumption in the human brain (≈10 to 100 fJ for one single synaptic event). [79,87,131] In the WSe 2 /In 2 Se 3 heterostructure device, the energy consumption was measured to be 2 pJ per event, which was ≈10 times smaller than that of the In 2 Se 3 device (Figure 19e,f).
The authors claimed that this low operation energy of the WSe 2 /In 2 Se 3 heterostructure device originates from the p-n junction formed at the heterostructure interface. The lowest observed operating energy was 258 fJ per event at V DS = −1 V, which is similar to the energy consumption in the human brain. This lowenergy operation is expected to contribute to the realization of multifunctional visual nervous systems based on 2D materials.

Ferroelectric Semimetallic Channels
It has been suggested that polar metals with atomically thin geometry can be used for device channels. In contrast with semiconductor channels, a thin polar metal channel can result in lower contact resistance and higher mobility, [132] which is beneficial to the transistor, memory, and synaptic device applications. We note that synaptic devices with polar metals have not yet been reported, but it is considered promising to use polar metals for neuromorphic computing. For example, atomically thin semimetals such as WTe 2 , MoTe 2 , and Mo 0.5 W 0.5 Te 2 are polar metals, and their polarizations can be switched by an external electric field.
The first device with a channel based on a polar metal (WTe 2 ) was reported by Fei et al. [55] Using few-layered graphene electrodes as the top and bottom gates, a vertical electric field was applied for the polarization switching. In Figure 20a, a sudden change of conductance is shown at certain vertical electric fields in the devices with trilayers and bilayers of WTe 2 . The sudden conductance drops demonstrate the switching of the polarization in the channel. The authors observed such ferroelectric switching near the zero electric field at all temperature ranges. There was no ferroelectricity in the WTe 2 monolayer because interlayer sliding is not allowed in the single-layer geometry.
Eshete et al. reported ferroelectricity in polymorphic Mo 0.5 W 0.5 Te 2 , enhanced by an atomic manipulation using chemical pressure and charge density modulation ( (Figure 20b). [56] The Mo 1−x W x Te 2 systems exhibit various crystal structures (i.e., polymorphism), which can be tuned by charge density and strain. In a series of Mo 1−x W x Te 2 alloys, there are two non-polar phases (hexagonal [2H] and monoclinic [1T′]) and one polar phase (orthorhombic [T d )]). By controlling the stoichiometric ratio, the phase transition from 2H to 1T' or from 1T' to T d could be realized.
As shown in Figure 20c, the non-polar 2H and 1T′ phases appear in a range of 0 ≤ x ≤ 0.4. The two phases are centrosymmetric crystal structures, where no spontaneous electric polarization can be formed. In a range of 0.4 ≤ x ≤ 1, the T d phase appears, which exhibits intrinsic ferroelectricity originating from its non-centrosymmetric crystal structure. (right in Figure 20c). The authors realized tunable polar lattice structures and robust ferroelectricity up to T = 400 K with a constant coercive field in an atomically thin material. The atomic and electronic manipulation of the ferroelectric polymorph of Mo 0.5 W 0.5 Te 2 produced larger polarization and a higher critical temperature than those with defective MoTe 2 or pristine WTe 2 . [55,56,118,[133][134][135][136] The first-principles calculations, conducted by the authors, supported the increase in polarization, compared with pristine WTe 2 (Figure 20d).
An optically controllable FeFET with a ferroelectric defective 1T'-MoTe 2 layer was reported by Park et al. [57] The defective 1T'-MoTe 2 area was fabricated by laser irradiation of 2H-MoTe 2 ; the resulting device channel partly contained semiconducting 2H-MoTe 2 . The device structure is shown in Figure 21a, where the multilayered 1T'-MoTe 2 also acts as a photoactive layer. Thus, a photocurrent is generated under light illumination.
An additional poling process could be achieved using a vertical electric field, which induces an extra built-in potential by Figure 20. 2D ferroelectric semimetal channels. a) Conductance G of undoped trilayer and bilayer devices as the vertical electric field is swept (black arrows). Reproduced with permission. [55] Copyright 2018, Macmillan Publishers Ltd., part of Springer Nature. b) A schematic cross-section of the device, showing the gate bias (V t and V b ) geometry and the application of vertical electric fields (E ⊥ ). c) Evolution of the crystal structure of Mo 1−x W x Te 2 with different stoichiometric ratios of tungsten (x). The polar lattice structures of the T d phase with two lattice configurations (polarizations) are shown in the right panel. d) First-principles calculations of bilayer WTe 2 and Mo 0.5 W 0.5 Te 2 . Compared to WTe 2 , the polarization increased by 100% in Mo 0.5 W 0.5 Te 2 in the calculation. Reproduced with permission. [56] Copyright 2022, Wiley-VCH GmbH. ferroelectric polarization. Two opposite poling directions were applied with an +8 and −8 V bias in the top and bottom graphene electrodes under light illumination ON/OFF. Under the negative bias, the photocurrent became larger, while the photocurrent was reduced under the positive bias ( Figure 21b). As shown in Figure 21c, photogenerated electrons and holes readily moved to each source and drain electrodes due to the band bending with a negative bias (−8 V); a ferroelectric polarization field was applied along the same direction with a built-in potential. The positive bias reverses the process.
Superconductivity was observed in a polar metal, bilayer T d -MoTe 2 . [58] The unique ferroelectricity can be used for conceptually new devices, by electrically controlling the superconductivity. For undoped bilayer T d -MoTe 2 , a clear superconducting transition is observed at a curie temperature (T c ) of 2.3 K. By sweeping an out-of-plane electric field in the system, the authors demonstrated a hysteretic switching in resistance between the normal (4 K, above the T c ) and superconducting states (1.7 K, below the T c ); by varying the electric field, the superconducting state exhibited a sudden drop in resistance to zero, as shown in Figure 21d.
The resistance remained zero (i.e., superconducting state) until the electric field switched the polarization at a displacement field D < −2 V nm −1 (Figure 21d). With reverse sweeping, a similar hysteretic behavior was observed. To explore the correlation between ferroelectric switching and superconductivity, the authors performed further electric field sweeps at different temperatures (Figure 21e). The superconducting critical temperature was clearly changed by the electric field sweeps in both directions and by the polarization emerging from the ferroelectric state. Their study means that the superconductivity is affected by the internal electric field from the ferroelectric state.
The authors further verified that the independent electron and hole carrier densities of the bilayer also determined the critical temperature (Figure 21f), which shows that the maximum T c occurs at the compensation point of the material, electrostatic doping Δn = 0. Upon doping the bilayer T d -MoTe 2 with electrons, the T c decreases until Δn = 2 × 10 13 cm −2 is reached, beyond which superconductivity disappears completely. They concluded that superconductivity disappears when the chemical potential is raised above the hole pocket, which means that both types of carriers are needed to produce the superconducting state. The coexistence of ferroelectric switching and superconductivity in a single material is a promising feature for developing a superconducting switch that can be driven by an external electric field. Exploring the transition between superconducting and normal states is expected to lead to low-power transistors and tunable qubits in the future. to the top source graphene electrode. c) A schematic band diagram across the vertical junction in the negatively poled photodetector. Reproduced with permission. [57] Copyright 2022, Wiley-VCH GmbH. d) Butterfly loops with coupled ferroelectric and superconducting states. e) Temperature-dependent ferroelectric and superconducting behaviors at a fixed charge carrier density. f) Carrier concentration as a function of Δn by fitting a two-band model to the Hall resistance. Reproduced with permission. [58] Copyright 2023, The Author(s), published by Springer Nature.

Ferroelectric Channels Based on Moiré Stacking
A breakthrough in achieving ferroelectricity has been reported, using stacked (non-ferroelectric) 2D layers. [59,137] In this section, we introduce emerging ferroelectric semiconductors, formed by stacking the same or different 2D materials as the active channel in FeFETs. As explained in Section 2, non-ferroelectric semiconductor TMDs or graphene can possess ferroelectricity in certain vertical heterostructures. We note that this emerging ferroelectricity is still in the initial stage of research, and many studies are focused on understanding the phenomenon. [115,138] Nevertheless, the promising aspect of this new ferroelectricity has been suggested for memory and synaptic device applications in recent research. [59][60][61][62]114] Two recent papers [59,114] reported an h-BN/bilayer graphene/h-BN system that shows ferroelectric features from the moiré lattice (Figure 22a). The two studies demonstrated ferroelectricity using transport measurements. As shown in Figure 22b,c, the charge neutrality point (Dirac point) of the graphene exhibits a hysteretic feature in the two direction-sweeps of the gate bias, [59] which could be considered strong evidence of the polarization flip. An interesting difference between the two studies [59,114] is the operating temperature; the ferroelectricity was demonstrated only at T = 4 K [59] and at room temperature [114] in the two studies.
In addition, an anomalous screening effect has been observed in a certain gate range in the ferroelectric active channel. [114] Various TMDs materials (e.g., MoS 2 , MoSe 2 , WS 2 , and WSe 2 ) have been investigated to realize ferroelectricity in vertically stacked structures [60,61] (Figure 22d-h). In contrast with the moiré lattice-driven ferroelectricity, the two studies used bilayers with the same TMDs, which exhibit orbital distortion in the stacked structures. Wang et al. confirmed ferroelectricity by PFM and measured the polarization using a graphene sensor, as shown in Figure 22e. [60] On the other hand, Weston et al. demonstrated ferroelectricity using a method of back-scattered electron channeling contrast imaging and KPFM and the ferroelectric bilayer was used as an active channel in their transport measurements (Figure 22f-h). [61] The most extensively studied TMDs, MoS 2 , exhibit ferroelectricity in the rhombohedral structure (in the bilayer geometry) and their ferroelectricity is robust up to room temperature. Besides two pioneering works, another important report on ferroelectricity from a similar origin was made with a bilayer of MoS 2 /WS 2 (heterobilayers). [62] In the study, tunneling transport and PFM were used to confirm the ferroelectricity, as shown in Figure 22i-l; the HRS and LRS could be observed with a ratio of ≈10 2 to 10 3 in the tunnel current, as shown in Figure 22k,l. This new type of robust ferroelectricity would provide a way to achieve Figure 22. Moiré stacking-based ferroelectric device channels. a-c) An h-BN/bilayer graphene/h-BN stacking device and demonstration of ferroelectricity in the device. Reproduced with permission. [59] Copyright 2020, The Author(s), published by Springer Nature. d) Stacked bilayer TMDs systems. The rhombohedral phase is made in the TMDs systems. Reproduced with permission. [60] Copyright 2022, The Author(s), published by Springer Nature. e) Ferroelectricity in the TMDs systems, measured by transport. f-h) Transport results of stacked bilayer MoS 2 systems in the rhombohedral phase. Reproduced with permission. [61] Copyright 2022, The Author(s), published by Springer Nature. i-l) PFM and transport results of a CVD-grown heterobilayer MoS 2 /WS 2 system. Reproduced with permission. [62] Copyright 2022, The American Association for the Advancement of Science. atomically thin memory and synaptic devices based on 2D materials.

Ferroelectric Tunnel Junction Devices
FTJ devices, with a thin ferroelectric potential barrier sandwiched between two electrodes, have been extensively investigated for non-volatile memory devices due to their promising performance with high data density, fast operation, and low operation energy. [47,[139][140][141] FTJ-based devices can non-destructively write and read information using the quantum tunneling effect. Figure 23a shows a crossbar memory array for high FTJ integration. [47,141] The crossbar architecture has the advantage of high density, but the sneak current issue still remains. [142] Thus, a conventional FeRAM contains one transistor and one ferroelectric capacitor, as shown in Figure 23b. Figure 23c shows the operating mechanism of the FTJ with a structure of metal/ferroelectric insulator/metal under external electric fields. Considering the Thomas-Fermi screening theory, the finite screening length of mobile charges in the metal electrodes in the FTJ produces incomplete screening of the bound charges at the interface with the ferroelectric layer. Therefore, a depolarization field is generated in the ferroelectric junction by the incomplete compensation (screening), providing a modified tunnel barrier height.
The tunneling current through the FTJ is determined by the barrier height (Δ ) which is proportional to the difference in screening length, shown in Figure 23c. [143,144] The distinct resistance states are generated by the direction of the FE polarization, which is called the tunneling electroresistance (TER) effect; the lowest and highest resistance state corresponds to the ON and OFF states in the FTJ device, respectively. In Figure 23c, the M2 electrode has a longer screening length compared to that of the M1 electrode. Thus, the polarization direction toward the M2 electrode accumulates electrons over a longer distance, which induces a larger band bending (compared to the band bending in M1). The larger band bending results in a lower average barrier height (the red dotted line in Figure 23c) than the initial one (grey dotted line in Figure 23c), which corresponds to the ON state. The opposite polarization direction toward M1, with a shorter screening length, creates the OFF state.
2D materials and their vdW heterostructures are promising to resolve the remaining issues in FTJs, such as incomplete screening of external electric field and the Fermi level pinning, Therefore, in this chapter, we will survey recently developed FTJbased memory devices with several structures including emerging 2D materials: M/FE/SC (metal/ferroelectric/semiconductor), M/vdW FE/Gr (metal/van der Waals ferroelectric materials/graphene) and FM/FE/FM (ferromagnet/FE/ferromagnet), as described in Figure 23c.

Ferroelectric Tunnel Junction Devices with Asymmetric Interfaces
An FTJ with an asymmetric structure of M/FE/SC achieves an enhanced ON/OFF ratio by effectively modulating the barrier height, as shown in Figure 24a. [145] In an n-type SC, electrons are accumulated by the bound charges in the FE layer at the interface between the FE and SC under a forward electric field. The accumulated electrons can tunnel through the FE layer, which makes the ON state. Under a reverse electric field, electrons are depleted at the interface and ionized donors in the SC screen the bound charges in the FE layer, which makes the OFF state. Accordingly, the high and low tunneling probability can be designed with asymmetric M/FE/SC junctions. A working mechanism similar to a p-type SC is described in the lower panels in Figure 24a. Compared with the FTJ with both metal electrodes, doping manipulation with the SC allows a larger TER modulation.
www.advancedsciencenews.com www.advelectronicmat.de Asymmetric synaptic FTJ devices with a structure of M/FE/SC have been reported with various semiconducting materials. For example, a non-volatile memory effect has been reported in the FTJ of Pt/HZO/Nb:STO with a high TER of 800. [41] Figure 24b shows that writing can be performed using voltage pulses and the resulting resistance state can be read with a voltage pulse of 0.1 V in the Pt/HZO/Nb:STO. The OFF state is changed to the ON state, when the amplitude of voltage pulses is larger than 2.7 V and is switched back to the OFF state when the amplitude of the voltage pulse is smaller than −2.7 V.
The calculated TER reached 80 000%, which is more than 100 times larger than symmetric M/FE/M devices. [71] Another study achieved non-volatile resistance switching using a semiconducting layer BTO with a high TER. [42] As shown in Figure 24c, the tunneling resistance of the FTJ increased from ≈1 × 10 3 to 3 × 10 7 Ω using different writing pulses varying from −1.2 to −4.4 V. As shown in Figure 24d, by applying consecutive negative pulses, the tunneling resistance of the device increased gradually from the ON state and then saturated at a certain value. On the other hand, with a series of positive pulses, the resistance continuously decreased from the OFF state. The dynamic, non-volatile resistance states shown in Figure 24d suggest synaptic functions such as LTP and LTD.
Ferroelectric vdW with graphene electrodes further enhanced TER from 10 3 to 10 7 . The giant TER in the FTJ with vdW ferroelectric and graphene electrodes was explained as follows: 1) the Fermi-level pinning effect becomes weak in the presence of graphene. In a metal/insulator heterostructure, Fermi-level pinning is known to easily occur due to the mid-gap states, which might create unexpected tunneling channels. However, in the presence of graphene, the Fermi-level pinning is removed and even modulated by the FE layer. 2) Graphene has an almost zero density of state near its charge neutrality point. This enables sufficient doping effect by the electrons for the ON state, and the holes for the OFF state.
The out-of-plane polarization of CIPS has been confirmed by PFM measurements, where the switching ratio exceeded 6 × 10 3 (Figure 25a,b). [43] Wang et al. reported a high TER of ≈10 7 with graphene electrodes, as shown in Figure 25c, [44] where writing pulse voltages with a period of 200 ms were applied to the Au electrode directly in contact with CIPS. The tunneling current was measured under a series of sweeping voltages. As shown in Figure 25d, when a 3.4 V pulse was applied, the device switched to the ON state at 10 −4 A, while −3.6 V pulses made the device return to the OFF state at 10 −11 A (i.e., the initial OFF state).
Wang's group discovered that the TER could exceed 10 10 by inserting a monolayer MoS 2 into the CIPS/graphene interface (Figure 25f). Since the exfoliated MoS 2 monolayer is an n-type semiconductor, the MoS 2 acts as a thin conductor with a certain polarization in CIPS (the left panel in Figure 25f), which does not significantly affect the tunneling process (ON state). However, in the other negative polarization (right panel in Figure 25f), the n-type MoS 2 is depleted and becomes an insulator, which acts as an additional tunnel barrier for the OFF state.

Multiferroic Tunnel Junction Devices
Multiferroic tunnel junction (MFTJ)devices are composed of two magnetic electrodes separated by a thin FE as the tunnel barrier and can be considered to be multi-functional devices. MFTJs operate based on the coupling between the magnetic configurations of the FM electrodes and the polarizations of the FE. Depending on the FM alignments (parallel or antiparallel), tunneling probability and tunneling resistance exhibit distinct states, which is called tunneling magnetoresistance (TMR), similar to TER, determined by the polarization direction.
By coupling TER and TMR, more resistance states can be demonstrated. As shown in Figure 26a, [46] four non-volatile states are available in conventional MFTJs, containing two TMR states and two TER states, arising from external magnetic and electric fields. Combining the functional properties of the TER and TMR effects, MFTJ devices read the multiple resistance states and detect the magnetoelectric effect, which is promising for nextgeneration multifunctional spintronics and memory devices.  (Figure 26b). [146] Eight stable and non-volatile resistance states were obtained, using magnetic shape anisotropies of the LSMO. As shown in Figure 26b, the top LSMO electrode has a magnetic crystalline anisotropy with an easy axis along the <110> direction and a hard axis along the <100> direction. On the other hand, the bottom LSMO electrode has an easy axis along the <110> direction and a hard axis along the [010] and [010] directions because it was designed along the [100] axis.
The resistance-magnetic field curves, corresponding to magnetism and polarization configurations, were obtained by applying a magnetic field along the [110] and [010] directions, which are the easy and hard axes of the bottom LSMO. Thus, stable resistance states were measured when the angle differences between ⃗ M top and ⃗ M bottom were 0°45°, 90°, and 180°, as shown in the lower panel in Figure 26b. The highest TMR ratio occurred at 0°and 180°with a value of 82%.
In 2017, Haung et al. fabricated LSMO/BTO/LSMO MFTJbased crossbar-structural memories as shown in Figure 26c. [47] Their multi-resistance states were achieved by electrical pulse voltage and the configuration of magnetizations (such as parallel and antiparallel). A series of pulse voltages of 0 V → +V p → −2.0 V → 0 V was applied to the MFTJ devices, and +V p max increased from +1.17 to 2.2 V for different series (Figure 26d). The highest MTJ device resistance was achieved at +2.2 V and the lowest state at −2.0 V, which had the largest TER of about 510%. In addition, after applying a pair of pulses with a reset voltage of −2.2 V and different positive +V p (+2.0, +2.1, or +2.2 V) with a duration time of 6 ns, memory loops were measured at 0.3 V under the magnetic field sweep, as shown in Figure 26e. The two different resistance states of each memory loop were determined by the parallel and antiparallel magnetizations of the top and bottom LSMO electrodes. This means that at least eight resistance states could be demonstrated in the unit cell of the crossbar array by combining electric and magnetic effects. Reproduced with permission. [43] Copyright 2020, Wiley-VCH GmbH. c) FTJ with Au/CIPS/Graphene. d) The ON (red) and OFF (blue) tunneling currents. Reproduced with permission. [44] Copyright 2022, Elsevier B.V. e) Band diagram of Au/vdW FE/Graphene. f) Band diagram of Au/CIPS/MoS 2 /Graphene. Reproduced with permission. [45] Copyright 2022, Elsevier Inc.
FTJs based on traditional perovskite ferroelectric materials have known issues with the critical thickness of the FE, high contact area resistance, and defective non-uniform interface, which limit their applications. However, MFTJs with 2D vdW materials have solved the critical thickness issue of perovskite.
Chen et al. calculated the performance of a Ni/bi-layer In 2 Se 3 /h-BN/Ni (NIBN) device and found enhanced TER and TMR by inserting h-BN as a buffer layer (Figure 27 (a)). [48] For conventional MFTJs, four resistance states were implemented using the polarization and magnetization directions of the FE and FM, respectively. However, four distinct electrical states (without magnetic states) could be obtained by the polarization configuration of each -In 2 Se 3 layer in the study. [48] TMR and TER combinations were found to have eight distinct states using different magnetization configurations of the FM in the NIBN device, as shown in Figure 27b. Having more states enables more multi-level memristors. Due to the presence of h-BN, the interaction between the second In 2 Se 3 layer with Ni was weaker than that of the first In 2 Se 3 layer with Ni. Thus, the h-BN layer created an asymmetric electrode environment, which induced changes in the interfacial charge distribution and internal potential, improving TER.
Further enhanced TMR could be designed using minority spin channels. Larger differences in minority spin injection were observed in the NIBN junction-based device, compared to a NIN (without h-BN) junction-based device, as the Ni electrodes were switched from M ↑↑ to M ↑↓ . In the FE polarization configuration  [46] Copyright 2012, Materials Research Society. b) Schematic of an LSMO/BTO/LSMO device and an H-R curve, depending on the magnetization of LSMO electrodes. Reproduced with permission. [146] Copyright 2015, WILEY-VCH GmbH. c) Schematic of an LSMO/BTO/LSMO crossbar array and d) its unit cell change of high resistance states according to maximum of V p under a series pulse 0V → V p → −2.0V → 0V. e) Multilevel states of the LSMO/BTO/LSMO device with different maximum voltages under magnetic sweep. Reproduced with permission. [47] Copyright 2018, WILEY-VCH GmbH.  [48] Copyright 2022, The Royal Society of Chemistry. c) Atomic structures of PtTe 2 /Fe 4 GeTe 2 /In 2 Se 3 / Fe 3 GeTe 2 /PtTe 2 . Reproduced with permission. [49] Copyright 2021, American Chemical Society.
for the largest TMR (P → → in NIBN and P ← → in NIN), a ten times larger change in the transmission coefficients of the minority spin in NIBN led to a TMR that was ten times larger than that of NIN. Accordingly, using 2D materials to improve TER and TMR in MFTJ-based devices is considered promising for multi-level state memristors.
Su et al. investigated the spin-dependent transport properties of vdW MFTJ with a 2D ferromagnetic Fe n GeTe 2 (FGT) and ferroelectric In 2 Se3 barrier layers. [49] Figure 27c shows an MFTJ with PtTe 2 /Fe 4 GeTe 2 /In 2 Se 3 /Fe 3 GeTe 2 /PtTe 2 , used in the study with two important parameters: polarization of the ferroelectric In 2 Se 3 and two magnetic configurations of the FGT. As shown in Table 1, non-volatile resistance states, determined from the two ferroelectric polarizations and parallel or antiparallel magnetic orderings, exhibit a TMR of up to 89%.
In conventional MFTJs based on perovskites, LSMO has primarily been used as the ferromagnetic electrodes, despite the difficulty of controlling their spin polarization at high temperatures. To make practical MFTJs for synaptic devices, however, will require new ferromagnetic materials for electrodes that can operate at room temperature in a reliable way. In addition, additional multiple states should be conceived using 2D materials, so that highly integrated synaptic devices can be developed based on spin and FE polarization in energy-efficient ways.

Comparison and Benchmark of 2D Ferroelectric Devices
In this chapter, we summarize key aspects, benchmarks, and comparisons of the three types of devices described in Sections 3, 4, and 5. Conventional FeRAM has a structure of 1T-1C (one transistor and one capacitor), which includes ferroelectric materials  in the capacitor. FeRAM inevitably exhibits a destructive reading process that induces short retention time; thus, alternative device architectures, FeFET and FTJ, have been suggested as promising ferroelectric devices. FeFET and FTJ have partly resolved the retention time issue through their non-destructive reading process, but reliability (or endurance) issues appeared in the two types of device operation. For example, depolarizing fields and leakage (and sneak) currents limit the device operation in FeFET and FTJ, as we discussed in Section 2.1.1. Various 2D materials have been proposed to resolve the complex issues in ferroelectric devices. We could minimize interface effects that degrade device performances by utilizing 2D materials in FeRAM, as shown in Table 2. In the case of Fe-FET, 2D materials have been newly used as channel materials for improved device performances. In addition, a high ON/OFF ratio could be obtained in FTJ with a tunneling barrier based on 2D ferroelectric materials. We note the trade-off between the thin geometry and sneak current in the use of the 2D fer-roelectric tunnel barrier. FeSFET is a new type of ferroelectric device based on emerging 2D materials. Long-standing issues in FeFET and FTJ, such as depolarization and leakage current, have been expected to be resolved by the new device architecture.
The high polarization in 3D ferroelectric materials could be achieved by heterostructures of 2D ferroelectric and non-ferroelectric materials. [53,54] Such heterostructures have shown novel device operations, particularly with optical signal processing. [53,54,57] Despite the intriguing device features, we note that 2D ferroelectric materials-based devices are still premature, and thus, cannot yet exhibit better performances than traditional 1T-1C structure devices. We compared their performances in Table 2. Based on the advantages of 2D ferroelectric materials, more optimized combinations for heterostructures are required.
Integrating 2D ferroelectric materials-based devices remains another critical issue. On the laboratory scale, exfoliated flakes www.advancedsciencenews.com www.advelectronicmat.de from high-quality single-crystalline layered bulk materials have been used in most studies. While there has been extensive research in high-quality, wafer-scale synthesis of 2D functional materials, fabricating highly-performing devices based on largescale 2D materials has been limited for simple field-effect transistor applications (e.g., mobility and contact resistance characterization). Thus, we require large-scale ferroelectric 2D materials with fewer defects, better uniformity, and high stability, which can largely improve the device's performances. In addition, we note that optimized transfer processes of 2D materials on arbitrary substrates are critical for industrial applications. Cracks and wrinkles, formed during transfer processes, seriously degrade ferroelectric features in 2D materials, which should be resolved in practical applications.

Conclusion and Outlook
We have explored the unique functions and working mechanisms of memory and synaptic devices based on 2D ferroelectric materials. The 2D ferroelectric materials provide clear benefits of high integration and low operating energy in the device structures of FeFETs, active ferroelectric channels, and FTJs. In addition, as interest in neuromorphic technology continues to grow, 2D ferroelectric materials are promising for synaptic devices with multiple conductance states beyond conventional memory cells.
The key to the multiple conductance states is the presence and control of local FE, called FE domains. The local FE domains can be switched using a fast write process involving spike signals with various voltages and time ranges. [35,38,47,126] Further attempts to engineer smaller-scale FE domains for highly-integrated neuromorphic devices are expected to encounter the domain wall issue, which has long been a critical scientific subject. Manipulating the local features and dynamics of domain walls is a bottleneck in improving the integration density and synaptic performance of devices based on 2D ferroelectric materials. Extensive recent fundamental and engineering studies have suggested that 2D ferroelectric materials with domain walls are promising for diverse conductance states in atomically thin and narrow geometry. [29,[51][52][53][54] Accordingly, we conclude our review by discussing the new aspects of emerging 2D ferroelectric materials: small FE domains and their boundaries (walls). Until now, three types of conductance engineering using FE domain walls have been clarified as follows; [147,148] 1) Domain walls change electronic conduction by attracting or repelling point defects created by donor or acceptor states in the bandgap. [149,150] 2) The intrinsic electronic band structures of domain walls are locally different from the averaged bulk electronic structures. [151,152] 3) At domain walls, the polarization is discontinuous, so that the local carrier distribution is modified. [153,154] We believe that, besides the above mechanism, there could be more effective and feasible ways to achieve various conductance states based on the FE domain engineering of 2D materials. Thus, we encourage researchers to investigate and exploit memory and synaptic devices based on the emerging materials group.