A minireview on 2D materials-enabled optoelectronic artificial synaptic devices

Current computing systems based on the von Neumann architecture have been rapidly developing for over half a century to meet increasing demands on data processing and mathematical calculations.^1–4 As integrated circuits are continuously becoming denser following the prediction from Moore’s law, the number of transistors incorporated in them is increasing exponentially toward fulfilling the growing data quantity.⁵ However, with increasing demands for handling “big data” nowadays, computing hardware requires the capability of storing and processing a massive amount of complex information in an accelerated manner, which leads to the exploration of deep learning and artificial intelligence (AI) with lower power consumption.⁶ Remarkable technological advancements, such as autonomous vehicles and human-interactive robotics, governed the prerequisite of swiftly mimicking visual perceptions. For example, autonomous vehicles require optoelectronic sensory devices to control, navigate, and drive them without a human presence, wherein the devices are able to learn, memorize, and process visual perceptions like a human retina. Conventional complementary metal–oxide–semiconductor (CMOS)-based optoelectronic devices operating by the von Neumann principle transmit accrued visual data to separately located processing/storing units like cloud computers.⁷ Accordingly, their performances are intrinsically limited due to the physical bulkiness of computing and memory units as well as their inevitable physical separation.^2–4 Furthermore, they process and transfer the data in a sequential manner through communication but often cause process latency and colossal energy consumption, which becomes more problematic in processing complex visual data, such as speech, image, and video.⁸ To overcome this von Neumann architecture’s bottleneck toward drastically improving optoelectronic device performances, extensive research has been put through in various directions. Some of them include the extreme scale-down of constituent transistors toward their physical size limit—i.e., near-atomic thickness—achieving a more energy-efficient operation. In addition to this physical size scaling approach, the ingenious idea of mimicking the functional architecture of the human brain and implementing it into computing hardware, namely, neuromorphic computing, has recently emerged. The human brain consists of a complex network of ∼10¹¹ neurons connected by ∼10¹⁵ synapses, achieving an unparalleled capability of computing.⁹ In addition, each neuron acting as a receiving/integrating unit for electrochemical signals enables to collect and analyze surrounding information with only ∼20 W of power consumption.^10,11 The neuromorphic systems artificially mimicking the human brain’s neurons and synapses are beneficial in handling a massive amount of complex data, such as self-learning and object recognition.¹² The implementation of such neuromorphic functions is enabled by the conversion of electrical and/or optical signals, mostly achieved with CMOS-based transistors. Particularly, optoelectronic synaptic devices that bi-directionally record optical inputs and convert them into electrical signals hold great potential for learning and pattern recognition, as their process is akin to the optic nerve in human eyes.¹² Earlier initiatives for this approach explored optic-neural synapses by integrating prepared synaptic electrical devices into optical-sensing devices separately. In this device structure, pattern imaging units are physically separated from pattern recognition circuitries resulting in an unwanted bulkiness of integrated systems. In addition, these optoelectronic synapses are limited in implementing neuromorphic functions as they do not respond to light stimuli directly.¹³ A variety of opto-electrical materials were explored as switching media in optoelectronic synapses to alleviate these limitations. Particularly, chalcogen-based phase-change materials were demonstrated for optoelectronic pattern recognitions by realizing a complete set of essential synaptic characteristics,^14,15 i.e., short-term plasticity (STP), long-term plasticity (LTP), long-term depression (LTD), and spike-timing-dependent plasticity (STDP). Furthermore, conventional light-sensitive semiconductors were also employed in optoelectronic synaptic devices triggered by various working mechanisms, including charge trapping/detrapping, ions/vacancies migration, and floating-gate.^11,16 Nevertheless, device performance issues, such as immense energy consumption and demand for complex broadband wavelength conversion, remained unresolved mainly due to intrinsic property limitations of these conventional materials, i.e., physical bulkiness, narrow wavelength coverage, indirect bandgap, weak light–matter interaction, and surface/interface defects.^17,18 Recently, two dimensional (2D) materials with anisotropic van der Waals (vdW) molecular bonding and their heterostructure variations have drawn significant attention owing to their intrinsic properties particularly suitable for optoelectronic synaptic devices; these include near atom thickness, strong light–matter interaction, superior electronic and optical tunability, defect-mediated surface, low voltage maneuver, and excellent mechanical robustness/flexibility.^12,14,19,20 In fact, resistive switching-driven optoelectronic synaptic characteristics were realized with various 2D materials, operated by a variety of working principles including filament formation, charge trapping/detrapping, and ions/vacancies migration.^12,18,21

This Research update overviews recent progress in 2D materials-enabled optoelectronic synaptic applications, covering a range of 2D materials explored for them as well as their working principles, as illustrated in Fig. 1. The 2D material property suitability for the synaptic devices is discussed, and their operational principles are reviewed in detail, followed by the introduction of up-to-date device demonstrations. Finally, we share our perspectives on the current challenges and forward-looking outlooks in this research field.

In the traditional computing systems based on the von Neumann architecture, memory and computational units are physically separated and are connected by a data bus, as illustrated in Fig. 2(a).^2–4 The “von Neumann bottleneck” arises from the requirement for using this data bus, which makes the device architecture significantly inefficient when dealing with complex data.^2–4 This fundamental limitation is becoming more significant as machine vision technologies are now becoming a fundamental part of various AI systems. Neuromorphic computing is to mimic the human brain, which performs pattern recognition with remarkable energy efficiency keeping the memory and process units in the same physical location, as illustrated in Fig. 2(b). Optoelectronic synapse is a new concept of emerging neuromorphic device technologies with tremendous potential to overcome the limitation of von Neumann architecture.^4,12,19 It combines a photodetector and an electronic synapse in a single device platform, just like the optic nerves of a human retina, thus implementing data perception, processing, and memorization in one location without requiring the data bus, as illustrated in Fig. 2(c). Optoelectronic synaptic devices exhibit multiple nonvolatile conductance states when subjected to pulses of light due to light–matter interaction—analogous to a photodetector with memory capability. Similar to electronic synapses, they can mimic all essential biological characteristics, including STP, LTP, and STDP, upon optical stimuli in the form of pulses.^14,15 STP and LTP are distinguished by current decay characteristics upon input signals, i.e., STP represents that the post-synaptic conductance state rapidly relaxes back to its initial conductance state upon an application of a weak input signal, while LTP represents the gradual decrease of the post-synaptic conductance, which eventually maintains a high conductance state above the initial state upon an application of a strong input signal.^22,23 In addition, STDP defines the intercoupling between pre-synaptic and post-synaptic activities owing to the relative timing of their actions, manifested by the change in synaptic weights.¹³ Furthermore, the devices can be integrated into neural network hardware for pattern recognitions responding to multiple wavelengths of ultraviolet (UV) to infrared (IR) range.^14,24

Efficient optoelectronic synaptic devices require a large set of performance metrics, such as high bandwidth, low crosstalk, substantial data processing speed, and low energy consumption.²⁵ With that objective, conventional metals or metal oxides have been mainly employed as the switching media in the devices where the photoexcited carrier trapping/detrapping was modulated by various means—e.g., the floating gate, the density of oxygen vacancy and intrinsic defects, and channel/gate oxide interface.^24,26,27 However, these bulk materials-based neuromorphic optoelectronic synapses are far from overcoming the von Neumann bottlenecks, such as high power consumption and process latency, due to their constrained optical, electronic, and mechanical properties. For example, conventional optoelectronic synapses have been demonstrated with CMOS-based memristive devices employing phase-change memory (PCM) or resistive random-access memory (RRAM). These devices exhibit multi-state conductive states—essential characteristics for neuromorphic computation—due to the intrinsic characteristics of their switching media materials. However, their device performances are still significantly limited in realizing the true potential of optoelectronic synapses, mainly due to restrictions associated with their switching media materials, e.g., their nonlinear/asymmetric synaptic conductance due to abrupt on/off switching as well as high operating currents increase the energy consumption per switching event. These drawbacks are attributed to the limited properties of conventional switching materials such as bulky phase-change semiconductors or metal oxides containing a large concentration of uncontrolled structural and chemical variations. Intrinsically layered 2D materials exhibit unprecedented advantages for alleviating these limitations^28,29 and, thus, are highly suitable for realizing high-performance optoelectronic synaptic devices. From the structural property point of view, their extremely small thickness (typically, 10 Å only) offers short pathways for the electric field-driven diffusion of charge carriers, essential for rapid electrical switching. Given the energy consumption is proportional to the volume of the active device area in optoelectronic synapses,¹¹ 2D materials are projected to be highly energy-efficient due to their extremely small volume for their large surface area.^11,16 2D basal planes possess dangling-bonds saturated surfaces with minimum chemical variations, which ensures high controllability of spatially defined conductive pathways within switching media. This unique aspect of defects-mediated surfaces inherent to 2D materials enables their heterogeneous integration forming unique vdW heterostructures overcoming the lattice match constraints in thin-film technologies.^11,30,31 This vdW heterostructure assembly scheme can pave the way for bandgap engineering toward enhanced optical or electronic properties.^11,16,32,33 For example, mechanically exfoliated bare graphene exhibits a mobility of ∼10 000 cm² V⁻¹ s⁻¹, whereas the identical graphene encapsulated with hexagonal boron nitride (h-BN) exhibits mobility up to 100 000 cm² V⁻¹ s⁻¹.^32,33 Furthermore, the large in-plane strain limit and mechanical stiffness of 2D materials offer opportunities for exploring mechanically reconfigurable optoelectronic devices of various form factors.^17,27,34,35 From the optoelectronic property point of view, 2D materials are uniquely suited for optoelectronic synapses due to their highly tunable photoresponse in a wide range of wavelengths as well as superior carrier dynamics.^36,37 For example, the confinement of electron and hole (e-h) pairs in 2D materials enhances the Coulombic interactions due to the reduction of dimensionality and dielectric constant.^38,39 Subsequently, it can lead to enhanced exciton binding energy/lifetime and light–matter interaction⁴⁰ accompanying quantum confinement-driven carrier transports.^41,42

In general, optoelectronic neuromorphic devices adopt light stimuli as input signals in modifying their electrical conductance to control and tune resulting synaptic responses.^21,43 The primary working principle for a majority of 2D materials-based optoelectronic synaptic devices relies on the charge trapping/detrapping phenomena within the 2D materials functioning as channels or at/near their interfaces.¹¹ Various factors govern these charge transport behaviors, such as the density of charge trapping sites at the interfaces of 2D materials/substrates and charge transfers from/to neighboring adsorbates (e.g., oxygen ions from the environment).^44,45 2D materials-enabled optoelectronic synaptic devices can fall into following categories broadly defined by the types of components contributing to the charge transports-driven switching characteristics, i.e., (1) intrinsic 2D material channels with/without functional oxides interface, (2) nanoparticles (NPs) or quantum dots (QDs)-incorporated 2D materials, (3) 2D vdW heterostructures, and (4) 2D ferroelectric materials.¹¹ Figure 3(a) shows schematic illustrations of their working principles, i.e., charge trapping/detrapping within 2D materials or polarization control by functional oxides, electron–hole pair generation by NPs or QDs interfaces, charge trapping/detrapping at 2D vdW hetero-interfaces, and polarization switching by ferroelectric 2D materials. Details will be discussed in Secs. III A–III D.

Intrinsically semiconducting 2D materials responding to light stimuli of various wavelengths have been explored as channel materials in phototransistors that realize optoelectronic synaptic characteristics.^{13,46,48,58,61} For example, He et al. utilized the chemical vapor deposition (CVD)-grown monolayer 2D MoS₂ in n-2D MoS₂/p-Si heterostructures in realizing ultrathin memristive synapses.⁴⁶ The monolayer 2D MoS₂ employed in the heterojunctions efficiently trap/detrap the photogenerated charges there, exhibiting optical and electric synaptic functions.⁴⁶ 2D layers combined with functional oxides have been extensively studied to achieve optoelectronic artificial synapses. The oxide materials offer intrinsic charge trapping sites when interfaced with the 2D layers or exhibit their own optical tunability suitable for the conversion of optical-to-electrical signals—e.g., polarization orientation control inherent to ferroelectric oxide materials. Oxide interfaces have been employed in conventional neuromorphic photonic devices as they can lead to the persistent photoconductivity phenomenon, i.e., retention of prolonged photogenerated carriers even after a termination of input, which is essential for emulating neuromorphic functions.⁶² The persistent photoconductivity phenomenon is attributed to the charge trapping/detrapping mechanism resulting from various factors, e.g., oxygen vacancies within the oxides or absorption of water/oxygen molecules on the oxide surface.^63,64 For example, Kim et al. developed germanium (Ge)-gated 2D molybdenum disulfide (MoS₂) phototransistors performing both optical sensing and synaptic operations on a single device that responds to the visible-to-IR (520–1550 nm) wavelength, as shown in Fig. 3(b).⁴⁷ In this device geometry, Ge functions as the gate material that absorbs IR light and induces band shifting due to its small bandgap energy, while 2D MoS₂ layers detect visible light as its bandgap energy matches the wavelength. Negative gate pulses decreased/increased the drain current within the 2D MoS₂ channel, reflecting the charge trapping/detrapping in it, respectively. The trapping sites at the interfaces of 2D MoS₂/silicon dioxide (SiO₂) lead to the positive/negative shifts of threshold voltage, V_th, via the photogenerated charge trapping/detrapping, respectively, as shown in Fig. 3(c).⁴⁷ Islam et al. reported optoelectronic synaptic characteristics in monolayer 2D MoS₂ field-effect transistors (FETs),¹³ where the photogenerated carrier trapping/detrapping at 2D MoS₂/SiO₂ interfaces was employed to emulate the synaptic characteristics. In these devices, the photogenerated holes at the interfaces enable optically synaptic operations, which can be further modulated with an electrically applied gate bias.¹³ In addition, Ahmed et al. reported optoelectronic characteristics in few-layer black phosphorus (BP) by exploiting its oxidation-related defects.⁴⁸ As the BP layers exposed to ambient conditions adsorb oxygen from the environment forming phosphorus oxide (PO_x) layers, these defective layers can easily trap charges to imitate excitatory and inhibitory post-synaptic action potentials. The thin amorphous PO_x layers from the natural oxidation of BP layers due to their ambient preparation conditions led to distinct photoresponse with varying wavelengths. Synaptic weights of these few-layer BP devices were modulated only through optical stimulation without employing electrical input signals.⁴⁸ In addition to the naturally formed oxides, functionally engineered ferroelectric oxides are also suitable as optoelectronic synaptic components due to their intrinsic optical or electrical tunability of polarization orientation. This optoelectronic modulation can be employed to introduce optically driven memristive switching in 2D layers, achieving optoelectronic neuromorphic characteristics.

Incorporating optically active NPs into 2D layers can improve optoelectronic performances by creating a large density of charge trapping sites at 2D layer/NP interfaces.^50–52 Pradhan et al. demonstrated optoelectronic synaptic characteristics in a hybrid structure of perovskite quantum dots (PQDs)-incorporated graphene, as shown in Fig. 3(d).⁵⁰ The photogenerated electrons trapped in the photo-absorbing PQDs lead to a negative photogate effect by inducing more hole carriers in the carrier transporting graphene channel through capacitive coupling.⁵⁰ This hybrid structure was fabricated by directly growing PQDs on the graphene lattice by a defect-mediated crystal growth technique and analyzed by transmission electron microscopy (TEM), as shown in Fig. 3(e).⁵⁰ Figure 3(e) shows the spherical shape of the PQDs grown on graphene.⁵⁰ Sun et al. incorporated heterostructure semiconductor QDs composed of cadmium selenide (CdSe)/zinc sulfide (ZnS) into 2D MoS₂ layer-based transistors.⁶⁵ Under 470 nm wavelength illumination, photogenerated electrons transfer into the 2D MoS₂ channel, whereas photogenerated holes become trapped in the QDs due to their distinct bandgap energy and defect states. This asymmetric transport increases the electron–hole pair recombination time, leading to the persistent photocurrent utilized for achieving STP and LTP.⁶⁵ Furthermore, Ni et al. reported hybrid optoelectronic synaptic devices based on 2D tungsten diselenide (WSe₂) layers and a thin film of silicon (Si) NPs—film thickness: ∼1.53 nm and NP size: ∼6 nm.⁵¹ The strong broadband optical absorption resulting from boron (B)-doped Si NPs in a wide range of wavelengths—i.e., from UV to near-infrared (NIR), was combined with the efficient charge transport of 2D WSe₂ layers. Specifically, the photogenerated holes from the B-doped Si NPs were efficiently transferred to the 2D WSe₂ layers upon 1342 nm laser spikes.⁵¹ The devices yielded optically stimulated synaptic responses, such as STP, LTP, paired-pulse facilitation (PPF), STDP, as well as other essential synaptic characteristics, accompanying a low energy consumption of ∼75 fJ.⁵¹ The underlying principles for the broadband optical absorption in the Si NPs were attributed to the enhanced plasmonic effects due to their reduced dimension.^66–68 

2D vdW heterostructures composed of vertically stacked heterogeneous 2D layers with relaxed lattice mismatch and epitaxial constraints have been projected to exhibit intriguing electrical and optical properties.^45,69,70 The charge trapping/detrapping occurring at the interfaces of distinct 2D layers inherent to these vdW heterostructures were employed to emulate optoelectronic synaptic behaviors. For example, Seo et al. reported an artificial optic-neural synapse for color and color-mixed pattern recognition by implementing synaptic and optical-sensing functions together on vdW heterostructures composed of h-BN/WSe₂, as shown in Fig. 3(f).¹⁴ In this device geometry, an h-BN/2D WSe₂ photodetector was connected with a 2D WSe₂/weight control layer (WCL)/h-BN device in series. The WCL modulating the charge trapping/detrapping on the surface of h-BN was formed by its surface treatment with oxygen (O₂) plasma treatment, as shown in Fig. 3(g).¹⁴ In addition, Qin et al. developed hybrid-phototransistors composed of graphene and atomically thin single-walled carbon nanotubes (SWNTs), demonstrating light-stimulated synaptic behaviors.⁵² The devices incorporating SWNT films sandwiched between graphene and SiO₂ emulated the depression plasticity of biological synapses successfully, i.e., STP and LTP, well as exhibiting nonvolatile memory characteristics.⁵² These optoelectronic synaptic behaviors were attributed to the positive/negative photoresponse of the hybrid graphene/SWNT layers, which were further employed to enhance learning effects by high-repetition training pulses.⁵² Moreover, Wang et al. reported multifunctional synaptic transistors based on 2D MoS₂/perylene-3,4,9,10-tetracarboxylic dianhydride (PTCDA) hybrid heterojunctions by exploiting the band alignment of the constituent materials.⁵³ The trapping/detrapping mechanism-driven electron transfers at the interfaces of top PTCDA/bottom 2D MoS₂ layers yielded STP and LTP synaptic behaviors as well as enhanced PPF ratio.⁵³ In addition, Kim et al. demonstrated optoelectronic memory devices employing 2D WSe₂/graphene vertically stacked heterostructures with 2D MoS₂ layers as the top floating gate.⁵⁴ In these devices, electrons or holes are allowed to be trapped in the 2D MoS₂ floating gate by the polarity of gate voltage pulses, accompanying rectified nonvolatile memory properties.⁵⁴ Specifically, upon repeated 532 nm laser pulses, the tunneling of electrons or holes between the 2D WSe₂ channel and the 2D MoS₂ floating gate was observed, manifested by the positive/negative staircase-type photoconductivity change.⁵⁴ In addition, recently, Islam et al. reported a FET structure with IR-sensitive PtTe₂ as a buried gate electrode and MoS₂ as the channel and demonstrated a multi-wavelength optoelectronic synapse that operates from 300 nm in UV to 2 μm in the IR spectrum.²² 

2D materials with intrinsic ferroelectricity or combined with ferroelectric oxides provide an additional degree of freedom to control synaptic characteristics.^58,71 Their ferroelectric polarization can be temporarily or permanently modulated, which can tailor their carrier transports to achieve targeted synaptic properties.^58,71 For example, Luo et al. demonstrated artificial optoelectronic synapses by integrating 2D tungsten disulfide (WS₂) layers on ferroelectric gate dielectric PfZr_0.2Ti_0.8O₃ (PZT) epitaxial thin films (∼5 nm) on SrRuO₃ (SRO) conducting layer buffered SrTiO₃ (STO) substrates.⁵⁸ In this device geometry, the PZT films function as the underlying ferroelectric gate dielectric that is optically and electrically tunable, while the 2D WS₂ layer channels exhibit voltage- and light-controllable memristive characteristics.⁵⁸ The optical absorption-driven ferroelectric polarization switching in the PZT films assists in the accumulation of positive charges in the 2D WS₂/PZT interfaces, causing the charge trapping/detrapping-enabled synaptic functionalities. The change in the polarization was observed to be more pronounced with increasing the optical illumination time. In addition, the higher responsivity achieved with the optical illumination of shorter wavelengths was demonstrated to induce more robust charge accumulation at the interfaces for faster polarization switching.⁵⁸ These 2D layer/ferroelectric heterostructures were demonstrated to exhibit STP and LTP transition by modulating device parameters, such as the light power and the number/frequency of electrical pulses.⁵⁸ In addition to the 2D layer/ferroelectric heterostructures, 2D materials with intrinsic ferroelectricity were also pursued as the stand-alone media to realize optoelectronic synaptic characteristics. Guo et al. demonstrated optoelectronic synaptic devices using intrinsically ferroelectric 2D indium(III) selenide (α-In₂Se₃) and gallium(II) selenide (GaSe) vdW heterostructure.⁷¹ Their distinct optical absorption due to distinct bandgap energies was employed for achieving plasticity characteristics and the change in their polarization orientation also led to distinguishable photocurrents.⁷¹ 

A variety of 2D materials-based optoelectronic synaptic devices have been fabricated and reported. One essential feature for efficiently mimicking biological synapses is their ability to demonstrate STP and LTP, as these two characteristics are directly related to the learning process in the human brain for achieving short/long-term memory. The basic operational principle for achieving such characteristics by optoelectronic means is to incorporate synaptic characteristics into photodetectors. The resulting optoelectronic synaptic devices function in a similar manner to an artificial retina, and their application ranges from pattern recognition to AI. This section will focus on introducing up-to-date developments of 2D materials-enabled optoelectronic artificial synapses and reviewing their device characteristics and performances.

Islam et al. demonstrated three-terminal optoelectronic synaptic FET devices by employing the CVD-grown 2D MoS₂ monolayer as the conductive channel and the SiO₂/MoS₂ interface as the charge trapping/detrapping site, as shown in Fig. 4(a).¹³ The charge trapping/detrapping at the interface was modulated with an applied gate bias, achieving the distinct operation of the FET device, i.e., it was operated as an optoelectronic synapse at a negative gate voltage and as a photodetector at zero or positive gate voltage. The persistent photocurrent characteristics of the devices were implemented to successfully emulate essential optoelectronic characteristics, including STP, LTP, PPF, and STDP. Transfer characteristics of the 2D MoS₂ FET device exhibited n-type behavior with an on/off current ratio of 10⁵ in the dark and under optical illumination. Particularly, the photogenerated carriers resulting from the illumination (450 nm wavelength) caused a negative shift in V_th and increased the drain current, confirming the photogating effect of the FET device [Fig. 4(b)].¹³ The device did not retain the photocurrent in the dark, but it was potentiated with five light pulses at V_D = 1.0 V and V_G = −2.0 V and was able to retain its conductance state [Fig. 4(c)]. The retention of the device was observed for at least 10⁴ s even after the light pulses were withdrawn. Furthermore, the optical tunability of the potentiation and conduction retention was demonstrated with another gate dielectric, i.e., aluminum oxide (Al₂O₃), and similar behaviors were observed [Fig. 4(d)].

In addition to the demonstration of optically driven electrical plasticity, combining the mechanical and optical plasticity can synergistically offer additional opportunities for mixed-modal neuromorphic devices. This approach of incorporating biomechanical motion into visual information in mimicking the perception and cognition ability of the human brain is directly applicable to AI systems for interactive human–machine interfaces, artificial retinas, and robotics.⁷² Figure 4(e) illustrates the conceptual demonstration of a biological tactile/visual sensory system employing the mechano-photonic artificial synapse. Yu et al. reported bioinspired mechano-photonic synapses employing graphene/2D MoS₂ vdW heterostructure devices [Fig. 4(f)].⁵⁵ In this device structure [the inset in Fig. 4(f)], a graphene/2D MoS₂ vdW heterostructure is integrated onto a triboelectric nanogenerator (TENG) that generates the Maxwell displacement current as the driving potential to modulate and harness optoelectronic synaptic plasticity, i.e., the mechanical displacement is generated from the TENG composed of copper (Cu)/polytetrafluoroethylene (PTFE)/Cu and PTFE/Cu layers connected to the transistor gate [Fig. 4(f)]. Upon optical illumination, the photogenerated charge carriers from the 2D MoS₂ layers are transferred to the top graphene layer. This leads to the generation of persistent photocurrent as the graphene/2D MoS₂ interface barrier efficiently prevents the electron–hole recombination [Fig. 4(g)]. Figures 4(h)–4(j) illustrate the working mechanism of the mechano-optoelectronic synapse under various operational conditions. The initial state (D₀) of the device is kept at electrostatic equilibrium by grounding the transferred charge between the Cu electrode and the PTFE/Cu friction layers [Fig. 4(h)], and there is no triboelectric potential at the initial state. An induced −V_TENG is coupled to the transistor gate, which shifts the Fermi level of the graphene downward, leading to its electrostatic doping with holes once the two friction layers are separated at D₊ distance [Fig. 4(i)]. Under light illumination and with the separation of TENG until −V_TENG < V_T, the photogenerated electrons are ejected from the conduction band of 2D MoS₂ layers into the low energy level of the hole-doped graphene, leading to the resistance increment of the graphene. Conversely, when the two friction layers are brought closer (D₋), +V_TENG is induced and coupled to the transistor gate due to the excessive positive charge, and the 2D MoS₂ layers exhibit the near-metallic behavior (gating by +V_TENG > V_T). Finally, the excess electrons on the 2D MoS₂ layers travel to the graphene and equilibrate the carrier distribution in the graphene/2D MoS₂ heterostructure [Fig. 4(j)]. Multimodal plasticity in the mechano-photonic artificial synapse can be modulated by mechanical displacement (D+/D−) and light illumination. Figure 4(k) provides the negative increment of the photo-activated post-synaptic current (−∆PSC) resulting from the synergic effect of mechanical and optical modulation of the device under a fixed D at 1 mm. The peak value of −∆PSC increases from 26 to 35 µA, where LED power intensity (P_LED) increases from 0.73 to 13.5 mW cm⁻². Under the dark condition, a slight decrease of the current is observed, which confirms the repeatable LTD synaptic behaviors of the device [Fig. 4(l)]. In addition, the influence of multiple consecutive light pulses on the mechano-photonic artificial synapse under various displacements was also verified [Fig. 4(m)].

In addition to the main stimuli inputs such as optical illumination and mechanical displacements as discussed above, various factors also can influence optoelectronic synapses characteristics; these include operational environment, gate dielectric, gating medium, and 2D layer preparation methods. For instance, influences of operational environmental and gate dielectrics on optoelectronic synaptic characteristics were studied with 2D MoS₂/SiO₂/Si FET-based devices, as shown in the scanning electron microscope (SEM) image in Fig. 5(a).⁴⁵ Specifically, the environment-driven adsorption of charge carriers on optoelectronic synaptic characteristics was verified with the device operating in ambient and vacuum conditions. The devices operating in the air exhibited strong hysteresis loops in their transfer characteristics both in the dark and under illumination as opposed to those operating in the vacuum, which was attributed to the absorption of water and oxygen molecules from the environments. When the negative gate voltages (V_G < 0) were applied, the device conductance increased for each incident light pulse, resembling the potentiation of a synapse, as shown in Fig. 5(b).¹³ The devices still emulated all essential characteristics of optoelectronic synapses irrespective of their operational conditions, such as STP, LTP, PSC retention, as well as STDP. For instance, highly symmetric STDP characteristics are presented in Fig. 5(c).¹³ Such characteristics were also obtained in the devices operating with Al₂O₃ as the gate dialect in place of SiO₂, confirming they are intrinsic characteristics of 2D MoS₂ layers.⁴⁵ Similarly, He et al. demonstrated optoelectronic synapses with CVD-grown 2D MoS₂ monolayers directly integrated on p-Si substrates.⁴⁶ The devices exhibited memristive characteristics with a large self-rectification ratio accompanying essential sets of optoelectronic synapses, such as potentiation/habituation, STP, and LTP. These characteristics are similar to those observed with the afore-mentioned 2D MoS₂/SiO₂/Si FET devices, irrespective of the dissimilarity in their device structures.

In addition, Yin et al. fabricated hybrid memtransistors employing mechanically exfoliated 2D MoS₂ tri-layers and reported a large on/off ratio of ∼10⁶ achieved with electric and light dual gates, as shown in Fig. 5(d).⁴⁹ The devices exhibited an enhancement of drain-source current by ∼3 orders of magnitude upon optical illumination, as shown in Fig. 5(e), as well as an increase in their low-resistive conductance state with increasing the illumination power.⁴⁹ In addition, the conductance steadily increased during the illumination of 60 s and exponentially decayed upon terminating the light, while it remained nearly constant in the dark, as shown in Fig. 5(f).⁴⁹ The devices operated based on the gold (Au)/2D MoS₂ Schottky barriers, which modulated the concentration of electrons within the 2D MoS₂ layers, i.e., with negative (positive) gate voltages, the Schottky barriers were increased (suppressed), leading to the decrease (increase) of the electron concentration, respectively. Furthermore, the reduction of the Schottky barrier height was also achieved with light illumination, which led to the injection of photogenerated carriers into the 2D MoS₂ layers increasing the source–drain current.⁴⁹ Finally, influences of gating media on optoelectronic synaptic characteristics were verified with 2D MoS₂ layers-based FETs employing electrolyte solution⁵⁶ or an ionic liquid.⁵⁷ For instance, John et al. demonstrated optoelectronic synaptic characteristics in 2D MoS₂ layers-based FETs where ionic liquids were employed as the gating media.⁵⁷ Persistent photocurrents were observed under the application of optical pulses (λ₀ = 445 nm), which was attributed to the defect or trap-associated slow recombination of electrons driven by the ionic liquid gate. Finally, optoelectronic synaptic characteristics of 2D materials were also demonstrated to be tunable with ferroelectric substrates instead of conventional SiO₂/Si substrates.^59,73 Du et al. employed 2D MoS₂ layers-integrated STO substrates where ferroelectric La_0.8Sr_0.2MnO₃ (LSMO) and BaTiO₃ (BTO) films were deposited as the gate electrode and dielectric, respectively, as shown in Fig. 5(g).⁶⁰ They investigated the dependence of the light-induced conductance states with the light pulses of three different light wavelengths (i.e., 450, 532, and 650 nm), as shown in Fig. 5(h).⁶⁰ Furthermore, the influence of the pulse number with a fixed light (10 mW cm⁻²) on resulting optoelectronic characteristics was examined. Specifically, the significant imaging contrast between the blue features and the other features in Fig. 5(i) was obtained with the devices operating with a different number of applied pulses, implying their implementation into a vision sensory array.⁶⁰ These 2D layer/ferroelectric heterostructures exhibited optoelectronic synaptic characteristics modulated by the light-tunable ferroelectric properties. For instance, the devices achieved a series of light-driven LTP and LTD characteristics under various illumination conditions, including illumination time, the number of the pulse, as well as electrical triggering.⁶⁰ Table I summarizes the performances of various optoelectronic synaptic devices based on 2D materials.

In conclusion, recent progress in the optoelectronic operations of 2D materials for neuromorphic applications is overviewed with a focus on their artificial synaptic properties and device-level operation mechanisms. A variety of 2D materials-based systems adopted for such applications are discussed, and their up-to-date implementations of proof-of-concept optoelectronic synaptic devices are outlined. Despite the rapid progress in verifying the projected promise of 2D materials for artificial synapses, several technical challenges still need to be resolved to realize their true potential for practically relevant neuromorphic applications. While various manufacturing strategies for developing wafer-scale single-crystalline 2D materials with diminished defects are gradually emerging, the resulting structural homogeneity and electrical uniformity of the materials are still far from optimization. This limitation often leads to device-to-device and cycle-to-cycle variability in individual synaptic devices, which will become more pronounced as the devices are integrated into system levels to perform complicated data processes. Furthermore, reducing the device energy efficiency and enhancing the component integration density still await further solutions, which are also associated with the quality of 2D materials. Specifically, the wafer-level scalable integration of 2D layers with all essential footprints precisely controlled/maintained at the atomic-scale still calls for breakthroughs to achieve reliable optoelectronic synaptic operations with long-term reliability. In addition, from the perspective of electrical-switching mechanisms in 2D materials, the factors that contribute to determining the charge trapping/detrapping process need to be further clarified in terms of the materials’ interfacial and chemical homogeneity. This issue is related to enhancing the intrinsically low predictability of neuromorphic devices by better controlling their stochastic operational nature. Overall, holistic approaches spanning from wafer-scale 2D materials growth and characterization to system-level device integration should be pursued for the realization of practically implementable brain-like computation systems operational with optoelectronic artificial synapses.

T.R. and Y.J. acknowledge support from AFOSR (Award No. FA8651-20-1-0008).

The authors have no conflicts to disclose.

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