In-sensor neuromorphic computing using perovskites and transition metal dichalcogenides

With the advancements in Web of Things, Artificial Intelligence, and other emerging technologies, there is an increasing demand for artificial visual systems to perceive and learn about external environments. However, traditional sensing and computing systems are limited by the physical separation of sense, processing, and memory units that results in the challenges such as high energy consumption, large additional hardware costs, and long latency time. Integrating neuromorphic computing functions into the sensing unit is an effective way to overcome these challenges. Therefore, it is extremely important to design neuromorphic devices with sensing ability and the properties of low power consumption and high switching speed for exploring in-sensor computing devices and systems. In this review, we provide an elementary introduction to the structures and properties of two common optoelectronic materials, perovskites and transition metal dichalcogenides (TMDs). Subsequently, we discuss the fundamental concepts of neuromorphic devices, including device structures and working mechanisms. Furthermore, we summarize and extensively discuss the applications of perovskites and TMDs in in-sensor computing. Finally, we propose potential strategies to address challenges and offer a brief outlook on the application of optoelectronic materials in term of in-sensor computing.


Introduction
In today's society, with the rapid advancement of technology, the continuous integration of artificial intelligence (AI) and sensor technology has become an inevitable trend.Sensor technology, as the bridge between the physical and digital worlds, is becoming increasingly prevalent and intelligent [1].However, traditional sensor devices typically only capture environmental information, and data storage, processing, and analysis often rely on remote cloud servers or physical separated computing devices [2].The time latency, data breach risks, and significant energy consumption associated with remote or separated processing have gradually led to a demand for sensor systems with greater intelligence and autonomous decision-making capabilities [3].In this context, the field of in-sensor neuromorphic computing has emerged.In-sensor neuromorphic computing introduces the concept of neuromorphic computing into sensor devices, giving sensors themselves intelligence and the ability to perform data processing and decision-making locally [4].The core idea is enlightened by the functioning of biological neural systems, where the complex connections and information transfer between neurons have led researchers to realize that this neuromorphic computing approach can effectively simulate, learn, and adapt to the environment [5].
The human learn and perceive the world through the senses of visual, auditory, tactile, olfactory, and gustatory.Among them, visual sensing is the main approach to obtain information.The visual information acquired in daily life constitutes most of the sensory input received by individuals, encompassing a significant portion of their overall perceptual experience [6,7].Therefore, simulating the visual sensing function of biological neural networks is particularly important for the researches of ANNs.
As a basic component of ANNs, artificial neuromorphic devices provide an important platform for the realization of in-sensor computing.Artificial synaptic devices show great advantages in simulating biological synapses, and at the energy consumption level of pJ or even fJ [8].The artificial synaptic devices possess the capability to perform various functions akin to biological synapses, encompassing STP/STD, LTP/LTD, and STDP.
The neuromorphic devices can be broadly classified into two categories: two-terminal memristors and three-terminal neuromorphic transistors.Each kind of device has its own advantages and disadvantages.Memristor exhibits simple structure and small cell size, which make it possible to realize super high-density integration, and simulation of many synaptic functions with low power consumption [9].But the memristors are facing the challenges of low sensing efficiency and high device variability.On the contrary, though three-terminal neuromorphic transistors suffer from the issues of integration, they commonly exhibit the excellent sensing ability, and have the separated reading and modulating terminals, facilitating the emulation of complex and diverse functions.Up to now, both of these two neuromorphic devices have widely applied in sensory perception, including visual, auditory, tactile, olfactory, and multisensory [10].
In this review, we first introduce the crystal structure and band gap of two common optoelectronic materials, perovskite and TMDs.Subsequently, we discuss the fundamental concepts of neuromorphic devices, including device structures and working mechanisms.Based on the above basic concepts, we summary and discuss the recent applications of perovskites and TMDs in optoelectronic neuromorphic devices and sensory perception.Finally, we propose the potential strategies to address challenges and provide a brief outlook on the application of optoelectronic materials in the domain of in-sensor computing.The overall diagram for this article is shown in figure 1.

Perovskites
Due to its unique physical and chemical characteristics, perovskites have been extensively employed in the research of optoelectronic devices, and have laid a solid material science foundation for the research of neuromorphic computing [67].

Structures of perovskites
Perovskites usually refer to a specific crystal structure adopted by various compounds with the general formula ABX 3 [68].The 'A' cations are typically larger and located at the corners of the unit cell, forming the vertices of the corner-sharing octahedra (figure 2(a)).Common 'A' cations include organic molecules or relatively smaller metal cations.The 'B' cations are smaller and are located at the center of the corner-sharing octahedra.The characteristics of this crystal structure critically determine the electrical and optical properties of perovskite materials [69].Common 'B' cations include transition metals and rare earth elements.'X' represents an anion, typically oxygen or halogen, coordinated to the B cation.In this arrangement, the anions form a coordinated structure around the B cations. Figure 2 illustrates the ideal cubic closed-packed perovskite structure, where B-, X-and A-sites occupy the corners, the octahedron centers, and the framework centers, respectively [68].The crystal structure can be visualized as a three-dimensional network of corner-sharing BX 6 octahedra [70].The A and B cations, along with the X anions, contribute to the stability and unique properties of perovskite materials.The cubic close-packed layers formed by the A-site cations and X-site anions further contribute to the overall structural integrity of perovskites.What distinguishes the perovskite structure is its flexibility and adaptability.The size of the A-site cation can vary, allowing for a wide range of chemical compositions.Goldschmidt proposed a 'tolerance factor' (t) [72], which is used to evaluate the stability of the perovskite lattice, where r A , r B , and r X , are the ionic radii for the ions A, B and X, respectively.When the tolerance factor t is closer to 1, the structure of perovskites is more stable, so the A-position ions have a larger ionic radius (for example, MAPbI 3 , FAPbI 3 , CsPbI 3 ) [73].Usually, when 0.9 < t < 1.0, the cell structure of perovskites is a three-dimensional cube (figure 2(a)).As t increases, disrupting the stability of the B-X bond, perovskites will be reduced to a two-dimensional (2D) or quasi-2D layered structure (figure 2(b)).Compared to 3D perovskites, 2D perovskites are more stable and have anisotropic properties, resulting in natural quantum well effects due to the segmentation of long chain molecules [74].This series of theories and concepts promotes the diverse range of applications for various perovskite materials in electronic devices, while elucidating and assessing the stability and performance characteristics of these materials through parameters such as tolerance factors.
Due to its outstanding achievements in the field of novel solar cells, perovskite materials have attracted the attention of many research fields [75].However, a critical challenge that cannot be overlooked pertains to the toxicity associated with the heavy metal lead [71].To addressing this issue, Yang et al embedded a cross-linking supramolecular complex to reduce the toxicity of lead [76].Developing high performance Lead-free perovskite with low toxicity is paramount for the broader acceptance and integration of perovskite technology into large-scale market applications.
Perovskite exhibits a high degree of charge carrier mobility, thereby facilitating the efficient transportation of both electrons and holes within the material.This high mobility is crucial for electronic devices like transistors [77].However, the presence of defects and trap states may affect the electronic properties of perovskites [78,79].

Optoelectronic property of perovskite
Perovskite materials have modulable energy gaps, which enable them to absorb light efficiently over a wide spectrum.Its high absorption coefficient is particularly beneficial for solar cells to achieve high overall efficiency.In addition, perovskites also exhibit strong photoluminescence properties, which emit light under external light sources.This property facilitates its application in light emitting diodes (LEDs) and other electroluminescence fields [80].Perovskite materials offer strong photosensitivity and rapid charge transfer, enabling efficient signal processing in sensing and computing applications across visible, ultraviolet and near-infrared spectra [81][82][83].With these combined benefits, perovskites present promising opportunities for advancing sensor technologies, facilitating enhanced detection capabilities and enabling more efficient and reliable in-sensor computing processes.

TMDs
Due to their unique chemical and physical characteristics, 2D materials have found widespread applied in the research of optoelectronic devices, providing a significant material platform for the realization of in-sensor neuromorphic computing.As an important member of the 2D material family, TMDs have garnered extensive attention from both scientific and industrial circles due to their diverse crystal structures, ultra-thin thicknesses, and bandgap correlations that make them ideal for applications in optoelectronic devices [84].The structure of TMDs materials is generally defined as MX 2 , where M represents the transition metal atom of IV-VIII, and X represents the chalcogen, including sulfur, selenium, or tellurium.

Structures of TMDs
TMDs materials can be broadly classified into four types: semiconductor TMDs (s-TMDs), metallic TMDs, semi-metallic TMDs, and insulating TMDs, due to differences in the electronic energy bands and their electrical conductivity.Totally, insulating TMDs and s-TMDs have the energy band between the valence band (VB) and conduction band (CB), and the Fermi level (E f ) is located at the band gap [85].If the Ef of the TMDs material falls within the mid-band gap, indicating a balanced concentration of electron and hole carriers, then the material exhibits intrinsic semiconductor properties [86].The bottom of the CB and top of the VB have partial coincidence in semi-metallic TMDs.Therefore, semi-metallic TMD has almost no band gap, and the energy level density near E f is negligible [87].In this subsection, we mainly discuss TMD materials whose properties are close to semiconductors.
In general, at least five different phases of TMDs materials have been observed, including 1T, 1T ′ , 2H, 3R, and 1Td [88] (figure 3).These five structures can be grouped into two categories, T phase (1T, 1T ′ and 1Td) and H phase (2H and 3R).The H phase is a triangular prism structure, which is the symmetrical arrangement of upper and lower tetrahedra, with metal atoms as symmetric points.In contrast, the T-phase is an octahedral structure, which is obtained by rotating the upper tetrahedron or lower tetrahedron of the H phase by 180 degrees.The basic structure of the 1T phase is a hexagonal system, in which one layer X moves relative to another layer X ′ , resulting in the structure of XMX ′ , where X and X ′ are the chalcogen layers.The TMD of 1T phase is mainly metallic and a small part is semiconductor.For example, 1T phase HfS 2 and HfSe 2 have bandgap widths of 1.45 eV [89] and 1.1 eV [90], respectively.The 1T ′ phase and the 1T d phase are generally formed by the deformed deformation of the 1T phase [91].For example, in the 1T ′ conformation of MoTe 2 , the Mo and Te atoms are removed from their original octahedral coordination positions and arranged in a zigzag pattern along the [010] crystal direction [92].In the 1T d conformation of TMDs materials, the disulfide compounds exhibit a one-dimensional (1D) structure, such as MoTe 2 nanosheets arranged along the [001] crystal direction [93].Nay, a similar structure was also found in WTe 2 [94].2H and 3R phase TMDs materials can be obtained by stacking H monolayers in the order of ABA and ABC (A, B and C are the different layers in the crystal of TMDs emphasizing the difference in their order) respectively.Viewed from the top down, the basic structure of the 2H phase TMD material is a hexagonal lattice formed by alternating two stacked X atoms and one M atom.Femtosecond pulsed lasers [95], electrical pulse signals [96] and gate voltages [97] can trigger the changes between different phase states.Monolayer MoTe 2 can undergo reversible phase transitions between 2H and 1T ′ under the control of ionic liquid gates [98].Inserting an alkaline layer between MoS 2 and MoSe 2 can cause a phase transition from a 2H structure to a 1T structure [99].These changes are non-volatile and result in a change in the resistive state of the material, which is beneficial for the research of resistance random access memory (RRAM) and synaptic devices [100].Memristor phase transitions have also been validated in 2D 1T-TaS 2 crystals [101], offering potential for neuromorphic computing.

Optoelectronic property of TMD
TMDs exhibit a tunable bandgap based on the material and its thickness, allowing for precise control over electronic properties.With the decrease of thickness to monolayer, the bandgap of TMDs transition from indirect to direct, resulting in enhanced light-matter interactions [103][104][105][106].This property of tunable bandgap is crucial for the applications in optoelectronic devices, such as photovoltaics, photodetectors and LEDs.The presence of a bandgap makes TMDs suitable for applications in FET as semiconductors.TMDs also display strong spin-orbit coupling, resulting in the emergence of spin-valley coupling.This property opens up possibilities for exploiting the freedom valley degree in information processing [107][108][109][110][111][112] and quantum computing [113][114][115][116]. Additionally, TMDs can emit light in the visible range, enabling their use in displays and other optoelectronic devices.The nonlinear optical properties of TMDs are noteworthy, allowing for applications in frequency conversion and generation of new optical frequencies [117][118][119][120].

S-Y Li et al
Nonlinear effects such as sum-frequency generation and second-harmonic generation have been observed in TMDs, providing opportunities for the development of nonlinear optical devices [119,[121][122][123][124].

Neuromorphic devices
In neuroscience, the basic units in biological neural networks (BNNs) are synapses and neurons.A synapse functions as a junction that connects an afferent nerve terminal with a corresponding efferent neuron.The strength of connection is called synaptic weight.Synaptic weights can be modulated in response to pre-synaptic/post-synaptic neuron activity, which is also known as synaptic plasticity and the basis of learning and memory [125].Compared to traditional computer, human brain exhibits less power consumption and high computing efficiency, and has the ability of fault-tolerent, event-driven, in-memoy computing, and parallel processing [126][127][128].Therefore, constructing hardware-based neuromorphic computing systems based on neuromorphic devices through emulating the structural characteristics of BNNs is an efficacious approach to break the limitations imposed by von Neumann bottleneck, and realize high-efficiency and low-energy data processing.Various kind of neuromorphic devices, including memristors [129][130][131][132][133], transistors [51,[134][135][136][137][138][139], memtransistor [140][141][142][143][144][145][146], spintronic devices [147][148][149][150][151], and phase-change memory [152][153][154][155][156][157] had been developed.Memristor and neuromorphic transistor are two most common devices for mimicking the synaptic behaviors.Memristor has the advantages of simple structure and small size, facilitating the array of cross-structure for large-scale integration, and providing a good hardware basis for the neuromorphic computing.However, since there is only one way in the memristor to transmit the signal before and after the synapse, the transmission and processing functions are difficult to modulate simultaneously, so they can only simulate the relatively simple biological synaptic functions.On the contrary, neuromorphic transistors have attracted more and more attention because of their advantages such as more input ports, easy to control test parameters, good cooperative control, and can realize spike logic operations and logic modulation with multiple presynaptic inputs.However, it is undeniable that the three-terminal neuromorphic transistors are still limited by integration issue because of their larger device size.In this section, the device structure and working mechanism of the two-terminal memristors and three-terminal neuromorphic transistors are introduced.

Device structure
Memristor usually exhibits typically sandwich structure including TE, resistive switching (RS) layer, and BE.According to the geometry of TE and BE, the memristors can be divided into four types: cross-bar (figure 4(a)), cross-point (figure 4(b)), lateral (figure 4(c)), and common BE (figure 4(d)).According to the electronic behaviors under the action of voltage sweep, the memristor can be divided into two categorie: digital-and analog-type [158,159].The digital-type memristor (figure 5(a)) has abrupt resistance switching properties, enabling the widely application in the next generation of non-volatile memory [160,161].The analog-type memristor (figure 5(b)) has the characteristic of gradual RS, and attracted more and more attention in the application artificial synapse and neuromorphic computing [162,163].
When the memristors are applied for mimicking synaptic behaviors, one electrode is treated as the neuron at the front of the synapse, and the other electrode is used to receive electrical signals and transmit response current to the postsynaptic neuron.Because the memristor has obvious advantages in simple device structure, small size, low energy consumption and high operation speed, it is suitable for the preparation of high-density electronic and large area integration device array.

Mechanism
The working mechanism of the memristor can be roughly divided into ion migration and charge trapping.In this sub-section, these two working mechanisms are briefly discussed.
In the system of ion migration, the mechanism can be subdivided into ECM and VCM (figure 4(e)).The memristor with ECM mechanism is generally composed of an oxidizable TE (anode, such as Ag and Cu), a relatively inert BT (cathode, such as W and Pt), and a dielectric RS layer (such as SiO 2 [164], Al 2 O 3 [165]) sandwiched between TE and BE (figure 4(a)).Under the action of electric field, the TE atom (M) can be oxidized into M n+, and then migrates towards the inert electrode across the dielectric layer.On the inert electrode side, M n+ obtains electrons to undergo a reduction reaction: M n+ + ne − → M, transforms into metal M, and eventually form metallic CFs, leading to the RS behaviors [166][167][168].
Different from ECM-based memristors, VCM (figure 4(f)) based memristors are mainly composed of inert metal electrodes/RS layer/inert electrodes, in which the dominant factor is mostly anion (oxygen, sulfur, and so on) ions or vacancies.In order to simplify the model, oxygen vacancies are usually used to describe the RS process.Under the action of electric field, the oxygen vacancy migrates, resulting in the  uneven composition of the RS layer material and the change of valence state of the metal elements (figure 4(b)).Simultaneously, arrangement of oxygen vacancies gives rise to formation of non-metallic CFs, resulting in a modification in device conductance.Moreover, the presence of oxygen vacancies at the interface can induce alterations in the Schottky barrier height, thereby influencing the RS process.Therefore, the RS mechanism of VCM-based devices is often the joint action of CFs and interface barrier.The important feature of the device is that the resistive layer itself participates in the redox reaction, so it generally has more stable resistive properties and better fatigue properties [169,170].
The other main working mechanism is charge trapping [171][172][173][174][175]. Memristors based on charge capture and release are usually related to charge traps.When the carrier is captured by a defect in the film, the distribution of trap energy levels within the energy band will undergo changes, leading to alterations in resistance.The introduction and modulation of charge trapping sites can be achieved through strategies such as doping nanomaterials, constructing heterostructures, and engineering interfaces [176][177][178][179][180][181].One of the typical conductance mechanisms in the charge transfer process is SCLC.By adjusting the size of the applied voltage, the J-V characteristics of the device can be switched between the three regions to achieve resistance [182].

Neuromorphic transistor
Neuromorphic transistor can solve some of the natural simulation problems of two-terminal memristors, because their information transport terminals and the synaptic weight modulation terminals are physically independent of each other.This property allows the three-terminal transistor to have a more stable operating current than the two-end memristor, thus giving the three-terminal transistor higher stability [190].Here, we introduce the device structure of neuromorphic transistor by taking FET as example.The classic of FET consists of four basic components: In a neuromorphic FG-FET [191,192], When the G voltage is applied to the gate electrode, carriers are injected into the FG to create an internal electric field that regulates the conductivity in the semiconductor channel [191].Gate voltage pulses effectively modulate the charge captured in the FG and change the channel conductance, so it can be applied for simulating the regulation behaviors to synaptic weights in biology.Artificial synapses made from FG-FET have the advantages of large capacity storage and advanced manufacturing technology [193].However, the problems of its instability under high voltage operation [194] and the difficulty of simulating long-term synaptic plasticity [195] need to be solved urgently.
The channel current of EDLT [196] can be regulated by the change of gate voltage through the capacitive field effect mechanism at the interface of channel and electrolyte.Due to their high coupling rate between gate and channel, synaptic devices based on EDLTs can achieve lower operational voltages [197], thus reducing energy consumption in neuromorphic circuits [198].Wan's group has designed many EDLT-based synaptic devices, confirming that planar devices with polymer electrolytes can realize multiple synaptic functions [199].Though EDLTs have been applied in intelligent ANN systems, there are still many problems to be solved in large-area integration and uniform device array preparation.
After a large number of ionic substances are injected into the redox material (electrochemical doping), the electrical conductivity of the ECT can be changed [200].Although ECT based electronic synapses have good non-volatility in order to regulate conductance, problems in device stability, switching speed and electrolyte stability are still important factors that plague the application prospects of ECT.
Ferroelectric materials are ideal for the preparation of Fe-FETs due to their spontaneous polarization [201].The Coulomb interaction of the carriers in the channel and the polarization effect in the ferroelectric insulator allow the device to adjust the carrier concentration in the FET by changing the gate voltage [202].Due to the advantages of multiple conduction states, large on/off ratio and high stability, Fe-FET have shown great potential in simulating biological synapses.However, similar to FG-FET, the issue of device size scaling-down is also a difficult problem that needs to be solved.
Memristors and transistors have been implemented including LTP/LTD, STP/STD, excitatory/ IPSC(E/IPSC), transition from short-term memory to long-term memory (STM to LTM), STDP, SRDP, and PPF/PPD (figure 7).These synaptic functions play an important role in training neuromorphic systems.As the important characteristics in BNN itself, such as EPSC, PPF, and STDP, the realization of these synaptic functions can directly reflect the ability and extent of a device or an ANN to simulate BNN.Because of the linearity of vector matrix multiplication in ANN [203], the recognition accuracy and efficiency of the ANN depend on the linearity and symmetry of the weight update trajectory and the number of effective conductance states [204].The parameters of linearity and symmetry extracted from LTP/LTD curves can be added to the ANN as initial variables.On the basis of these parameters, the constructed ANN is able to be trained for executing the task like pattern recognition.
In general, memristor and neuromorphic transistor based on optoelectronic materials have irreplaceable advantages and can play an important role in in-sensor neuromorphic computing.

Application
Mimicking the information processing behaviors of the nervous system provides a strategy for developing computing systems with low energy consumption.The design and fabrication of neuromorphic devices are important for the construction of ANNs and in-sensor neuromorphic computing system.At the same time, the integration of neuromorphic devices into arrays and the testing of their corresponding perceptual recognition capabilities are also an essential part of the realization of in-sensor neuromorphic computing S-Y Li et al ability.Optoelectronic materials provide a solid foundation for the fabrication of various artificial synaptic devices due to their excellent optical capability, and unique physicochemical properties.In this subsection, we will introduce the applications of perovskites and TMDs in neuromorphic devices and sensory perception.

Perovskites for artificial synapses 4.1.1. Perovskite for memristors
Both of the ion migration and charge trapping mechanism have been applied for designing and developing Perovskite based optoelectronic memristors (table 1).Taking advantage of ECM mechanism, in the study conducted by Ku et al a memristor was meticulously crafted with a distinctive Ag/MAPbI 3 /FTO structure [205] (figure 8(a)).
The intriguing RS characteristics of this device were primarily attributed to the formation and disruption of Ag CFs within the AgI interlayer.The interaction between Ag and I ions at the interface of Ag/MAPbI 3 induced the creation of a β-AgI phase.This intermediary β-AgI layer acted as a reservoir for silver, influencing the generation of mobile Ag + and modulating the extent of Ag + migration (figure 8(b)).The memristor exhibited a distinctive bipolar RS behavior, characterized by a remarkable R ON/OFF ratio of approximately 10 3 under the compliance current (I CC ) of 10 mA (figure 8(c)).Notably, this device managed to emulate several fundamental synaptic features observed in biological systems.These included SRDP, PPF, PTP, the transition from short-to long-term plasticity and STDP (table 1, figures 8(d) and (e)).The operational energy of the MAPbI 3 -based memristor was estimated to be exceptionally low, reaching about 47 fJ/µm 2 .This energy efficiency was attributed to the diminutive cell area, which resulted in a low responsive current.In essence, the designed memristor not only showcased impressive RS characteristics but also demonstrated its ability to mimic key synaptic behaviors observed in biological systems, making it a promising candidate for the applications in neuromorphic computing.
VCM mechanism can also be applied for developing perovskite based memristor for the high RS effect, fast response time and non-volatile memory [214][215][216].Yang et al fabricated a device with structure of Ag/PMMA/MA 3 Sb 2 Br 9 /ITO [208] (figure 8(f)).The spontaneously generated metallic filament in the as-synthesized MA 3 Sb 2 Br 9 was facilitated by the presence of initially formed antimony (Sb) and bromide vacancies, leading to the forming-free behavior (figure 8(g)).This phenomenon closely resembles the synaptic weight change observed in biological synapses due to the movement of calcium ions (Ca 2+ ).Remarkably, MA 3 Sb 2 Br 9 demonstrates stable RS behavior across 300 cycles, with an impressive R ON/OFF of approximately 10 2 , long retention time (∼10 4 s), and excellent endurance (300 cycles) (figures 8(h) and (i)).Some common synaptic behaviors, including E/IPSC, LTP/D, and STDP, can be mimicked by using this memristor (figure 8(j)).Moreover, this device can obtain a relatively low energy consumption of 117.9 fJ µm −2 due to the low corresponding current.
Methylammonium lead halide (MAPbX 3 ) perovskite is an organic-inorganic hybrid perovskite material and widely employed as RS layer for memristors.However, this kind of materials are always instable due to the hygroscopic nature of the CH 3 NH 3 + (MA).In contrast, formamidinium lead halide (FAPbX 3 ) perovskite exhibits enhanced stability compared to MAPbX 3 [217].This improved stability is attributed to the superior bonding interactions between the FA cations and the inorganic matrix [218].Das et al designed and fabricated an artificial synapses device (Al/FAPbBr 3 /ITO) [206] (figure 9(a)).The device working mechanism is considered as the migration of vacancies and halide ions in the perovskite film under the influence of the electric field (figure 9(b)).The device replicated a range of synaptic functionalities, including PPF, EPSC, SVDP, SFDP, and SDDP (figures 9(c)-(f)).In addition, the synaptic behaviors can also be replicated by using light as stimulation which can promote current conduction by creating electron hole pairs [219] (figure 9(g)).Based on the charge trapping mechanism, Ma et al engineered a memristor featuring a structural composition of Au/CsPbBr 3 /CuSCN/PEDOT:PSS/ITO [213] (figure 9(h)).The CsPbBr 3 layer promotes the optical response of the device.The charge trapping/de-trapping process at the CuSCN/PEDOT:PSS interface dominates the device RS behaviors.Optoelectronic synapses possess a remarkable memory retrieval capability [220], allowing them to retrieve historical optoelectronic information and replicate various synaptic functions such as PPF, short-/long-term plasticity, and the transition from STM to LTM.As the time interval ∆T decreases, the PPF index shows a gradual increase, aligning with the characteristics of synaptic PPF behavior (figure 9(i)).Under the stimulation of UV light pulses (371 nm, 14.3 mW cm −2 ) with fixed direction (figure 9(j)), the PPF effect can also be replicated.A1 represents the device current observed after the initial exposure, while A2 corresponds to the device current measured after the subsequent exposure.Compared to the electric stimulation, the light stimulation PPF demonstrated by optoelectronic synapses closely mirrors the behavior observed in biological synapses (figure 9(k)).The inset shows a peak current at a 0.9 V pulse.(f)-(j) Reproduced from [208] with permission from the Royal Society of Chemistry.

Perovskites for neuromorphic transistors
In three-terminal neuromorphic transistors, the gate electrodes function as the input receptors, akin to presynaptic membranes that receive external stimuli.On the other hand, the conduction channel serves as the output receptors, similar to postsynaptic membranes.The variations in channel conductance dynamics represent the synaptic plasticity.This unique structure enables simultaneous signal reception and reading, facilitating the intricate adjustment of synaptic weight under diverse signals.Up to now, various perovskites, including MAPbI 3 , CsPbBr 3 , and MAPbBr 3 , have been applied in optoelectronic neuromorphic transistors (table 2).In 2018, Wang et al fabricated an artificial synaptic transistor with a device structure of Au/Pentacene/PMMA/CsPbBr 3 /SiO 2 /Si [221] (figure 10(a)).The device mechanism involves the generation of charges under light illumination, rapid charge separation/transfer under the bias, and hole-electron recombination.The effective separation of excitons at the interfaces is achieved due to the type II band alignment between pentacene and CsPbBr 3 QDs [222].The observed phenomenon serves as the fundamental basis for optically triggered charge trapping and electrically induced charge release in flash memory utilizing CsPbBr 3 QDs.Leveraging this, nonvolatile synaptic plasticity can be emulated in the perovskite device when subjected to photonic or electric stimulation (figure 10(b)).The device successfully emulates synaptic functions such as short-/long-term plasticity, PPF/PPD, and SRDP (figures 10(c)-(e)).These functionalities serve as fundamental building blocks for artificial computing at the device level.Moreover, the electrical habituation and photonic potentiation are implemented and the synaptic weight of the device exhibits responsiveness across multiple wavelengths, specifically at 365, 450, 520, and 660 nm.The energy consumption for the synapse transistor is remarkably low at 1.4 × 10 −9 J per event due to the no electrical interconnect power loss and large optical bandwidth.
In 2022, Wang et al developed optical artificial synaptic devices with a structure of Au/Pentacene/MAPbBr 3 -Rhodamine B (RhB) /SiO 2 /Si (figure 10(f)) [223].The observed memory effect is likely attributed to the presence of charge trapping sites located at the interface between the pentacene layer and the MAPbBr 3 -RhB layer, which play a crucial role in capturing photogenerated charge carriers and thereby reducing the recombination rate of photogenerated holes and electrons.The incorporation of RhB into the composite synaptic devices resulted in a marked enhancement of photosensitivity (figure 10(g)).The devices successfully replicate fundamental functions of biological synapses, including PPF, short-term plasticity and long-term plasticity (figures 10(h) and (i)).Furthermore, these devices exhibit a notable response to light signals even at low intensities, as low as 1.1 µW cm −2 , and operate at modest working voltages (figure 10(j)).The channel currents of these devices can be effectively modulated by exposure to light stimuli, demonstrating satisfactory synaptic behaviors.Importantly, under a drain-source voltage of −50 µV, the devices showcase an impressively low energy consumption of 1.25 fJ (figure 10(k)).

TMDs for neuromorphic devices 4.2.1. TMDs for memristors
Ion migration is the most common used mechanism for developing TMD based neuromorphic devices (table 3).The stepwise construction of metal CFs enables the device transition between distinct multilevel resistance state.Recently, Yu et al designed and fabricated a two-terminal device in the structure of Ag/MoTe 2 /ITO (figure 11(a)) [227].The synaptic device demonstrates stable non-volatile bipolar resistive properties attributed to the formation and rupture of CFs.By subjecting the device to continuous modulation through a pulse train with a pulse width of 40 ns, the energy consumption for completing a write operation and generating an initial spike is reduced to 74.2 pJ and 1.22 pJ, respectively.Furthermore, this device can effectively emulate various biological synaptic behaviors including STP to LTP transition.It is worth noting that the decimal calculation, one of the most widely used arithmetic [228], was introduced into  the device to realize the operation function of addition (figure 11(b)) and multiplication (figure 11(c)).The realization of this function promotes the study of decimal arithmetic functions beyond binary arithmetic in TMD based synapses to complete the increasingly complex computational tasks of neural networks.In addition to the modulation of synaptic devices by electrical pulses, TMD-based two terminal synaptic devices also have excellent corresponding ability to optical signals.For example, near-infrared (NIR) light irradiation can be used as a good light modulation signal to regulate the synaptic performance.Wang et al designed and fabricated a NIR photo RRAM based on quasi-planar MoSe 2 /Bi 2 Se 3 heterostructure [229], which realized abnormal NIR threshold switching and light reset operation (figure 11(d)).Under near-infrared light at 790 nm, the syndicate in the bridging silver wire is oxidized back to Ag + due to photogenic hole impact, thus transforming the device from a LRS) to a HRS (figure 11(e)).The transformation from HRS to LRS is through electrical bias, which reflects the characteristics of photoelectric cooperative regulation.At the pulse intensity of 0.5 V, the dependence of the pulse interval on the synaptic weight in short-term plasticity was revealed by adjusting the pulse interval, and the characteristics of PPF and PPD were also explored.In addition to NIR light, modulation based on visible light signals is also a common way.Ranganathan et al designed an optoelectronic synapse based on a vertically aligned MoS 2 photo memristor (figure 11(f)) [230].Due to the formation and fracture of silver CF, the synapse can still maintain the long-term promoting and inhibiting properties up to 350 • C under the regulation of 520 nm green light.Interestingly, it is found that the optical responsiveness of the LRS is higher than that of the HRS (>30 times) (figure 11(g)).
Different from ECM based optoelectronic synaptic devices, the dominant factor in VCM based optoelectronic synaptic devices is mostly vacancy.For example, based on light-sensitive rhenium vacancy and sulphur vacancy migration, Seo et al fabricated a flexible photoelectric synapse on 2D van der Waals layered rhenium disulfide [231] (figure 12(a)).When the applied optical peak energy exceeds the band gap of rhenium disulfide, photoexcited excess electrons will be generated by the formation of metastable peroxides or the ionization of vacancies (figure 12(b)).For the optoelectronic synaptic devices, the synaptic behaviors were commonly modulated by changing the power of light stimulation.By adjusting the intensity, width and frequency of the input optical signal (0.5-5 mW cm −2 ), the device has a photosensitive memory and can directly switch its 32 conductive states in response to light stimulation, thus successfully simulating the dynamic characteristics of a variety of biological synapses, such as PPF with the index of 132%.The device also exhibits good synaptic properties under bending condition, after 1000 bending cycles, the device still maintains a highly stable LTP characteristic (figure 12(c)).Meanwhile, with a writing operation energy consumption of only 0.56 pJ, the synaptic dynamics of the two terminal optoelectronic device exhibited remarkable tolerance to physical bending, underscoring its flexibility.
The devices working with the charge trapping mechanism can also be applied for designing TMD based artificial optoelectronic synaptic devices.For example, He et al prepared a single-layer MoS 2 based memristor on a p-type silicon substrate [233] (figure 12(d)).The device is excited by 310 nm UV light, and the photogenerated electrons and holes are separated by the built-in potential field at position near the surface of the MoS 2 /Si heterojunction, resulting in a photodiode-like behavior.The reverse bias current is significantly enhanced under light conditions, in which photogenerated charge carriers dominate the The enhanced behavior of device 2 in two different configurations is not responsive in configuration I, but is effectively enhanced in configuration II.(j) Evolution of the conductance in device 1 when the system is changed from configuration I to configuration II.(f)-(j) Reproduced from [237], with permission from Springer Nature.transmission, compared with the state with high self-rectification ratio (4 × 10 3 ) in the dark condition.When the lamp is turned off, the node exhibits a continuous photocurrent phenomenon.Due to the gradual increase in the conductivity of the device due to the photogenerated electrons and holes, a series of continuous photon pulses (0.11 mW cm −2 , 1 s) can be applied to simulate the increase in synaptic strength.In contrast, under negative electrical pulses (−8 V, 5 ms), the electrons in MoS 2 are driven towards the interface and subsequently captured by the trap site layer at the SiO 2 interface, resulting in a gradual decrease S-Y Li et al in device conductance with an increasing number of electrical pulses.This behavior corresponds to the synaptic strength adaptation pattern (figure 12(e)).By changing the frequency of the light pulse from a low frequency of 0.1 Hz to a higher frequency of 1 Hz, the device will also exhibit a switch from STP to LTP.Based on the heterojunction structure of MoS 2 and Si, the device has photoelectric cooperative neuromorphic function, which can also realize synaptic functions including STM/LTM and PPF.
Not just limited to the mechanism of charge trapping and ion migration, the phase change mechanism can also be applied for designing TMDs based memristors.Zhu et al reported a reversible photoelectronic synapse in which the transfer of Li + s is controlled by an electric field, resulting in a 2 H (semiconductor) and 1T' (metal) phase transition of local MoS 2 (figure 12(f)) [237].Based on the high (in-plane) diffusivity of Li ions in MoS + 2 multilayers, the authors can establish effective ion coupling in the device to simulate synaptic interactions in biological systems [239][240][241], including synaptic competition and cooperation [242].In biological neural networks, the diffusion of PRPs (for example, CaMKII, calmodulin-dependent kinase II) (figure 12(g)), which are essential for synaptic growth between neighboring synapses, can induce synaptic interactions [243].Analogous to competition for finite PRPs in biological neural networks, the authors designed two identical synapses to affect each other's activity by increasing the potentiation (depression) degree of one synaptic device under the condition of limited supply of lithium ions, simulating the synaptic competition behavior in biological networks (figure 12(h)).The simulation of the competitive process can run stably under the condition of at least 25 cycles.Using 500 pulses (6 V, 1 ms) respectively to stimulate one of the synaptic devices, compared to whether the structure established a connection with the other device, the 100% conductivity enhancement was observed (figures 12(i) and (j)), indicating that one of the devices can provide a key component (lithium ions) to the other, thus achieving synaptic cooperation.When only one synapse is stimulated, the conductivity of the non-stimulated devices will also increase or decrease, which verifies that the effective ion coupling between the memristor device network can be obtained through in-plane Li ion exchange, and the surface system can effectively simulate the synaptic interaction effect in the ANNs.

TMDs for neuromorphic transistors
Three-terminal devices based on TMDs materials, including MoS 2 , WS 2 , and WSe 2 , have important applications in the fabrication of artificial synapses (table 4).Luo et al reported an optical synaptic Fe-FET with the structure of BGTC for simulating the synaptic behaviors [244] (figure 13(a)).In this structure, PZT as a ferroelectric grid dielectric, the film can be switched and permanently polarized to adjust the charge transport channel in WS 2 (figure 13(a)).When the light intensity is increased from 0.19 to 28 µW, the relaxation time is increased by nearly 45 times, showing obvious transition STP to LTP.At the same time, multiple exposures at 532 nm, with a duration of 100 ms and an optical power of 0.19 µW, can gradually trigger LTM patterns.Synaptic behavior has been demonstrated in neuromorphic FET based on amorphous oxide semiconductors [245], and sustained photoconductivity influenced by gate voltage bias and the measurement environment is also a characteristic of optoelectronic synapses inherent in MoS 2 [246].Islam et al used a monolayer MoS 2 based FET to simulate most of the common optical synaptic behaviors [247] (figure 13(b)).Under the stimulation of light with 450 nm, this FET can simulate the synaptic properties such PPF, and STDP (figure 13(c)).By tuning the stimulation conditions, the nonlinear factors (potentiation/depression), and asymmetry can be optimized to 2.92/5.49and 2.57 respectively, facilitating their neural network applications to train photoelectric synapses [248].
Compared with TMD synaptic FET, the difference of FG-FET is that the FG is usually embedded in a dielectric layer.Tang et al prepared a synaptic FG-FET based on the fully 2D van der Waals material (MoS 2 /h-BN) [249] (figure 13(d)).The energy consumption of this device is estimated to be 18 fJ, which is much lower than that of conventional CMOS circuits (≈900 pJ) [250].The conductivity change in the LTP/LTD process increases as the pulse width (amplitude) increases and eventually saturates as the pulse width (amplitude) increases (figure 13(e)).3000 quasi-continuous states were programmed for the LTP/LTD process with a series programming/erasing pulse of ±15 V (40 ns pulse width) (figure 13(f)), indicating that there is a large number of states available for neuromorphic computation in this artificial synapse.When a ±13 V pulse is applied, the resulting on/off ratio is 20, while the conductance is saturated only after 20 pulses.The study revealed that a more optimal linearity of weight update behavior could be achieved through reduction in both the number of states and the on/off ratio.The quantitative analysis of linearity of weight update behavior in the process is shown in figure 13(g), enabling the application for constructing neural networks with high accuracy and reliability [251].
In addition to electron/hole based charge transfer, protons can also be used as carriers of charge transfer [258].He et al reported a transistor-based synapse based on a 2D WSe 2 transverse heterostructure with proton charge transfer [254].By conducting hydrogen ions back and forth between WSe 2 and WO 3  interfaces to catalyze hydrogen evolution at the layered TMD interface, the device can achieve synaptic functions such as long-term plasticity and short-term plasticity with the consumption of 2.7 pJ.

Sensory perception
The study of synaptic devices is the most important foundation in the research of in-sensor neuromorphic computing.The ANN based on synaptic array is a leap from single device computing to array computing.By constructing an artificial synaptic device array with sensing ability, researchers can build an ANN that simulates some functions of biological neural networks, and thus realize some co-functions of perception and recognition.
In the aspect of perception recognition based on in-sensor neuromorphic computing, single-device computing is the basic content of current research [259][260][261][262][263][264][265][266][267][268][269].The artificial synapse based on a single device can realize the recognition of partial picture information including handwriting through simulation.For example, Wang et al simulated visual sensing with a three-terminal synaptic transistor based on a MoS2 material [259] (figure 14(a)).Using CrossSim software, the authors built a three-layer feedback supervised ANN with sizes of 64 × 36 × 10 and 784 × 300 × 10 (figure 14(b)), respectively.Using the probability distribution of the device's conductance change as the test value, they identified the small handwritten image of 8 × 8 pixels (figure 14(c)) and the large image of 28 × 28 pixels of MNIST (figure 14(d)).The network accuracy can achieve 90%, exhibiting a deviation of less than 10% from the results obtained with ideal floating-point numeric precision.In addition to simple recognition of individual images, single-device based perceptual recognition can also perform classification tasks for images based on data sets.Liu et al used NeuroSim multilayer PNN simulator to verify the application of VP-MoS 2 heterojunction in image classification [264,270] (figure 14(e)).The network, which contains 100 hidden neurons and 10 output neurons, classifies two datasets: MNIST and Fashion-MNIST [271].On the basis of 20 × 20 and 28 × 28 pixels of MNIST and fashion-MNIST, the recognition accuracy reached 95.23% and 79.65% respectively (figure 14(f)), which is close to the accuracy of ideal devices (95.47% and 79.95%).In addition, individual synaptic devices can also be applied for simulating pupil constriction-an unconditioned reflex in mammals-in terms of visual sensing [272].Chen et al simulated the PNN with CsFAMA-based optoelectronic devices (figure 14(g)) [268], which consisted of a 32 × 32 × 3 Flatten input layer, a 60 × 52 pixels sensing layer for real-time MAC of perovskite devices, and a dense output layer.Images of faces, aircraft, cars and birds obtained from the CIFAR dataset of the Keras software package are used for training and recognition testing.About 100% ideal accuracy can be achieved for face recognition using single-layer PNN.
The ion migration spontaneous relaxation through photovoltaic field induced by photocurrent of CsFAMA device is very similar to the image sensing adaptive behavior in biological vision system, and the background noise of overexposed images can be removed by pre-filtering, thus being processed using the same multilayer PNN based on a perovskite photovoltaic sensor array.By utilizing the Boltzmann function to extract relaxation dynamics through curve fitting of device responsivity, sensor-filter neurons can progressively achieve high-fidelity reproduction of captured images depicting aircraft, cars, and birds.Consequently, this approach effectively rectifies misidentification issues pertaining to target objects within the images.For example, an overexposed image of an unprocessed aircraft subject to the same PNN will be mistaken for a bird, and the two objects have similar pattern features.The recognition rates for all images of vehicles, aircraft and birds were reduced to 88, 62 and 38%, respectively (figure 14(h)).In adaptive imaging, although the recognition was wrong at first, the recognition rate rose rapidly as the adaptation period increased, and the PNN network was able to correctly identify the aircraft in the image within 0.9 s.For vehicles and birds, correct identification takes longer, but both can be done in no more than 3.3 s of adaptation.All images can be accurately identified in 4.2 s.
Although single-device based synaptic elements show great potential in perceptual recognition, their recognition accuracy will be greatly limited in the face of interference from external mechanical simulations.Simultaneously, the information processing capacity of a single device is limited [273], and array-based processing can process more information even in the face of external interference [274], so it is particularly important to study the array-level in-sensor computing.For example, Seo et al prepared a flexible van der Waals photo synaptic array unit consisting of 25 optoelectronic synaptic devices based on ReS 2 and hexagonal boron nitride heterostructures [231].An eight-layer CNN consisting of optoelectronic synaptic devices identifies the CIFAR-10 dataset (figure 14(i)) prepared by the DNN+ NeuroSim integrated benchmarking framework [275].Thedataset includes 10 object classes (such as dogs, horses, trucks, etc.), consisting of a training set of 50 000 images and a reasoning set of 10 000 images [276], and each image pixel contains three color channels: red, green, and blue.Through 50 000 training sessions and 10 000 reasoning sessions, the maximum recognition rate of the HW-NN is proved to be 89.4%, which is close to the ideal value of 91.1%.When the bending radius is less than 5 mm, the recognition accuracy reaches 89.2% (figure 14(j)).At the same time, in the flat state after 1000 bends, the accuracy reached 89.3%.The small accuracy gap with the completely flat array proves its primary computational properties of bending resistance.Besides the visual sensing, the optoelectronic synaptic devices can also be applied in the sensory fusion.Yu et al propose a graphene/MoS 2 heterojunction based artificial synaptic device with mechanical and optical co-plasticity (optoelectronic synaptic plasticity aided by MD) [252].The MD occurring between the two friction layers in the (Cu/polytetrafluoroethylene (PTFE)/Cu) TENG induces tribological potential coupling in the transistor.This directly impacts the charge transfer/exchange within the transistor channel, specifically in the graphene/MoS 2 heterostructure, thereby regulating the photocurrent of the optoelectronic synapse.When MD = 1 mm and the device is illuminated by a green light (pulse width 50 ms, intensity 3.5 mW cm −2 ) pulse and adjusted to different MD values, its long-term dynamic photocurrent response demonstrates the implementation of collaborative mechanical and optical modulation for LTP.In terms of perceptual recognition, the change of MD will also significantly affect the recognition accuracy.The simulated ANN comprises 28 × 28 input neurons, 100 hidden neurons, and 10 output neurons.Meanwhile, the training sample used between 0.5 and 1.5 mm.By fixing MD at 1.5 mm (the number of synapses is 600 000) and increasing the size of the training sample (from 360 to 60 000), the recognition accuracy can be increased from 54% to 92%.The combination of friction motors and artificial optoelectronic synapses reflects the interaction between vision and touch in biological neural networks.Mechanical optoelectronic artificial synapses can recognize certain mechanical tactile and optoelectronic signals, which broadens the scope of identifiable information.At the same time, the mutual correction between different senses can realize mixed mode interaction and simulate complex biological nervous system while improving the recognition rate, which has great potential in interactive in-sensor computing.
In general, the application of optoelectronic material (perovskites and TMDs) based neuromorphic devices in perception and recognition mainly remains in early stage, and the perceptual recognition capacity is still mostly relied on ANNs.The current in-sensor neuromorphic computing functions realized are mainly low-level computing capabilities based on single devices, such as noise reduction, contrast enhancement, and adaptation, while the relatively high-level computing capability based on large scale hardware device array has rarely been realized.On the other hand, neuromorphic devices based on optoelectronic materials also find application in the domain of visual sensing, and few of them can be combined with other senses, such as tactile.Because the current related studied ANN systems are still in the exploratory stage in terms of hardware, the simulation and realization of multi-sensory combination similar to BNN is still in the bottleneck stage.

Challenge and outlook
In this review, we have introduced the crystal structure and band gap of two common optoelectronic materials, perovskites and TMDs.Subsequently, the fundamental concepts of neuromorphic devices, including device structures and working mechanisms, are briefly discussed.Then, the recent applications of perovskites and TMDs in optoelectronic neuromorphic devices (memristor and transistor) and sensory perception are summarized and discussed.Although the perovskite and TMDs play a crucial role in applications within the realm of in-sensor computing, there are still some challenges need to be solved.
Materials.For TMDs, the material variety is still relatively rare.According to the definition of TMDs material structure (MX 2 ), there are at least dozens of suitable compounds within the range allowed by the periodic table of elements and without considering different phase states [86].Up to now, most of the sulfides are used, and molybdenum and tungsten are mostly used in the choice of metal types.This is partly because these compounds have better photoelectronic properties and strong stability [277].The more important reason is that the stability of other TMDs materials in synthesis process is still facing challenges.TMDs with single or few layers are usually synthesized by CVD.Besides, TMD materials, particularly those with metallic properties obtained from CVD, are prone to be oxidized, such as single-layer 1T ′ MoTe 2 .Therefore, it is necessary to combine the vacuum oxygen-free system during processes of the material synthesis, characterization, device fabrication and testing.
For 2D perovskites, the decay problem of under light conditions will seriously affect the electron transport characteristics and optical corresponding ability of the device.The 2D layer number, composition and lattice structure of perovskite materials are the key factors affecting the photoelectric properties.An appropriate number of layers can improve the absorption rate of the photodetector without affecting the electron transfer characteristics of the device.Therefore, the design of the material structure and the controlling to the preparation conditions should be well optimized for obtaining excellent optoelectronic properties.
Device.First, although two-terminal devices have a simple device structure that is easy for integration, the underlying physics of their working mechanism has not been thoroughly explored, and reliable control methods are lacking.Some suggestions have been made to improve the device performance such as modifying the electrode geometry [278] and built-in packaging [279], but a comprehensive and effective control mechanism for this phenomenon has remained elusive.Although three terminal devices demonstrate superior control capabilities, challenges related to material requirements and manufacturing complexity pose obstacles to their practical implementation in real systems.
Second, the quality of each contact interface in the devices exhibits a great influence to the electric performance.The addition of h-BN as an intermediate layer between the interface is proved to be effective in reducing Schottky barrier potential [280].However, substantial experimentation is required to validate the efficacy of this method.Graphene is also a good candidate for transparent electrodes, and can be applied for constructing van der Waals contact.Therefore, employing graphene as an intermediate layer to mitigate contact resistance represents a highly viable approach [281].
Third, although the nanoscale integration of optoelectronic devices can be achieved by nanomechanical methods such as lithography, the integration problems such as crosstalk and contact between manual and large-scale integration still exist.Besides, the application of novel materials including perovskite and TMD also presents problems such as EPSC fluctuations, long-term stability, response speed and environmental specifications.
Therefore, the future research should concentrate on understanding the fundamental working mechanisms of the underlying devices to inform the design of a more rational device structure.Simultaneously, the contact interface and large-scale integration process of the device should be optimized to reduce the performance degradation.
Sensory perception.First, although photoelectric synaptic devices have been implemented, the majority of applications for these devices are currently confined to fundamental model research, such as Pavlovian experiments and pattern recognition.To achieve more intricate neuron functions, it is imperative to judiciously integrate large-scale equipment, leveraging the collaborative efforts of computer science and electronic engineering.Besides, up to now, the computing results are generally obtained through simulation based on software.Therefore, achieving neuromorphic computing through extensive integration of arrays in hardware is a challenging journey that lies ahead [282].
Secondly, in terms of simulating the recognition ability in BNNs, synaptic devices based on photoelectric materials can only realize part of the function of visual perception, and are rarely applied to other senses.Only few works involve the studies about multi-sensory interaction and.Consequently, exploring and advancing multi-sensory interaction and fusion should be identified as a pivotal avenue for future research.
Despite the numerous difficulties and challenges, once these obstacles are overcome, neuromorphic devices based on perovskites and TMDs will assume a pivotal role in the domain of in-sensor neuromorphic computing, which can offer a viable avenue to enhance computational power while simultaneously reducing energy consumption within computing systems and realizing sensing ability.

Figure 1 .
Figure 1.Overall diagram of in-sensor neuromorphic computing using optoelectronic materials.

Figure 4 .
Figure 4. Four structures of memristor based on electrode type classification.(a) Cross-bar.(b) Cross-point.(c) Lateral.(d) Common BE.(e) Schematic diagram of ECM mechanism.(f) Schematic diagram of VCM mechanism.(g) Schematic diagram of ferroelectric polarization.(h) Schematic diagram of phase change.
(1) source/drain (S/D) electrodes for measuring the channel current; (2) a semiconductor layer that acts as a conductive channel and is affected by an external field to provide mobile carriers; (3) dielectric layer for regulating carrier migration; (4) gate (G) electrodes that can adjust the carrier concentration of the semiconductor layer.The length (L) and width (W) of the channel determine the size of the device.The thickness of each component ranges from a few to several hundred nano-meters, generally determined by the functional requirements of the device.According to the position relationship of each component in the device, there are four basic FET structures: BGTC (figure 6(a)), TGTC (figure 6(b)), BGBC (figure 6(c)), and TGBC (figure 6(d)).On the basis of the structures of FET, several neuromorphic transistors have been developed, including FG-FET (figure 6(e)), EDLT (figure 6(f)), ECT (figure 6(f)), ferroelectric FET (Fe-FET) (figure 6(g)).

Figure 8 .
Figure 8.(a) Schematic of the Ag/CH3NH3PbI3/FTO device.(b) Schematics of RS mechanism of the Ag/CH3NH3PbI3/FTO memristor, which highlights the upper process involving electroforming and the lower process that does not.(c) Corresponding PPF and PTP behaviors of Ag/CH3NH3PbI3/FTO memristor under the stimulations with different pulse intervals (d) and heights (e).(a)-(e) Reprinted from [205], Copyright (2020), with permission from Elsevier.(f) Schematic of perovskite based memristor (Ag/PMMA/MA3Sb2Br9/ITO). (g) Schematics of RS mechanism of the MA3Sb2Br9 based memristor.(h) and (i) Endurance (h) and retention (i) performance of MA3Sb2Br9 based memristors (endurance, writing voltage: −0.5 V, erasing voltage: 1.2 V, reading voltage: 0.01 V; retention: writing voltage: −0.5 V, reading voltage: 0.01 V).(j) IPSC characteristics observed under the pulse simulation of 0.9 V, 500 µs.The inset shows a peak current at a 0.9 V pulse.(f)-(j) Reproduced from[208] with permission from the Royal Society of Chemistry.

Figure 9 .
Figure 9. (a) Schematic illustrations of a biological neuron and FAPbBr3 based two-terminal artificial synapse.(b) Schematic illustrations of charge trapping mechanism in thermal equilibrium, voltage and light exposer.(c) PPF performance replicated by FAPbBr3 based artificial synapse.(d) and (e) The EPSC response under different pulse amplitude (d) and duration (e).(f) long-term plasticity achieved by applying 30 consecutive pulses.(g) The EPSC obtained in absence and presence of light illumination.(a)-(g) Reprinted from [206], with the permission of AIP Publishing.(h) Schematic illustration of Au/CsPbBr3/CuSCN/PEDOT:PSS/ITO.(i) Electrical PPF performance of device response when two positive pulses (2 V, 100 ms) are applied for ∆T = 1000 ms.(j) Schematic of the device measured under light illumination.(k) Typical photo-responsive PPF characteristic under a light pulse pair with a 5 s interval.(h)-(k) [213] John Wiley & Sons.[© 2020 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim].

Figure 11 .
Figure 11.(a) Device structure of the Ag/MoTe2/ITO memristor.(b) Graph of the result of addition operation.(c) Multiplication result graph.(a)-(c) Reprinted from [227], Copyright (2023), with permission from Elsevier.(d) Schematic configuration of RRAM device, showing MoSe2/Bi2Se3 nanosheets heterostructure sandwiched between the TE and BE.(e) Schematic of the light-modulated RRAM device.(d) and (e) [229] John Wiley & Sons.[© 2019 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim].(f) Structure diagram of the device and resistance mechanism diagram of ECM.(g) Responsivity vs laser power show the linear response of the device for laser at both HRS and LRS.Insert depicted for responsivity ratio between LRS and HRS for each power.(f) and (g) [230] John Wiley & Sons.[© 2020 Wiley-VCH GmbH].

Figure 12 .
Figure 12.(a) Structure of the ReS2 memristor and the simulated biological synapses.(b) Top-down diagram of sulfur vacancy and rhenium vacancy migration when considering density functional theory density (DFT) [238].A dashed black line represents a cell, and a solid red line represents a zigzag Re chain.(c) LTP characteristic diagram of the device under different bending times.(a)-(c) [231] John Wiley & Sons.[© 2021 Wiley-VCH GmbH].(d) Structure of the memristive synapse based on monolayer MoS2.(e) Photonic potentiation and electric depression in the W/MoS2/p-Si memristive synapse.(d) and (e) [233] John Wiley & Sons.[© 2018 WILEY-VCH Verlag GmbH & Co. KGaA, Weinheim].(f) Device structure diagram and phase transition diagram between 1T ′ and 2 H, under the premise of controlling lithium ion migration.(g) A schematic diagram of PRPs diffusion across multiple synapses in a biological neural network.(h) Synaptic competition outcome tested under 100 positive and 100 negative stimulation pulses (amplitude: 6 V, duration: 1 ms (i)) The enhanced behavior of device 2 in two different configurations is not responsive in configuration I, but is effectively enhanced in configuration II.(j) Evolution of the conductance in device 1 when the system is changed from configuration I to configuration II.(f)-(j) Reproduced from [237], with permission from Springer Nature.

Figure 13 .
Figure 13.(a) Schematic configuration of the Fe-FET and the corresponding working mechanism.Reprinted with permission from [244].Copyright (2019) American Chemical Society.(b) Schematic diagram of back-gated monolayer MoS2 FET as optoelectronic synapse.(c) The STDP behavior achieved based on MoS2 based synaptic FET.(b) and (c) Reproduced from [247].CC BY 4.0.(d) The structure diagram of the FG-FET, where MoS2 is the semiconductor channel layer, h-BN is tunneling layer, and the few layers graphene (FLG) are the floating gate and contact electrodes.(f) Post-synaptic current (PSC) versus pulse number, demonstrating LTP/LTD characteristics emulated by MoS2 based FG-FET device.3000 distinct states obtained through a series program/erase pulses with the amplitude of ±15 V. (g) Nonlinearity analysis on the weight update of the experimental potentiation/depression results.(d)-(g) Reproduced from [249].CC BY 4.0.

Figure 14 .
Figure 14.(a) Device structure diagram formed on a substrate by the diffusion of patterned ions.Containing a synaptic transistor with Na + and a conventional transistor without Na + .(b) Diagram illustrating a three-layer neural network consisting of an input layer, an operation layer, and an output layer.(c) and (d) Recognition accuracy of small (c) and large (d) images with 8 × 8 pixels following the training process.(a)-(d) Reprinted with permission from [265].Copyright (2021) American Chemical Society.(e) Schematic of device I (right, VP-MoS2 heterostructure) and device II (left, MoS2).(f) Recognition accuracy of two data sets.Red lines: ideal device; blue lines: large-dynamic-range VP-MoS2 device with 128 conductance states; violet line: large-dynamic-range VP-MoS2 device with 30 conductance states, respectively.The findings suggest that the heterostructure device exhibits minimal deviation from the ideal device in both classification tasks.(e) and (f) Reproduced from [264].CC BY 4.0.(g) Device structure of CsFAMA photovoltaic sensors.(h) Evolution of image recognition accuracy of automobile, aircraft and bird with adaptive time.(g) and (h) Reproduced from [268].CC BY 4.0.(i) CIFAR-10 dataset and CNN comprising convolutional, pooling, and fully connected layers.(j) Recognition rate for the CIFAR-10 dataset.(i) and (j) [231] John Wiley & Sons.[© 2021 Wiley-VCH GmbH].