Researches on artificial neural networks (ANN) which could offer both storage and computing have been carried out extensively in order to obtain a deeper understanding of human vision and brain^{1, 2}. For example, synaptic transistors, memristors³, phase change memories, etc. have been widely applied to simulate biological neurons^4-6. Furthermore, the spiking neural networks (SNN), as the third generation of ANN, are the biologically-realistic models of living nervous systems taking discrete events instead of continuous values as stimuli⁷. From a single device to an array of devices, increased number of more complex functions could be realized, such as the field of artificial intelligence (AI) related to vision^8-11. In addition, associative learning based on the interaction between thoughts and experience is one of the most important capabilities of human brain cells¹². The classical Pavlovian conditioning links two stimuli together to produce a new learned response in a person or animal¹³. Recently, different materials and various structures are used to stimulate the effective performance of biological synapses^{1, 14, 15}. The main significant figure of merit are sensitivity and power consumption for evaluating neural devices¹⁶. The ideal synaptic transistors have strong synaptic plasticity and lower power consumption than bio-synapses which consumes about 10 fJ per spike^{1, 17, 18}. Synaptic plasticity is divided into long-term synaptic plasticity (LTP) closely related to learning and memory and short-term synaptic plasticity (STP) processing the temporal information on the relevant time scales such as milliseconds or seconds^{12, 17}. According to Hebbian theory, also called spike timing-dependent plasticity (STDP), persistent or repeated stimuli give rise to a higher synaptic weight which represents the strength of connection between two nodes^19-21. In addition, spike-duration dependent plasticity, spike-number dependent plasticity, spike-frequency dependent plasticity, learning-experience behavior and forgetting behavior are also important parameters to measure the properties of synaptic devices^{17, 22}.

Photonic synapses, coupling optical sensing and synaptic plasticity, are expected to reduce power consumption and achieve ultrafast signal transmission^{23, 24}. The light-sensitive photosensors are similar to the photoreceptors of biological retina because they both can convert light signals to electrical signals²⁵. This kind of devices, which could mimic the biologic eyes and work as a standalone platform, would pave the way for machine vision systems in the future²⁶. Moreover, Lee’s group have extended the human visual sense beyond visible light to UV light and this work may be further applied to healthcare via combining UV selectivity and dose-calculation²⁷.

Optical modulation as a novel gate-control mode for achieving multi-terminal neuromodulations could decrease thermal loss and operate simply¹⁷. The light-stimulated synaptic device offers more possibilities to broaden the functionalities and system complexities of the current neuromorphic systems²⁴. Two-dimensional (2D) transition metal dichalcogenides (TMDs) are intriguing and promising optoelectronic devices due to their superb electrical and optical properties^{28, 29}. Many kinds of 2D TMDs could be synthesized in large-area by chemical vapor deposition (CVD) or other methods, making practical devices possible^30-34. Molybdenum disulfide (MoS₂) has been investigated as an efficient charge-trapping layer and light absorption layer in photonic synapses recently due to fast charge-trapping rate, good trapped charge-retention, good optical absorption^{35, 36} and high carrier mobilities^{1, 17, 18, 37, 38}. Yoo’s group³⁷ has confirmed that the MoS₂ layer could act as a high-efficiency charge-trapping layer via control experiment. Although the two-terminal devices are easier to form a crossbar-array structure, the inhibitory synaptic behaviours cannot be achieved in these devices³⁹. Furthermore, three-terminal gate-control synaptic transistors are able to possess another degree-of-freedom to modulate the neuromorphic function apart from optical modulation^{8, 40}. For instance, Chu’s group⁴⁰ have shown a gate tunable artificial synapse to mimic the modulation of drug dosage. In addition, MoS₂ with intrinsic flexibility could be applied to curved image sensor array for the machine vision⁴¹.

The proposed mechanism for photonic synapses is charge trapping/detrapping or phase change^{5, 39}. The former could achieve low-energy consumption instead of the requirement of high power supplies for the latter³⁹. In addition, the trap centers capturing and releasing carriers could be material defects, surface dangling bonds, potential wells³⁵, etc⁴². The persistent photocurrent/photoconductivity caused by these trap centers make the device be able to mimic some specific functions of biological synapses²⁶. Given that, the TMDs with outstanding photoelectric conversion properties and various defects as carrier traps show great potential acting as optically active layers for photonic synapses³⁹. Additionally, the von Neumann bottleneck caused by separated memory units and processing makes synaptic devices with low power consumption more desired⁴³. However, electrical synapses under low-energy consumption¹⁷ could be achieved via ultrashort electrical pulses while the photonic synapses with ultralow power consumption to attojoule (aJ) level are still few, especially via ultrashort optical pulses⁴⁴.

In this work, the large-area growth 2D MoS₂ films and an array of MoS₂ synaptic transistors integrating the sensing-memory-preprocessing capabilities are demonstrated. The photosensing scope of our devices broadens humanoid vision to near-infrared light. Moreover, the photonic synapses under visible light possess hypersensitivity to transient single pulse and ultralow power consumption superior to bio-synapses. Specifically, the synaptic plasticity via photoinduced charge transfer is investigated including STP, LTP, paired-pulse facilitation (PPF) and the forgetting behavior²³. The classical Pavlovian conditioning can also be verified by our synaptic device. In addition, back-gate bias could enhance or reduce memory function exhibiting similar effects of emotions²⁵. Positive emotions offer a positive effect and vice versa. The image pre-processing is achieved by the large contrast between STP and LTP. Given that, we can conclude such synaptic devices are promising building blocks in AI, especially visual perception and brain storage^{26, 27, 45}.

Synapses, i.e. neuronal junction, could transmit electric nerve impulses between neurons and effectors to achieve signal transmission¹⁷. In order to mimic the properties of synapses, MoS₂ synaptic transistors are prepared via low pressure CVD (LPCVD) for improved thickness uniformity³⁴. The details of material growth and device fabrication are presented in the Experimental section. The accurate control of material growth condition including pressure, temperature, the amount of precursors, the position of substrates and so on could improve the reproducibility of channel material. The optical image of the polycrystalline MoS₂ film on a Si/SiO₂ substrate is shown in Fig. 1(a). A continuous and relatively uniform large-area film over 2 cm² is grown successfully. Figure 1(b) shows the optical microscopy image of a part of the MoS₂ synaptic transistor array. The devices are fabricated by standard lithography with Ti/Au electrodes as the source, drain, and silicon as the backgate. The corresponding Raman spectrum in Fig. 1(c) shows the typical E_2g (381.9 cm^–1) and A_1g (407.3 cm^–1) peaks of MoS₂ measured by the 532 nm laser^{28, 33, 46}. The frequency difference between E_2g and A_1g is about 25.4 cm^–1 indicating multilayer growth, which was further confirmed by PL spectrum⁴⁷ in Fig. 1(d). Figure 1(e) and 1(f) show the corresponding Raman mapping around 381.4 cm^–1 and PL mapping selected randomly from Fig. 1(a). The material uniformity of as-grown MoS₂ films can be further confirmed by the uniform color contrast of Raman and PL signals in Fig. 1(e) and 1(f)⁴⁸. Figure 1(g) shows a schematic of the visual perception of human eyes and information transmission to the brain based on our MoS₂ synaptic transistors. The retina in human eyes could detect the visible light and then transmit the information to brain via a number of synapses²⁷. Under external stimuli such as optical and electrical pulse, the source and drain electrodes are used to liken pre- and post-synaptic terminals, respectively²⁵. Additionally, gate modulation can serve as another form of signal input for three-terminal artificial synapses^{8, 40}.

Photonic synapses which integrate sensing and synaptic function could mimic the visual perception of human retina²⁷. The bias voltage (V_ds) of 1 V was applied for the following measurements without special instructions. The visible light detection is depicted in Fig. 2(a) and three different wavelengths are chosen including 405 nm, 520 nm and 635 nm. Surprisingly, near-infrared (NIR) light such as 808 nm and 850 nm could also be detected without the upconverting nanoparticles mediation which could upconvert NIR to visible light or UV¹⁸ shown in Fig. 2(b). Because multilayer MoS₂ photodetectors could be responsive to NIR light while monolayer MoS₂ ones could only detect visible light^{49, 50}. This reveals the capability of photoreception of our device could be extended to NIR as human eyes could only detect the visible light. The broadband detection from visible to NIR light⁵¹ at room temperature based on our MoS₂ synaptic transistors provides new opportunities in machine vision. The insert in Fig. 2(b) shows a good ohmic contact due to linear output characteristics between metal electrodes and MoS₂^{14, 52-56}. Taking 405 nm laser as an example, synaptic function-related measurements are investigated. Figure 2(c) displays the excitatory post‐synaptic current (EPSC, i.e. Ids−Idark ) of the synaptic device under different single pulse width, showing the spike-duration plasticity¹⁷. According to Fig. 2(c), the MoS₂ transistor shows typical synaptic properties, i.e. the effects after stimuli do not disappear immediately with the disappearance of the stimuli. The proposed working mechanism based on our photonic synapses is that the charge trapping centers provided by the channel and interfaces reserve the photoinduced carriers even after removing the optical stimuli⁵⁷. That is to say, the persistent photocurrent/photoconductivity phenomenon is attributed to slow recombination of defect or trap centers³⁵. The building potential wells between MoS₂ and source-drain electrodes could reserve/release charge carriers with the modulation of light stimuli³⁵. The trap sites in the potential wells may originate from the sulfur vacancies in the channel and the dangling Si-O bonds at the interface of MoS₂/SiO₂³⁵. The proposed operating process of our photonic synapses might be the state of thermal equilibrium without optical/electrical stimuli at the beginning and generation of photocarrier under optical stimuli followed by carrier trapping, carrier recombination and returning to the original state under dark⁴³. In addition, the photocurrent will enhance as the single pulse width prolong obtained from Fig. 2(c) due to the production of more photogenerated carriers. Furthermore, the synaptic transistor distinctly responds to a very short optical pulse with a width of only 5 μs in Fig. 2(d) which is four orders of magnitude faster than the instantaneous information transmission in the biological synapses (typically 50 ms)⁵⁸. Possible reasons for the ultrashort optical pulse detection capability are ultrafast optical signal propagation speed^{2, 35, 39, 59} and ultrasensitive optical detection⁶⁰ based on the polycrystalline MoS₂ films. As a result, our synaptic transistors could improve the recognition rate of single optical pulse significantly and achieve an ultralow power consumption of 40 aJ, two orders of magnitude lower than that of biological synapses. The power consumption of one synaptic event could be estimated by the equation of E=Ipeak×Vread×t where I_peak, V_read and t are the maximum EPSC after a pulse, reading voltage and pulse width, respectively⁶. Our power consumption could be reduced by shortening the pulse width with constant reading voltage of 1 V, while many researches achieve low power consumption via decreasing the reading voltage^{44, 59, 61, 62} such as 10 μV. So far, the minimum power consumption of a single optical pulse is about femtojoule (fJ) level, well above that of ours (Fig. 2(e))^{5, 14, 17, 42-44, 59, 61-68}. The ultralow power consumption and ultrafast speed suggest the prepared device a promising candidate for an energy-efficient neuromorphic computing system¹⁷.

STP, also called short-term memory (STM), is a critical source information for associative learning and sound localization like classical Pavlovian conditioning¹². Figure 3(a) shows a clear and typical synaptic response at a single pulse width of 50 ms like the information transmission speed in biological synapses⁶⁹ and V_ds of 1 V. This optical pulse (405 nm) results in an acute current rise due to a sudden increase of photoinduced carriers⁷⁰. After the optical pulse is off, the EPSC exhibits a slow delay and relatively long relaxation time because the trapped holes by the surface states caused a low recombination rate of carriers^{12, 70}. Figure 3(b) and inset unfold the EPSC would enhance as the power density (P) of the single optical pulse (width: 100 ms). At the higher P, the faster decay time can be observed in Fig. 3(b) because the decay dynamics are dominated by shallow carrier traps rather than deep carrier traps at lower P⁷¹. Additionally, the sublinear relationship between EPSC and P in the inset of Fig.3(b) confirmed the presence of trap states further⁷². The sublinear relationship is attributed to the photogating effect⁷³ caused by long-lived charge trapping processes⁷¹. The PPF triggered by two consecutive presynaptic spikes at 2 s intervals is expressed in Fig. 3(c). A higher EPSC would be obtained after the second optical spike because of the effect of the first optical spike. The PPF index is defined as the ratio of A₂ to A₁ where A₂ and A₁ mean the EPSC triggered by the second or first light spike, respectively^{18, 74}. Figure 3(d) shows the PPF index increases as the pulse interval (Δt) decreases. This is in accordance with the behaviour of information processing in biological synapses¹⁴. Furthermore, the characteristic relaxation times of rapid and slow decay constants τ₁and τ₂ can be estimated according to the double-exponential decay function given by Eq. (1)^{18, 75}.

PPFindex=C1exp(−Δt/τ1)+C2exp(−Δt/τ2)+C0,

where C₀ equals to 1, which indicates the trend of PPF index is towards to 100% eventually⁶⁷. In our case, the values of τ₁ and τ₂ are 0.8 s and 9.2 s fitted from Fig.3(d), respectively. Additionally, the age-related learning characteristics could be emulated by our synaptic transistors thanks to the PPF characteristics which are varied with human age⁷⁶.

Compared with STP, the LTP, i.e. long-term memory (LTM), usually takes longer time which means the lasting potentiation of the synaptic weight¹². Figure 4(a–c) show the spike-number dependent LTP and forgetting behavior. The number of spikes could be regarded as the learning times⁷⁰. There is a continuous upward trend of current as the number of spikes increases due to more and more photoinduced carriers. This is similar to biological synapses because successive stimuli could induce more neurotransmitters⁶. This phonomenon could explain more and more high-efficiency relearning behavior of a human’s brain³⁹. The forgetting behavior is liken by the EPSC after the optical spikes¹⁴. At first, the forgetting behavior would decrease rapidly and then decay slowly which is coincident with human brain^{22, 70}. As seen in Fig. 4(c), the delay of the transient current after photostimulation is fast at the beginning and then becomes slow. The delay speed is dependent on three continuous stages including recombination, shallow and deep trapping procedures⁷⁷. The defect-induced recombination and traps make the corresponding fast and slow delay⁷⁸. In addition, the spike-frequency dependent LTP and forgetting behavior are shown in Fig. 4(d–f). The higher frequency of repeated stimuli could enhance the synaptic weight and improve the memory function or learning efficiency obtained from Fig. 4(d)^{17, 76}. Moreover, the MoS₂ synaptic transistor shows the characteristic of the biological high-pass filters which could be applied for filtering unnecessary information in signal processing as shown in Fig. 4(e)^{6, 14, 26, 70}. The EPSC would become saturated eventually even though the number or frequency of optical pulses increases continuously. This means the synaptic device would not cause excessive neural activity⁶. From Fig. 4(c) and 4(f), the forgetting processes follow the classical Ebbinghaus forgetting curve which indicates the memory decays over time^79-81. Furthermore, the higher-learning condition stimulated by more spike numbers or higher spike frequency leads to slower forgetting⁸². Note that the effect of single pulse duration time on synaptic plasticity is similar to spike number and the same is true for the Δt between two continuous stimuli and spike frequency⁹.

The Pavlovian conditioning, as unconscious or procedural memory, could help creatures to be prepared for something expected or likely¹³. This process, a classical example of the associative learning, is consist of the acquisition, extinction and recovery¹². The artificial synaptic devices could remember their past dynamic history, which is vital to mimic the Pavlovian conditioning^{12, 13}. Figure 5(a) shows the typical Pavlovian conditioning under one training. Here, the electrical spikes (E) of V_ds and light spikes (L) act as the bell and bones in the Pavlov’s dog experiment. The bell ringing would make the dog salivate after the training of feeding and ringing bells together¹³. In Fig. 5(a) and 5(b), the black dashed line represents the threshold value of salivation. The EPSC under E would not reach the threshold value initially. However, the later E could induce salivation for the next five minutes after the stimuli of E and L together (E + L) shown in Fig. 5(a). In addition, 10 successive E + L spikes are applied as repeated trainings in Fig. 5(b). The EPSC maximum is far more than the threshold value after 10 spikes and the salivation state would last at least half an hour. This means the classical Pavlovian conditioning could be mimicked successfully based on our MoS₂ synaptic device and more trainings could make better associative learning.

The backgate (V_g) could regulate the memory function of our synaptic device as well except for repeated optical stimuli or electrical stimuli of V_ds^{4, 14, 18, 83}. Figure 5(c) and 5(d) depict the memory function of the synaptic transistor could be modulated by changing the V_g simply. It functions like the emotional effect on the memory function of human brain^{84, 85}. Figure 5(c) shows the small negative V_g of –0.1 V could significantly enhance the transient current under 30 optical spikes than that at zero gate voltage. On the contrary, the positive V_g could make the transient current decrease drastically. This phenomenon of increased EPSC may be attributed to negative gate voltage assisted detrapping of electrons. Cheng’s group has reported that the free electron concentration or channel conductance increased due to released trapped electrons after negative gate voltage⁸⁶. On the contrary, the positive gate voltage would make the free eletrons in the channel easier be trapped by trap centers resulting in the reduced channel conductance⁸⁶. The inhibited current increment due to the applied positive gate voltage does not equate to the conventional inhibitory post-synaptic current^{24, 87}. The carrier trapping/detrapping under different positive/negative gate voltage is considered as the main operation mechanism of current switching behavior for the three-terminal artificial synapses^{1, 35, 57}. The negative V_g could act like the positive emotions, and vice versa when positive V_g for our device. It is well-known that someone could have a better memory function in a good mood. A more obvious gate-tunable phenomenon is shown in Fig. 5(d). The EPSC would increase to about five times at the V_g of –1 V than that without V_g. Thus, our synaptic device could be applied for mimicking the visual learning and memory process at different emotions.

Inspired by human retina, our synaptic device would be further applied for neuromorphic imaging and pre-processing due to the excellent light-stimulated synaptic properties and easy large-area growth⁴¹. Figure 6(a) shows the schematic diagram of human’s neuromorphic visual system. Given that, we stimulate the letter recognition of “I” under STP, LTP with or without -V_g shown in Fig. 6(b–d). The time needed for photocurrent decay to 1/e of its original state are 40 s and 2 min, respectively, for STP and LTP at zero V_g obtained from Fig. 6(b) and 6(c). The retention time of LTP is much longer than that of STP. The difference between LTP and STP is resulted from the different photo dosage which can make different effect on device conductance and decay dynamics³⁵. When the −V_g was applied on the synaptic device, the photocurrent improved significantly in Fig. 6(d). Additionally, the big difference between STP and LTP makes the MoS₂ synaptic device show a better contrast in the pre-processing which is very important for efficient machine vision^{15, 41}.

In summary, the uniform polycrystalline MoS₂ film could be a promising choice to mimic the visual perception and brain storage. The unique charge-trapping characteristics and facile synthesis of MoS₂ make it a favored candidate over other materials for bio-mimicking devices^{1, 17, 18, 37}. The photoreception capability of our device could expand the edge of vision from visible light to NIR light. Our photonic synapses could achieve information transfer and storage of the artificial brain with sensitive response to ultrashort optical pulses and the lowest power consumption under optical stimuli reported so far. The synaptic plasticity such as STP, PPF, LTP, forgetting behavior and associative learning are successful stimulated based on this proposed device. Moreover, the emotional effect on memory by regulating the backgate voltage is demonstrated. The channel conductance modulation is achieved by carrier trapping/detrapping in the channal and interfaces/surfaces¹. Eventually, like the neuromorphic imaging and pre-processing device, the prospective imaging effect is shown. This work which integrates sensing-memory-preprocessing capabilities provides an insight into artificial retinal system and artificial brain.

The polycrystalline MoS₂ was grown in a 1 in. quartz tube with a one-zone furnace. Mixed MoO₃ powder (8.7 mg) and NaCl (1.4 mg) together placed in the center of the furnace and S powder (255 mg) were used as the precursor for the MoS₂ growth. A pair of Si/SiO₂ substrates are placed face to face with the vertical distance about 2 mm in the downstream after rinsing by acetone and isopropanol. The whole CVD process is controlled at a low pressure of 0.4 Torr (1 Torr=133.3223684 Pa) with the Ar/H₂ (8∶1) as the carrier gas after purging out the air in the tube about 10 min. The S powder is ramped up to 220 °C using a heating belt when the temperature of the furnace center reaching 670 °C. Then, the furnace was maintained for 30 min when it was heated up to 750 °C. Finally, the tube was cooled to room temperature naturally after removing the furnace.

The Ti/Au (Ti: 10 nm, Au: 100 nm) metals were deposited by electron beam evaporation and then lifted off in acetone as the source and drain electrodes after the photolithography.

The optical images were captured by OLYMPUS microscope (LV100ND). The Raman and PL data were measured with a Raman-AFM confocal spectrometer (Witec, alpha300 RA) with a laser of 532 nm.

The optoelectronic properties of the photonic synapses were measured with the SemiProbe probe station and a semiconductor parameter analyzer (Keithley 4200) and Platform Design Automation (PDA, FS- Pro). Different duration of optical stimuli and different laser densities set by a Programmable DC Power Supply (Itech Electronic, IT6100B), a function generator and an irradiatometer were used to measure the photoresponse.