Ti₂C₃T_x MXene nanosheets were prepared by chemically etching the Al layers from purchased Ti₃AlC₂ powder (Jilin 11 Technology Co., Ltd.), followed by ultrasonic treatment. Specifically, 1.6 g LiF was added to 20 mL HCl (12 M) and stirred until dissolved. Then, 1 g Ti₃AlC₂ powder was slowly added to the above mixed solution and stirred in an oil bath at 40 °C for 24 h to completely etch the Al layers. Next, the reactants were washed by centrifugation with deionized water until almost neutral (pH ≥ 6). Subsequently, deionized water was added to the resulting sediment and sonicated in an ice bath for one hour to exfoliate the multilayer MXene. Finally, the mixture was centrifuged at 3500 rpm for 1 h and the supernatant containing single or few layers of MXene nanosheets was collected. The concentration of MXene colloidal solution was determined to be 6 mg mL⁻¹ by weighing the MXene film obtained by suction filtration. Violet Phosphorus dispersion (0.2 mg mL⁻¹) was purchased from Nanjing XFNANO Materials Tech Co., Ltd.

The prepared MXene colloidal solution was diluted to 0.6 mg mL⁻¹ with deionized water, and mixed with violet phospholene dispersion at a volume ratio (mL) of 0.01:1, 0.02:1 and 0.03:1, respectively. The resulting mixture was placed on a vortex mixer to mix it thoroughly. Au electrodes with a thickness of 60 nm were thermally evaporated onto poly(vinylidene fluoride) (PVDF) filter paper using a patterned shadow mask. The length and width of the channel are 50 and 2000 μm respectively. Then, the mixture of MXene and violet phospholene was coated onto the channel by suction filtration. Finally, the obtained device was dried in a nitrogen glove box at 100 °C for 10 min.

The morphology of the sample was characterized by scanning electron microscope (SEM, Zeiss Gemini500) and transmission electron microscope (TEM, JEOL JEM-F200). Raman spectrum was obtained by a Raman spectroscopy (LabRAM HR Evolution) equipped with a 532 nm laser. The chemical compositions and work functions of the samples were analyzed through X-ray photoelectron spectroscopy (XPS, Thermo Fisher ESCALAB Xi +). The steady-state PL spectroscopy tests were performed by a transient steady-state fluorescence spectrometer (Edinburgh FLS9) with a 532 nm laser. Environments with different gas atmospheres and RH were created by placing corresponding volatile liquids and saturated salt solutions in sealed test chambers. All electrical tests were performed through a digital source meter (Keithley 2410). Two lasers with wavelengths of 360 nm and 532 nm were used as illumination sources and the light power was calibrated by a NOVA II power meter (OPHIR photonics, Israel). The light pulse signal was generated by an electronic timer (GCI-73).

The device function of visual-olfactory crossmodal perception is manifested by different synaptic behaviors in different gas environments excited by light pulses, which requires rational material selection and device design at the begining. Figure 1 presents an overall schematic diagram of the multiscale design route that focuses on device function. Figure 1a is the designed device configuration, in which a hybrid film composed of abundant MXene/VP heterostructures are suction filtered onto a filter paper with gold electrodes. Details regarding material preparation and device fabrication are described in the experimental section. Figure 1f shows the corresponding optical microscope image of the device. It can be seen that the MXene/VP hybrid nanosheets fill up the channel and form a uniform film. The suction filtration process effectively avoids uneven aggregation and coffee ring effect, which is beneficial to the transport of carriers between nanosheets. The morphologies and structures of MXene and VP nanosheets were ascertained by TEM and Raman spectroscopy (Fig. S1). At the microscale, complex carrier dynamics processes such as intersheet hopping transport and interfacial trapping of photogenerated carriers in the heterojunction network composed of MXene and VP nanosheets lead to the PPC effect (Fig. 1b), thus enabling the device to act as an optoelectronic synapse. Figure 1g presents a SEM image of the hybrid film. It can be seen that MXene and VP nanosheets are interconnected or overlapped with each other, forming a heterojunction network.

The energy dispersive X-ray spectroscopy (EDS) elemental mapping confirms the homogeneous mixing of MXene and VP nanosheets (Fig. S2). At the nanoscale, a local van der Waals heterojunction with a built-in electric field is formed between adjacent MXene and VP nanosheets (Fig. 1c). The built-in electric field effectively promotes the separation of photogenerated carriers, and the excellent conductivity of MXene further improves the transport of carriers, thus significantly enhancing the photoresponse of VP. Meanwhile, gas molecules are easily adsorbed on the surface of 2D materials with large specific surface area, altering the optoelectronic response of the devices through the charge transfer with the nanosheets and the effects on the heterojunction barrier. Figure 1h is a TEM image of the hybrid film, from which a heterostructure of MXene/VP can be clearly found, exactly as envisioned in Fig. 1c. While at the atomic scale, the unique electronic structure of P atoms with a lone pair of electrons endows VP with high reactivity for gas adsorption (Fig. 1d) [30,39], thereby facilitating olfactory perception.

In order to analyze the energy band structure of the MXene/VP heterojunction and accordingly clarify the carrier transfer process, UPS, KPFM and steady-state PL spectroscopy tests were performed. Figure 2a and b shows the UPS spectra of VP and MXene. The work functions of VP and MXene can be obtained by subtracting the cut-off energy (17.09 and 16.85 eV) from the excitation energy (He I, 21.22 eV) as 4.13 and 4.37 eV, respectively. Furthermore, the difference between the Fermi level (E_f) and the valence band top (E_V) of VP is determined to be 1.77 eV by calculating the lowest binding energy, while that value of MXene is 0 eV due to its metallic nature. Combined with the band gap value of 2.31 eV of VP (Fig. 3a), we can depict the band structure diagram of MXene and VP before and after contact as shown in Fig. 2c. When they are assembled into van der Waals heterojunctions, the E_f tends to be the same, causing the conduction band (E_C) of VP to bend upward toward MXene and form a Schottky junction. In the depletion region of the Schottky junction, there is a built-in electric field directed from VP to MXene. Driven by this built-in electric field, photogenerated holes in VP are injected into MXene under illumination, while electrons remain in VP. As a result, the separation and transport of photogenerated carriers are effectively promoted, which improves the photoelectric response of VP.

To further confirm the formation of local Schottky junctions, we measured the contact potential difference (CPD) between MXene and VP by KPFM. Note that in order to facilitate the positioning of the MXene/VP heterojunction, a layer of Mxene film was spin-coated on the sapphire substrate, and then VP nanosheets were spin-coated on the MXene film. Figure 2d shows the atomic force microscope (AFM) image of VP nanosheets on the surface of MXene film for the KPFM measurement. The height profile along the white dashed line shows that the thickness of the VP nanosheet is about 200 nm. The KPFM image in Fig. 2e shows the surface potential mapping corresponding to the heterojunction in Fig. 2d. It can be clearly seen that the surface potential of MXene is obviously lower than that of VP and there is a CPD of about 75 mV between them. This result indicates that the work function of MXene is greater than that of VP and there is a built-in electric field from VP to MXene, which is consistent with the UPS analysis. Figure 2f shows the PL spectra of pure VP and MXene/VP heterojunction. Pure VP exhibits a high PL intensity resulting from strong radiative recombination, but it is clearly quenched when VP forms heterojunctions with MXene, implying that the recombination of e-h pairs in VP is suppressed.

As discussed above, the addition of MXene greatly improves the photoelectric response of VP, so we first investigated the photoelectric response of the hybrid film with a volume ratio of 0.02:1 (mL) of MXene and VP dispersions. Figure 3a shows the optical absorption spectrum of the VP nanosheet dispersion. The absorbance increases monotonically with decreasing wavelength, with obviously enhanced absorption below 600 nm. The inset shows that the optical band gap is calculated to be 2.31 eV by the Tauc plot method, which is close to the values reported in literatures [29,31]. In view of the stronger absorption of VP in the UV band, we first tested the photoelectric response of the device to 360 nm UV light. Figure 3b shows the I-V curves in the dark and under illumination with different UV light powers. The current under illumination is greater than that in the dark at any bias voltage and increases continuously with the increase of light power. The photocurrent under different light power densities at 1 V is plotted in Fig. 3c, and their quantitative relationship can be fitted as a power law by Eq. (1):

where I_ph is the photocurrent, P is the light power density and α (0 <  α < 1) is the power index reflecting the linearity between the photocurrent and light power density. The fitted power index is 0.32, which deviates from the ideal value of 1. The deviation from the linear response is attributed to the complex process of carrier generation, transfer, trapping and recombination in the hybrid film [40,41]. Responsivity is the core parameter to measure the photoelectric response level and can be calculated by Eq. (2):

where R is the responsivity and S is the effective illumination area. As shown in Fig. 3c, the responsivity is negatively correlated with light power density with a maximum value of 7.7 A W⁻¹.

In contrast, the pure VP device shows a photocurrent of only 100 pA at 10 V bias, and the responsivity was 0.36 μA W⁻¹ (see Fig. S3), which was 7 orders of magnitude lower than that of the MXene/VP device. Even compared with pure VP devices based on Si substrates reported in literatures, the responsivity of our device is still 3-6 orders of magnitude higher [31,33]. A detailed comparison of responsivity is shown in Fig. S4. Such a huge enhancement of the responsivity fully proves that forming heterojunctions with MXene can greatly improve the photoelectric response of VP. The specific detectivity (D^*, reflecting the ability to detect weak light) and the external quantum efficiency (EQE, representing the ratio of collected carriers to incident photons) were also calculated by Eqs. (3) and (4):

where q is the elementary charge, I_d is the dark current, h is the Plank constant, c is the light velocity and λ is the wavelength. As shown in Fig. 3d, the D^* value calculated at the minimum light power density is 6.73 × 10¹⁰ Jones, which is comparable with previous reported 2D materials-based devices [42,43]. The obtained EQE is as high as 2.67 × 10³%, which is attributed to the enhanced carrier separation and collection efficiency by MXene. We also investigated the dynamic response characteristics of the MXene/VP heterojunction device. Figure 3e shows the time resolved photoelectric response under intermittent UV illumination. The device demonstrates good repeatability and stability of photoswitching behavior. The rise (t_r) and decay (t_d) times can be obtained from a single switching cycle enlarged in Fig. 3e. As shown in Fig. 3f, the rise time is about 3.4 s, but it takes a relatively long time for the current to decay after turning off the UV illumination. The decay process can be well fitted by an exponential function with two relaxation times shown in Eq. (5):

where I₀, I₁ and I₂ are constants. The fitted time constants of fast component (t₁) and slow component (t₂) are 1.7 and 24.7 s, respectively. The PPC effect is associated with complex carrier dynamics, such as intersheet hopping transport, trapping and de-trapping at heterojunction interfaces and oxidation defects of VP surface (analyzed by XPS in Fig. S5) [33,44]. The effect of different mixing ratios of MXene and VP on the photoelectric response performance has also been investigated, as discussed in detail in Fig. S6. Although the increase in the proportion of MXene will further improve the responsivity, the dark current will also increase significantly. Therefore, the device with a moderate ratio of MXene to VP (0.02:1) were selected for overall performance testing. In addition, it can be seen from Fig. 3a that VP also absorbs part of visible light, so we also studied the response characteristics of MXene/VP heterojunctions to visible light with a 532 nm laser as light source, the test results and discussion are shown in Fig. S7.

As above described, the PPC effect were observed in MXene/VP heterojunction device, which enables the device to function as an optoelectronic synapse to mimic visual perception and memory, with multiple characteristics of biological synapses such as excitatory postsynaptic currents (EPSC), paired-pulse facilitation (PPF), short-term plasticity (STP), long-term plasticity (LTP) and “learning-experience” behavior. Figure 4a schematically illustrates the biological synapse. Synapse acts as a communication site that transmits electrical or chemical signals between two neurons [45]. In our proof-of-concept MXene/VP optoelectronic synapse, UV light pulses are regarded as presynaptic stimulus and channel current as synaptic weight. An applied presynaptic spike (light pulse) can trigger an EPSC in the MXene/VP heterojunction channel at a small bias of 1 mV, as shown in Fig. 4b. The current decays rapidly and then slowly over time when the light is turned off, and is still 2.4% higher than the initial value after 100 s. This phenomenon is consistent with the decay process of postsynaptic potential in biological synapses [46]. When two consecutive presynaptic spikes are applied, the amplitude of the postsynaptic current will increase continuously, and the rate of increase (A₂/A₁) is defined as PPF, as shown in Fig. 4c. PPF is a typical manifestation of STP, which is closely related to the temporal recognition and decoding of visual signals in biological systems [22]. PPF strongly depends on the time interval (Δt) between two presynaptic spikes, so we measured the PPF index (PPF index = A₂/A₁ × 100%) at different time intervals. As shown in Fig. 4d, the PPF index decreases from 135 to 112% as Δt increases from 0.1 to 10 s after applying light pulse pairs with a duration of 1 s. This is consistent with the fact that there is more recombination of the trapped carriers within the longer Δt. The decay curve of the PPF exponent with Δt matches well with the double exponential function shown in Fig. 4d, with relaxation times τ₁ and τ₂ of 1.19 and 18.73 s, respectively, comparable in scale to those of a biological synapse [47,48].

In biological brains, the ability of synaptic weights to be modulated by neural activity is known as synaptic plasticity, which can be divided into STP and LTP according to the length of retention time and are an important bases for learning and memory [49]. Here, the transition from STP to LTP was also mimicked in the MXene/VP optoelectronic synapse by different stimulus modes. Figure 4e and f shows the EPSC curves under varying light pulse intensity and width. Similarly, with increase of the light pulse intensity (4.92 to 34.35 mW) and pulse width (0.1 to 1 s), both the amplitude and retention time of EPSC increase correspondingly, which is in analogy with the transition from STP to LTP in biological synapses [50]. In addition, the increase in the number of light pulses can also lead to the transition from STP to LTP (Fig. 5a), which is similar to the fact that frequent stimulation can increase the synaptic weight in biological synapses. Based on the biomimetic characteristics of the device, the “learning-experience” behavior is demonstrated, as shown in Fig. 5b. For our device, the light-on state corresponds to learning or relearning and the light-off state corresponds to the forgetting process. During the first learning process, the EPSC gradually increases when successive 10 light pulses are applied to the device, and then decays spontaneously after the light is removed. The EPSC retains 25% of the increment after 20 s, which means that the memory of the first learning is forgotten to 25%. In the second learning process, the EPSC returns to the maximum value by applying only 4 light pulses, and maintains a higher value than the first learning after the same decay time. This result corresponds to the fact that it takes less time to relearn forgotten information and that the memory will be enhanced after learning, surprisingly similar to the learning and memory behavior of the human brain [51]. Low power consumption is one of the fascinating advantages of biological brains, so we also consider the energy consumption of the MXene/VP optoelectronic synapse. The energy consumption for per synaptic event is defined as I_peak ×  V ×  t, where I_peak, V and t refer to the peak value of the EPSC, the applied bias voltage and the pulse duration time, respectively. As shown in Fig. 5c, the energy consumption increases with the increase of the pulse duration time at 1 mV bias and the minimum energy consumption is 14.7 pJ when the pulse duration time is 0.1 s, which is lower than those reported 2D material-based optoelectronic synapses [52,53,54,55]. A more detailed comparison with reported optoelectronic synapses based on 2D materials and their heterojunctions is presented in Table S1.

Vision is the most important way for humans to obtain outside information, and thus visual memory is the most effective memory mode of the human brain. Artificial synapses for the construction of neuromorphic vision should exhibit a good brain-like visual memory behavior. To this end, the “XJTU” letter pattern composed of 7 × 25 pixels is used as the input signal to investigate the visual memory features of the MXene/VP photoelectric synapse through a single-pixel imaging system. Figure 5d shows the evolution of the pixel grayscale of the letter pattern with the number of light pulses and decay time. Note here that the pixel grayscale derived from the normalized EPSC represents memory strength, the pulse number corresponds to the number of learning times, and the decay process represents forgetting. It can be seen that the letter pattern after first study is not very clear due to the low memory strength, but the distinct target image can be recognized after 10 times of learning. Subsequently, the grayscale reduction of the image during the forgetting process was also recorded. As time goes by, the recognizability of the letter pattern gradually decreases, and the memory strength of the image is obviously weakened after 1 min. The Ebbinghaus forgetting curve illustrates how information is lost in the human brain over time, which can be expressed quantitatively by Eq. (6) [56]:

where P is the possibility of recall, t is time, and τ is the specific relaxation time and β is the exponent that ranges from 0 to 1. As shown in the inset of Fig. 5d, the forgetting process of the image fits the Ebbinghaus forgetting curve very well with the values of τ and β of 55 s and 0.54, which demonstrates the good brain-like learning and memory behavior of the MXene/VP optoelectronic synapse and its great potential in developing the human visual system.

The perception and memory processes of human are crossmodal, that is, the information input from different sense organs will interact in a synergistic or antagonistic manner, thereby strengthening the ability to perceive the environment or memory. For example, Tsushima et al. found that fragrance can change the human perception of the moving speed of pictures through psychological and physiological experiments [57]. Therefore, optoelectronic synapses with visual-olfactory crossmodal perception for developing biomimetic visual perception and memory are highly desired. Due to the excellent hydrophilicity of MXene, the large specific surface area of 2D materials and the unique electronic structure of VP, the gas atmosphere-dependent photoelectric response is expected for the MXene/VP optoelectronic synapse. Figure 6a illustrates the process of crossmodal interaction between vision and olfaction in the human perceptual system. Stimulation of specific odors during visual information input can result in enhancement or inhibition of visual perception. Figure 6b plots the EPSC curves stimulated by 10 UV pulses in different gas environments. It can be seen that amplitudes and decay times of the postsynaptic currents in different odor environments are distinct under the same visual stimulus, which is analogous to the visual-olfactory crossmodal perception process of human. The mechanism should be attributed to the different photoelectric gas-sensing processes of VP for various gases. Taking ammonia as an example, ammonia molecules with lone electrons are adsorbed on the surface of VP and act as electron donors to release electrons to conduction band of VP, which increases the carrier density and shrinks the depletion regions of the Schottky junctions, thereby reducing the resistance of the hybrid film and increase the photocurrent. In order to quantitatively explore the differences in visual perception in environments with different odor levels, the photoelectric responses were tested in environments with different relative humidity (RH), where different concentrations of water molecules were regarded as different odor levels. Figure S8 shows the variation of the baseline (dark current) of EPSCs with the RH in the absence of light. The RH has a significant inhibitory effect on dark current, because water molecules adsorbed on the surface of the nanosheets of the hybrid film expands the intersheet spacing, thereby increasing the tunneling resistance [58,59]. This interesting phenomenon inspires us that the inhibitory postsynaptic current (IPSC) can also be obtained in the MXene/VP optoelectronic synapse by applying humidity pulses, as usually this is difficult to achieve in planar two-terminal synaptic devices due to the absnce of gate [60,61,62]. As shown in Fig. S9, the long-term depression (LTD), a common and indispensable behavior in biological synapses, can also be achieved in the MXene/VP optoelectronic synapse by successive humidity pulses. Accordingly, LTP/LTD synaptic behaviors with light potentiation and humidity depression, which are normally only available in three-terminal synaptic devices, can also be realized in our two-terminal synaptic device, as shown in Fig. S10. In addition, the effect of the RH on the amplitude and decay time of the EPSC was also investigated. Figures S11 and 6c show the EPSC curves under the excitation of a single light pulse and 10 light pulses in different RH environments, respectively. Interestingly, the ΔEPSC (peak value minus baseline) increases but decay time decreases with increasing RH.

The increase of ΔEPSC is attributed to the reduction of dark current, while the reduced decay time may be related to the fact the H⁺ and OH⁻ produced by the dissociation of water molecules combine with electrons and holes, accelerating the disappearance process of the trapped photogenerated carriers. Figure 6d and e maps the recognition and forgetting processes of the letter “X” in environments with different RH. It can be seen that the higher the RH (that is, the higher odor level), the more rapidly the EPSC increases but also the faster it decays. If we still assumed that the pixel grayscale represents the memory strength, we can conclude a typical visual-olfactory crossmodal perception behavior of this photoelectric synapse, that is, visual information is perceived more strongly in environments with high odor levels, but is also forgotten more quickly over time after the visual stimulus is removed. This phenomenon is just like the memory behavior of the human brain discussed in Bjork's memory theory—retrieval strength of memory is negatively correlated with storage strength [63,64]. This is the first report of such complex memory behavior in optoelectronic synapses and demonstrates the bright application prospects of MXene/VP heterojunctions in crossmodal neuromorphic vision sensors.