## Introduction

Visual input, as one of our most important sensory functions, plays a critical role in human perception. More than 80% of the information received from the external environment is from vision1,2,3. Human vision is fundamentally a memory-based process, as the sensory neurons in the retina can not only detect light signals, but they also preform image preprocessing before more complicated visual information processing takes place in the visual cortex4,5. Existing CMOS-based artificial intelligence vision systems are composed of a photoreceptive chip, an analog-to-digital converter that transforms electrical input into digital signals, and an external artificial neural network (ANN) that preforms complex image processing tasks4,6. However, the physical separation of the functional components generates a large amount of redundant data during storage and transfer processing, which in turn leads to delays in data access and high power consumption. In addition, with the rapid growth of sensory nodes, bandwidth limitations make it difficult to send all data back to central or cloud computers quickly to realize real-time processing1,4,7,8. For this reason, the development of multifunctional electronic devices integrating sensing, memory, and processing functions is an effective way to improve the efficiency of artificial vision systems9. Optoelectronic neuromorphic sensors with both the sensing characteristics for light stimulation and nonvolatile multi-level storage characteristics provide a good choice for the development of artificial vision systems. Recent studies have shown that these sensors can perform image preprocessing and neuromorphic computing functions for machine vision systems4,6,10.

Most of the reported studies focus on the development of neuromorphic sensors operating in visible range, which are by design aimed to be alternatives to the human visual system. To ensure their survival and reproduction, most animal species have the capability of recognizing and perceiving ultraviolet (UV) light. For example, bees have developed an amazing ability to navigate and locate flowers using their UV-sensitive visual and nervous systems3, while reindeer can identify ground moss under snow in faint light by perceiving the intensity of the reflected UV light11. On the other hand, depending on its intensity, duration, and frequency of exposure and other factors, UV light can cause premature aging, skin cancer, macular degeneration, cataracts, and other ailments12,13. Since human beings cannot perceive this wavelength, the development of UV neuromorphic sensors can complement humans’ understanding of UV light and be instrumental for different applications such as biological sensors, healthcare devices, rocket early warning and missile detection11,14. However, the reported UV optoelectronic synapses are mainly based on the charge trapping/detrapping effect, which results in large writing non-linearity. Moreover, since these devices need to separate the photo-generated electron-hole pairs to achieve non-volatile memory characteristics, they are generally arranged as multi-layer structures, which increases the difficulty of large-scale industrial fabrication. Materials that respond to UV stimuli in conjunction with non-volatile phase transformation could open new avenues for the realization of high-performance neuromorphic sensors.

As an archetypal Mott material, vanadium dioxide (VO2) undergoes a typical phase transition from the low-temperature monoclinic (M1) phase to the high-temperature rutile (R) phase at the critical temperature of ~341 K15,16. During the phase transition process, VO2 exhibits a sharp change of resistance with several orders of magnitude and a pronounced optical switching in the infrared region. Benefitting from this phase transition, VO2 has been widely exploited in novel electronic and optical applications such as smart windows17,18,19, bolometers for infrared detection20,21,22, switching devices23,24,25, and neuromorphic devices26,27,28. In particular, optical control of VO2’s phase transition at room temperature has great potential for investigating the intrinsic physical mechanism and realizing optical modulation devices. Currently, optical pumping is used to induce the photoexcitation insulator-metal phase transition, which promises to allow vital insights into the nature of each state and may lead to metastable new phases under non-equilibrium conditions29,30,31,32. However, being an ultrafast excitation, such optical means cannot introduce stable phase transition; instead, a transient process on the picosecond scale is induced. Several works show that the electrical properties of VO2 can be modulated through various means of irradiation, such as electron beams33, X-ray34, and even UV light35. These results indicate the possibility of photo-controlled phase transition in VO2.

In this study, we present a novel neuromorphic sensor based on the optical control of phase transition in VO2 films with UV light, and demonstrate that this device can realize UV light perception and multi-level storage functions. The proportion of monoclinic phase in the film decreases with UV radiation dose, indicating the tunability of the phase transformation introduced by the optical stimulation. Based on this mechanism, the optoelectronic synaptic functions with integrated sensing and non-volatile multilevel storage features are successfully realized in VO2 grown on both Al2O3 and Si substrates. Using the optoelectronic synapses as sensing units, an ANN is constructed to realize the image sensing and memorization functions. The neuromorphic sensor array can extract the UV information from the surrounding environment, which significantly improves the image recognition rate on the MNIST handwritten dataset from 24% to 93%.

## Results

### Light-dosage-dependent synaptic plasticity

We grew epitaxial VO2 films with a transition temperature of about 341 K on r-Al2O3 substrates using pulsed laser deposition (PLD) technique. The high quality of VO2 films was confirmed using an atomic force microscope image (AFM) and through its X-ray diffraction (XRD) pattern (Supplementary Figure 1). Then, we fabricated the film into an optoelectronic transistor. The schematic diagram of the device structure is shown in Fig. 1a. More details about the device fabrication can be found in the Methods Section. Ohmic contact was exhibited between the source and drain electrodes (Supplementary Fig. 2). The temporal changes in the drain currents ID were measured under red (650 nm), green (532 nm), blue (450 nm), and UV (375 nm) light at an intensity of 64 mW/cm2. As shown in Fig. 1b, the transistor exposed to UV light exhibits non-volatility, while the ID irradiated under visible light returned to its initial state. The different behaviors of ID under visible and UV light are due to the different modulation mechanisms, which will be discussed in detail further below. Moreover, we investigated the effect of the light exposure on the channel current at different wavelength (Supplementary Fig. 3). As the illumination intensity increased, so did the photocurrent; however, only the device illuminated using UV light exhibited obvious non-volatile behavior. The transistor also exhibited weak non-volatile tunability under blue light, which can be due to the larger photon energy compared to the other visible lights used. It should be noted that this change was very small compared with that of UV illumination. In order to verify that the non-volatility is only dependent on the light wavelength, we irradiate the device under a stronger light intensity with 550 mW/cm2 at 532 nm (Supplementary Fig. 4). Although the device takes a longer relaxation time, it will eventually return to the initial state, showing a volatile characteristic. Since the transistor exhibited a synaptic property under UV exposure, we emulated other basic features of synaptic plasticity to simulate the learning and memory functions.

Figure 1c shows the stepwise increase of ID under illumination for six different durations using a constant light intensity of 84 mW/cm2. The durations were 1 s, 10 s, 50 s, 100 s, 150 s, 200 s, respectively, and the channel current was monitored at a small VD of 50 mV. The result indicated that ID increased along with the increase of exposure duration and good stability was demonstrated in each state. Then, we chose 10 s as the light pulse width while keeping the other conditions fixed, and measured the excitatory postsynaptic current (EPSC) response of neuromorphic transistor at different pulse numbers and different pulse intervals (Fig. 1d and Supplementary Fig. 5). It is found that both a pulse number increase and a pulse interval decrease lead to a significant enhancement of the synaptic strength. Furthermore, the pulse-switching characteristics of optical potentiation (light intensity of 84 mW/cm2, duration of 20 s) and electrical depression (voltage of −2.5 V, duration of 20 s) was studied in Fig. 1e. The channel current of the transistor can be reversibly switched between high- and low-current states dozens of times without significant degradation. The long-term synaptic plasticity, which includes the long-term potentiation (LTP) and long-term depression (LTD), was also simulated using our transistor (Fig. 1f). We applied 50 consecutive photonic pulses at an intensity of 84 mW/cm2 and a pulse duration of 10 s to emulate LTP. In contrast, the LTD appeared when 50 VG pulses were applied to the gate electrode (voltage varying from −1.5 V to −3.5 V, duration of 10 s). Here, electrolyte gating was utilized to achieve low voltage regulation due to its electric double layer effect, which can reduce device energy consumption effectively27,36. The results show that under optical writing and electrical erasing for programming, the device can be controlled continuously and in an adjustable multi-state non-volatile manner. The non-linearity values of potentiation and depression were calculated as 0.2 and 1.1, respectively. More details about the calculation formulas of non-linearity values and corresponding fitting parameters can be found in Supplementary Note 1 and Supplementary Table 1. Obviously, LTP exhibited high linearly, while LTD exhibited a decrease in linearity due to factors such as the internal dynamics of the ionic liquid. The non-linearity factors of this VO2-based neuromorphic transistor were significantly lower compared with those reported in previous works (Supplementary Table 2). In order to ensure good stability at each state, we examined the retention characteristics after writing and erasing operations (Supplementary Fig. 7), where it was found that the channel current remained constant for at least 4000 s after each operation.

Based on this long-term memory property, the smart sensing and image memorization of the letter V was realized using a 3 × 3 array consisting of VO2 transistors (Fig. 2). Laser light with the wavelengths of 650 nm and 375 nm at an intensity of 64 mW/cm2 were used to write this letter. Supplementary Figure 8 shows the simplified schematic of the illumination pattern. The changes of channel current (ΔID) were normalized to 0-1 for the initial input signal and expressed by the shade of color. The images of letter V were all successfully input into the synapse array after 500 s exposure duration using the two light sources. The overall color of the letter written using the red light was significantly lighter than that of the letter written using UV light, indicating the small ΔID obtained under red light illumination. After removing the light stimuli, the ΔID of the array excited by red light almost disappeared after 500 s, while the ΔID stimulated by UV light decreased slightly at 1000 s and remained unchanged at 2500 s. This phenomenon indicates that the VO2-based neuromorphic synapse array can store UV information selectively. In order to demonstrate the erasing/writing operations in a more intuitive way, we erased the letter using a voltage pulse (−2V for a duration of 100 s), and rewritten it using UV light under the same conditions. The result shows that after erasing, the channel current almost returned to its initial state and remained stable for the next 500 s. The ΔID of the letter V after repeated writing was almost the same as the previous time. The above processes suggest that the VO2-based neuromorphic sensor array has excellent image memory capability and visible-blindness feature for the non-volatile change.

### Photo-induced non-volatile phase transition

Next, we studied the underlying mechanism of selective memory property of VO2 at different wavelengths. The volatile response to visible light can be explained as the rapid recombination of photo-generated electron-hole pairs, while the non-volatile response to UV light could be ascribed to a photo-induced phase transition. To investigate the effects of UV light, we studied the temperature dependence of resistance at various UV light exposure durations (Fig. 3a). The as-grown VO2 film showed a typical phase transition, with resistance changing by three orders of magnitude. After UV light irradiation, the value of resistance in the low-temperature insulating phase gradually decreased, and was comparable to that of the metallic state after 30 h exposure. Moreover, we examined the response of channel current to UV light in different atmospheres (Supplementary Fig. 9). It was found that the ID had a wide range of changes and good retention characteristic after irradiation in nitrogen and vacuum conditions. On the contrary, the current increase under an oxygen atmosphere was not obvious and subsided quickly after the light was removed. We can speculate that oxygen plays an important role in the optical control of phase transition, as is indicated by the difference in results obtained under oxygen-enriched and oxygen-deficient environment.

A Raman scattering experiment was employed to determine whether the UV light irradiation process is accompanied by structural phase transition in the VO2 film. The as-grown VO2 film exhibited typical M1 phase characteristics, with Raman peaks at 146, 198 (Ag), 226 (Ag), 262 (Bg), 312 (Ag), 339 (Ag), 390 (Ag), 443 (Bg), 499 (Bg), 617 (Ag), and 827 cm-1 (Bg)38,39 (Fig. 3d). As the exposure duration increased, metallic domains were gradually formed in the film, which was reflected in the Raman spectra as a sharp rise in the luminescence background (position indicated by arrow)39,40. Although the films mainly maintain the M1 phase under the irradiation durations of less than 40 h, the intensity of the characteristic peak of this phase significantly weakened. Finally, after an exposure duration of 40 h, a broad band between 200 and 1,000 cm-1 appeared in the spectra, proving that the VO2 structure was completely transformed from M1 phase to the R phase. The M1 phase portion was estimated from the Raman results as a function of exposure duration (Supplementary Fig. 11). The results show that along with the electronic structure phase transition, a structural phase transition also appeared during the optical control process.

Then we discussed the physical mechanism of VO2 neuromorphic sensor as shown in Fig. 3e. Since the activation energy for creating oxygen vacancies was calculated to be between 3 and 3.5 eV35, 375 nm UV light with a photon energy of 3.35 eV should be capable to release oxygen from the VO2 film under an oxygen-deficient environment to create oxygen vacancies in the crystal lattice. Red and green light cannot release the oxygen from the lattice, since their photon energies are lower than the activation energy of oxygen vacancy41,42, regardless of their light intensities. With the appearance of oxygen vacancies, the V atoms lose a few electrons and release them to the neighboring V-3d states; these electrons partially occupying the d// and π* orbitals, leading to an electronic phase transition. Moreover, the oxygen vacancies in lattice and the differences in V ionic radius caused by the electrons’ release also lead to a strain in VO2, which transforms it from a low-symmetry monoclinic phase to a high-symmetry rutile phase and further induced the metallic phase43. This structural phase transition was further confirmed through the XRD pattern shown in Supplementary Fig. 12 and the related Supplementary Note 2. During the reset process, electrolyte gating could insert the oxygen ions back into the crystal lattice under a negative voltage44. With the decrease of oxygen vacancies in the channel, the VO2 structure gradually returns to its initial insulating monoclinic phase. In this manner, a reversible phase transition is achieved at room temperature through optical programing and electrical erasing. Since the transformation process is a photo-induced non-volatile phase transition and the metallic phase proportion increases almost linearly with the irradiation dosage, the device conductance shows good retention and linear dependency.

### Device performance on silicon wafer

We deposited VO2 film on a two-inch SiO2/Si wafer by magnetron sputtering technique, to further prove its silicon compatible potential. In order to study the structure of VO2 sputtered on Si substrates, we carried out a series of characterization experiments (Supplementary Fig. 13). The temperature dependence of resistance exhibits a significant change in 3 orders of magnitude, indicating that sputtered VO2 also has a typical phase transition characteristic. The phase composition is analyzed by powder X-ray diffraction and Raman spectroscopy. The film exhibits polycrystalline properties, mainly containing strong VO2 (011)M1 family peaks (space group P21/c) and a weak ($$\bar{4}02$$)M2 peak (space group: C2/m). This result can be further verified by Raman spectrum. The sputtered VO2 film exhibits strong M1 phase characteristics, and is accompanied by weak M2 phase (131.09 cm−1) and A phase (966.88 cm−1) peaks45. The VO2 film grown by PLD is pure M1 phase, and it is found that the photo-induced phase transition is caused by the transition from M1 phase to R phase. Although the films grown using two methods have some differences (for example, temperature window and crystal orientation), the VO2 film sputtered on Si substrates dominated by the M1 phase also exhibits UV photo-induced phase transition similar to the VO2 epitaxial film grown on Al2O3 substrates.

Then, a 3 × 3 device array was fabricated with the same device structure as prepared on r-Al2O3, each array having 103 devices (Fig. 4a). We conducted the same optical writing operations to verify the photo-induced phase transition characteristics of silicon-based devices. We randomly selected 100 devices from the arrays, and examined their channel resistance and response to UV light (Fig. 4b, c). The IV curves distribution of the devices are relatively concentrated, and the resistance histogram (Fig. 4b inset) shows that the overall device resistance on the Si wafer is ~2 MΩ, which reflects the uniformity of the film growth. The histogram of the statistical distribution of the photo response shows that after 100 s of UV irradiation at 84 mW/cm2, 96% of the devices have a channel current change of more than 2 nA. The fitting results show that the distribution of ΔID was a normal distribution. The selective memorization tested under different wavelengths of light showed that the silicon-based device also had the non-volatile memory characteristics for UV light only (Supplementary Fig. 14). In addition, the changes of channel currents were tested against UV exposure duration and UV light intensity as depicted in Fig. 4d and Supplementary Fig. 14, respectively. The results show that the multi-level memory feature of the device can be adjusted by controlling the UV irradiation conditions. Moreover, we carried out the optical programming and electrical erasing operations on the transistors (Fig. 4e), which showed reversibility and retention characteristics. To further characterize its non-volatile multi-level features, a series of UV light pulses (intensity of 84 mW/cm2, duration of 10 s) were used to program the device (Supplementary Fig. 15). Throughout the writing process, ID showed LTP synaptic plasticity and the channel current exhibited almost the same response to each UV pulse. We extracted the accumulation of ΔID and plotted it in Fig. 4f along with the UV dose, which can be calculated by the following equation: UV dose (mJ/cm²) = UV Intensity (mW/cm²) × Exposure Time (s). The curve of ΔID dependence of the UV dose is fitted well using a power function with a power of 0.92. The above results indicate that the VO2 grown on SiO2/Si wafer has the same perception and storage characteristics of UV light. The wafer-scale integration capability of VO2 lays a good foundation for future applications of neuromorphic UV sensors (Supplementary Table 2).

Furthermore, we investigated the effect of device size on the non-volatile channel current change under the same UV dose. Considering the effect of UV radiation on the channel in the out-of-plane direction, the channel conductance G can be described as: $$G=\frac{W}{L}{\int }_{0}^{H}\sigma \left(h\right){dh}$$, where W, H, L are the width, thickness, and length of the channel, respectively. σ(h) is the conductivity of VO2 channel, which is the function of the depth. The integration part is named as σs. The change of σs is related to the concentration of the induced oxygen vacancies determined by UV dose, and is independent on the lateral size of the device. The change in channel current ΔID can be represented by $$\triangle {I}_{D}={{V}_{D}\triangle \sigma }_{S}\frac{W}{L}$$. It can be seen that ΔID is independent of the device area. Besides, VO2 films down to nanoscale still have the phase transition characteristics28,46,47. Therefore, scaling down will not affect UV neuromorphic characteristics of the device. In addition, the device performance can be further improved by increasing the ratio of W/L.

### Image preprocessing and recognition

At present, most machine processing of visual information is in the light range which is visible to human beings48. This is because the visible light information is one of the main types of external information that guides human life. However, non-visible light also contains much important visual information, and this type of information plays an important role in guiding the behavior of creatures whose perceptible light range is different from humans’11. For example, the significant absorption of UV by nectar causes the stamen to be obviously darker than the petals in the UV range, and the perceptible light range of bees includes this UV part of the spectrum, which is imperceptible by human3. This characteristic UV information of nectar could help bees find the target flowers quickly during nectar collection. In addition, it is worth noting that when people try to identify certain characteristic information, redundant information will be automatically filtered out by the receptor, just like when people focus on a specific color, the rest of the color information will be mostly filtered out by their eyes. Such information extraction behavior can be defined by designing a suitable convolution kernel, which is a matrix of weight values used to perform a weighted average operation on pixels in a small area28,49. Since the proposed VO2 device shows a difference in its UV and visible light response, it can be used to simulate the behavior of bees focusing on UV information during nectar collection. A UV visual system with preprocessing (i.e. extraction of UV characteristic information) and recognition functions was modeled using computer simulation. The schematic diagram of its operation is shown in Fig. 5a. Based on the different functions implemented, the visual system was spatially divided into a convolution kernel array part for visual information preprocessing and an ANN part for image recognition after preprocessing.

To demonstrate the difference in image recognition with or without the ability to focus on UV information, the standard MNIST handwritten digital images (at a size of 28 × 28 pixels each) were used. An additional value independent of the RGB values was added in the computer simulation to introduce the invisible UV information into traditional RGB images. Based on the different responses of the device to 650 nm, 532 nm, 450 nm, and 375 nm light, each VO2 UV visual sensor was formed as a convolution kernel with a size of 1 × 1 × 4. Such convolution kernel performs weighted average processing on the four-color values (RGB and UV values) of a single pixel. After convolution, the resulting feature map reflected the scene that bees can observe when collecting nectar (i.e. an image with more abundant UV information and scarce visible light information). Subsequently, the preprocessed image was input into a fully connected (FC) ANN for recognition, which included an input layer (784 neurons), a hidden layer (300 neurons), and an output layer (10 neurons). The detailed operation mechanisms of the convolution kernel and the ANN are described in Supplementary Note 3.

For image recognition, the ANN part was firstly trained using the back-propagation algorithm and 60,000 images from the MNIST train dataset. Subsequently, three types of test datasets were fed into the ANN to compare the differences in image recognition accuracy under different conditions. These included the original MNIST test dataset, the same dataset with blurred visible light information, and the dataset after preprocessing, as shown in Fig. 5b. The second type of dataset was obtained by adding RGB Gaussian noise to the first dataset. Figure 5c shows the dependence of the recognition accuracy for these three datasets on the training epoch number. It can be seen that the recognition accuracy for images containing RGB Gaussian noise only reached about 24%, which was only slightly higher than the initial accuracy. This means that in this case, the recognition system can hardly recognize the characteristic information. In contrast, after the device preprocesses the UV information, the recognition accuracy of the image reached about 93%, which was the same as that obtained for the original MNIST. This result shows the effectiveness of the device in extracting ultraviolet information. In addition, this phenomenon was consistent with the fact that bees can identify which flower has nectar accurately, while humans cannot achieve this from visual information alone. Moreover, previous studies showed that the photoelectric response speed of VO2 is in the sub-picosecond scale50,51, ensuring its potential application in real-time monitoring systems.

## Discussion

In summary, we have successfully fabricated and demonstrated a VO2 optoelectronic synapse able to perceive and memorize UV light stimuli due to its photo-induced non-volatile phase transition. Benefitting from a phase conversion ratio linearly related to the light dosage, the device has linear writing and retention behaviors. Electrolyte gating was utilized as the electrical erasing process. Moreover, we fabricated a wafer-scale integrated neuromorphic sensor array and proved that the VO2 film on the silicon wafer also achieves optical control of phase transition, indicating the possibility of commercial mass production of neuromorphic sensors. In terms of high-density integration, it is worthy to further study the photo-induced phenomena of oxide materials with high phase transition temperature52. An ANN was simulated for the recognition of handwritten digit images from the MNIST dataset after the addition of random Gaussian noise. The results demonstrate that using the neuromorphic preprocessing process to reduce redundant data, the image recognition rate increased from 24% to 93%. Our work shows that VO2 has photo-induced non-volatile phase transition property and large-scale integration potential, which lays the foundation for the practical applications of neuromorphic sensor devices.

## Methods

### Sample preparation

The 20 nm VO2 thin films were epitaxially grown on r-plane ($$1\bar{1}02$$) Al2O3 substrates, using the pulsed laser deposition with a 308-nm XeCl excimer laser, an energy density of about 1 J/cm2 and a repletion rate of 3 Hz. The VO2/Al2O3 films were deposited at 485 °C in a flowing oxygen atmosphere with pressure 1.0 Pa. The samples were cooled down to room temperature at 20 °C/min. The deposition rate of VO2 films was calibrated by X-ray Reflection.

The wafer-scale VO2 film with a thickness of 20 nm was deposited on 2-inch SiO2 (300 nm)/Si wafer by RF magnetron sputtering using V2O5 target. It was performed with RF power of 150 W with a flow of 70 sccm Ar and working pressure of 7 mTorr. The oxygen content in the as-grown film was well-controlled after an annealing at 650 °C for 0.5 h under vacuum condition.

### Device fabrication

The thin films were patterned into channels with a coplanar gate structure using standard photolithography and argon-ion etching. The effective device area is 50 µm × 180 µm. The length between the gate electrode and channel is 10 µm. The 70 nm Pt layer was deposited as electrodes by RF sputtering. The transistor device was completed by dropping an ionic liquid N, N-diethyl-N-(2-methoxyethyl)-N-methylammoniumbis-(trifluoromethylsulphonyl)-imide (DEME-TFSI) on the channel and gate electrodes.

### Material characterization

X-ray diffraction patterns of the VO2 film was performed using a Rigaku SmartLab instrument with a 2θ range from 20 to 45° in step of 0.05°. XPS measurements were performed on ThermoFisher Scientific ESCALAB 250X under monochromatic Al Kα radiation with an energy of 1486.6 eV. XAS measurements were performed on via total electron yield method, and the background vacuum level was 6 × 10−7 Torr. Raman spectrum was analyzed using the alpha300 R microscope under 532 nm laser excitation. Powder X-ray diffraction pattern of the VO2 films sputtered on Si substrates was measured using a Rigaku Ultima IV instrument with a 2θ range from 20 to 60°.

### Device characterization

All the electrical characterizations were measured in a Lakeshore probe station with a Keithley 4200 semiconductor parameter analyzer in vacuum at room temperature. An UV laser with a wavelength of 375 nm were used for the optical switching in the experiment.