Introduction

Ultraviolet (UV) light not only plays an essential part in the survival and reproduction of animal species, but also holds significant research value in various fields, such as optical communications, healthcare, environmental monitoring, missile plume detection, combustion monitoring, and solar astronomy1,2. Due to the human eyes’ inability to sense UV radiation, UV photodetectors that enable direct detection of UV rays, have experienced rapid development3,4,5. For example, UV photodetectors (UVPDs) equipped on unmanned aerial vehicles can detect corona discharges in power transmission lines, providing early warnings of impending catastrophic electrical faults (Fig. 1a)6. However, the storage and processing of information carried by UV light requires specialized components, and the physical separation of the sensing and high-level processing functions undoubtedly increases the hardware overhead (Fig. 1b)7,8,9,10,11,12. Moreover, the processing of information generates a large amount of raw and unstructured data, leading to low frame rates and high energy cost8,13,14. Recent research has shown that machine vision systems based on in-sensor computing (ISC) architectures, which combine optical sensing with non-volatile tunable positive and negative weights, can effectively address the above problems (Fig. 1c)7,15,16,17,18. However, visible-blind UVPDs for ISC architectures have not yet been explored.

Fig. 1: Application of UVPDs in real-time processing tasks and schematic diagram of the required computing architecture.
figure 1

a Unmanned aerial vehicles equipped with UVPDs can be deployed on demand to locate and characterize faults. b Traditional UV sensory computing architecture flows with separate sensing, memory, and computing units, as well as high-cost optical filters. c Schematic diagram of an ISC architecture with various functional components integrated. The device needs to have tunable photoresponsivity and matrix-vector multiplication capability.

Full size image

Existing visible-blind UVPDs are mainly based on wide bandgap semiconductors, such as GaN, SiC, and SrTiO3 (STO)19,20. Their intrinsic electrical properties bestow them with natural visible blindness characteristics, thus eliminating the need for the integration of high-cost optical filters and complex optical systems. Among these materials, transition metal oxides have a unique functional ionic defect, namely oxygen vacancies. The concentration of oxygen vacancies can be modulated using an electric field, endowing the materials with a wide range of work functions21,22. Through this oxygen stoichiometry engineering, resistive swiching23,24,25, interfacial magnetism26,27, catalytic activity28,29, and switchable photovoltaic effect21,30 can be manipulated in oxide-based devices. As a typical transition metal oxide, STO not only has a wide bandgap (~ 3.2 eV) but also exhibits a high oxygen ion mobility31. These properties make STO an attractive material for developing ISC-UVPDs. However, the high oxygen ion mobility in STO also makes oxygen vacancies more likely to diffuse back to their initial positions31,32, leading to poor non-volatile properties. Recent research shows that CeO2 can be used as an oxygen reservoir33 because it has good oxygen ion conduction and the ability to store and release oxygen ions34. Through interface engineering, we engineered the CeO2/STO interface to confine oxygen vacancies within the CeO2 layer, thereby extending their recovery migration path. This integration of oxygen ion transport and storage within a single device ensures non-volatility and makes the design and fabrication of ISC-UVPDs feasible.

In this study, an ISC-UVPD with excellent visible-blind characteristics and an ultrafast nanosecond response is demonstrated based on CeO2/STO heterostructures. By manipulating the interfacial barrier through oxygen stoichiometry, the photoresponsivity of the ISC-UVPD can be linearly adjusted in both positive and negative directions. To showcase the device’s integrated sensing, storage, and computing functions, a 3 × 3 convolution kernel is built to realize real-time edge detection and image stylization. Moreover, a 3 × 3 × 4 ISC-UVPD array is designed to accomplish a standard pattern recognition task. It is also confirmed that the ISC-UVPD is capable of achieving complex pattern recognition and target localization tasks. This work establishes the framework for constructing non-volatile ISC-UVPDs through interface engineering.

Results

Tunable non-volatile photoresponsivity

CeO2 films were epitaxially grown on (001) STO substrates using pulsed laser deposition (PLD). The high quality of the grown single crystalline CeO2 film was confirmed by X-ray diffraction patterns and atomic force microscopy (Supplementary Fig. 1). The interdigital electrode structure35 was selected for fabricating the ISC-UVPD (Fig. 2a). Here, STO was employed as the photosensitive layer and CeO2 as the oxygen storage layer. More details on device fabrication can be found in the Methods section. During the testing process, the interfacial barrier was adjusted using voltage pulses (VP) on the left electrode while grounding the right side, and then the short-circuit current (Isc) of the device was measured under UV irradiation (a wavelength of 375 nm, an intensity of 10 mW cm–2). The I-V characteristics of the UVPD were investigated under dark and light conditions, showing a low dark current and a substantial light-to-dark current ratio of ~ 227 at 0.5 V (Supplementary Fig. 2). The contribution of STO and CeO2 to the device properties was studied by fabricating the same electrode structure on different substrates. Only the STO substrate showed photoresponse to UV light, although CeO2 films of the same thickness were grown on other substrates (Supplementary Fig. 3). Under UV illumination, devices fabricated on a pure STO substrate displayed negative/positive Isc after being subjected to 50 positive/negative VP pulses (± 10 V, each with a pulse width of 0.5 s) (Supplementary Fig. 4a). But, the positive and negative Isc decreased rapidly, making this device much less useful. The retention characteristics of the device were significantly improved by the additional CeO2 through interface engineering, and the Isc remained unchanged for at least 1000 s (Supplementary Fig. 4b).

Fig. 2: Optoelectronic properties of the ISC UVPD with tunable photoresponsivity.
figure 2

a Schematic of the device with the interdigital electrode structure stimulated using 375 nm UV light (light intensity fixed at 10 mW cm−2). b Photocurrent map measured via sweeping I-V characteristic curves after each voltage pulse programming (50 pulses, each with a pulse width of 0.5 s) from − 30 V to 30 V with a 5 V step. The device was reset to its initial state before each programming operation. c Long-term stability of the tunable states, extracted from (b) at 0 V. d Measured photocurrent versus Psource after different voltage manipulation (50 pulses, each with a pulse width of 0.5 s), ranging from − 20 V to 20 V. e Ultrafast photoresponse measured at 0 V bias using a 355 nm pulsed laser and an oscilloscope. The device was regulated with 100 pulses of +/− 80 V programming voltage (each with a pulse width of 0.5 s). f Photoresponse to light of different wavelengths (650, 532, 450, and 375 nm). The pink region represents the device’s photoresponse following modulation with + 10 V programming voltage pulses, whereas the blue region represents the photoresponse following modulation with − 10 V programming voltage pulses. In all cases, the light intensity was fixed at 10 mW cm-2. g Schematic circuit diagram for a 2 × 2 device array. h Time-resolved output photocurrent measured during the application of different illumination levels to the array.

Full size image

To examine the tunable photoresponsivity of the device, the I-V curve of the ISC-UVPD was measured after the application of 50 programming VP pulses with varying amplitudes (ranging from − 30 V to 30 V with a step of 5 V, each with a pulse width of 0.5 s) at a constant light intensity of 10 mW cm−2 (Fig. 2b). The device was modulated back to its initial state before each voltage programming operation. The detailed procedures for device programming and testing are shown in Supplementary Fig. 5. The results indicate that the device has a negative Isc and a positive open-circuit voltage (Voc) after positive programming and exactly the opposite behavior after negative programming. To obtain a more intuitive relationship between the photocurrent and modulation voltage, the variation of Isc with Vp was extracted and plotted (Supplementary Fig. 6). It is obvious that the photocurrent decreases monotonically as the modulation voltage increases from negative to positive. The photocurrents of the device also increase linearly with both the number and duration of programming voltages (Supplementary Fig. 7). Furthermore, each programmed state of Isc exhibits good retention characteristics (Fig. 2c). The initial increase in Isc during the illumination stage can be attributed to the photoinduced electronic transitions, that spatially redistribute the ionization states of oxygen vacancies (+ 2, + 1, and neutral states) in the STO substrate32. A large photoresponse range can also be achieved by applying a smaller programming voltage; however, this requires a longer programming time compared to using a larger programming voltage (Supplementary Fig. 8). The endurance characteristics of the device were evaluated with ± 10 V voltage pulse programming (Supplementary Fig. 9), and it was found that after 100 cycles, attenuation of both negative and positive photocurrents was negligible. The device also exhibits good photoswitching behavior (Supplementary Fig. 10). The measured Isc is plotted against illumination intensity (Psource) at various VP modulations (Fig. 2d). A linear fit with an average coefficient of approximately 0.99 illustrates the linear dependence of the data. The ISC-UVPD showed a fast photoresponse time (rise time) of about 10 ns (Fig. 2e), which is crucial for ultrafast image recognition. More details on ultrafast photoresponse measurement can be found in the Methods section. In addition, scaling the device size does not affect its ultrafast photoresponse and multi-state tunability (Supplementary Fig. 11).

To investigate the UV selectivity of the photodetector, its photoresponse to different wavelengths (650, 532, 450 nm and 375 nm) at the same light intensity (10 mW cm−2) was measured after ± 10 V VP programming (50 pulses of 0.5 s) (Fig. 2f). Apparently, the device only exhibits a significant photoresponse to UV light, with a calculated UV/visible rejection ratio reaching the impressive value of 1814, calculated as Isc(375 nm)/Isc(650 nm)36. The excellent visible-blind properties of the device can be attributed to its high transmittance to visible light (Supplementary Fig. 12a). The optical bandgap was determined from the intercept through the extrapolation of the linear portion of (αhυ)2 vs. hυ curves37,38, which gave a calculated value of 3.27 eV (Supplementary Fig. 12b). The spectral response measured under constant illumination from 300 nm to 800 nm at 0 V bias also confirms this conclusion. (Supplementary Fig. 13). More importantly, the device directly responded to the UV components of sunlight, indicating its potential for use in real-world scenarios (Supplementary Fig. 14, the UV index on the test day was 4). In this experiment, the device was modulated to produce more pronounced negative/positive photocurrents by applying 50 VP pulses (± 30 V, each with a pulse width of 0.5 s). Our visible-blind ISC-UVPDs exhibit good UV detection performance under 0 V reading voltage, including responsivities of ± 19 mA/W and a high detectivity of 2.4 × 1010 Jones under 375 nm light irradiation. The linear dynamic range is about 120 dB (ranging from 5 μW cm−2 to 5.7 W cm−2). Compared to the works in Supplementary Table 1, the present study shows good visible-blind characteristics with ultrafast photoresponse. These characteristics enable the device to focus on UV signals in complex lighting environments without the need for additional filters, which is of great significance for reducing hardware overhead and improving the performance, accuracy, and reliability of the system.

Then, the multiply-accumulate operation function of the ISC-UVPD was validated by connecting the array devices in parallel, as shown in the schematic of Fig. 2g. Considering that the output photocurrent Isc is directly proportional to the incident light intensity Psource, stated as Isc = R·Psource, the photoresponsivity R can be considered as the weight value15. The Isc and R corresponding to the values assigned to the device array are shown in Supplementary Fig. 15. The actual output photocurrent of the device array is the accumulation of the element-wise multiplication result of the illumination matrix and devices’ photoresponsivity matrix. Thus, irradiation with the same light intensity (10 mW cm−2) under different illumination patterns should result in different output photocurrents. The experimental results were consistent with the expected values (Fig. 2h), indicating that the ISC-UVPD array could perform multiply-accumulate operations. The negative and positive photoresponsivity can be increased by reducing the width of the channel (Supplementary Fig. 16).

Mechanism of tunable non-volatile photoresponsivity

Next, the tuning mechanism of the device’s non-volatile photoresponsivity was investigated. Kelvin probe force microscopy (KPFM) was employed to characterize the dynamic changes of the energy barrier before and after voltage programming39. Under dark conditions, both the positive (left side) and the negative (right side) terminals exhibited almost symmetrical potential distribution (Supplementary Fig. 17). Since KPFM is a surface probing technology, the results suggest that the Schottky barrier at the Pt/CeO2 interface remains stable before and after voltage programming, indicating that this interface does not contribute to the modulation. The measurement results under light irradiation show asymmetrical potential distributions, and these distributions were opposite for positive and negative voltage programming (Supplementary Fig. 18). Therefore, the charge accumulation of the left and right electrodes under light irradiation can be attributed to the change of the CeO2/STO interface. Supplementary Fig. 19 shows the dependence of Isc on poling time, implying the role of oxygen vacancy migration40.

To verify this hypothesis, scanning transmission electron microscopy (STEM) was utilized to observe the changes in the heterointerface beneath the positive terminal before and after voltage programming. The high-angle annular dark field (HAADF) image of the as-grown CeO2/STO interface is shown in Fig. 3a. The result indicates that CeO2 has high crystalline quality and an atomically sharp interface. It can be seen that the CeO2 film was rotated by 45° relative to the STO substrate to achieve minimal lattice mismatch (~ 2%). This leads to an orientation relationship of CeO2(001) | | STO(001) (interface normal) and CeO2[110] | | STO[100] (in-plane directions)41.

Fig. 3: Oxygen stoichiometry control before and after voltage programming.
figure 3

a STEM image of the pristine CeO2/STO interface. STEM-EELS scans of (b, d, and f) Ce M-edge and (c, e, and g) Ti L-edge: pristine state (b, c); negative regulation (1000 pulses of − 100 V voltage programming, each with a pulse width of 0.5 s) (d, e); and positive regulation (1,000 pulses of + 100 V voltage programming, each with a pulse width of 0.5 s) (f, g), across the interface area, as hinted by the orange frame in (a). Left: STEM-EELS 2D scan. Middle: the corresponding layer-by-layer EELS analysis. Right: M5/M4 and \(\frac{I({e}_{g}({L}_{3}))-{I}_{0}}{I({t}_{2g}({L}_{3}))-{I}_{0}}\) ratio distribution maps for each layer. h Schematic energy band diagrams and electron-hole pair generation after different voltage programming.

Full size image

With electron energy loss spectroscopy (EELS), a thorough investigation of the heterointerface was carried out with the goal of observing the changes in valence states of various elements before and after voltage programming. Six rows of atoms above and below the interface were selected for analysis, as shown by the orange frame in Fig. 3a. Figure 3b–g shows the corresponding horizontally integrated EELS data cube for the Ce-M edge and the Ti-L edge. The vertical axis is the scanning distance across the interfaces (in nm), and the horizontal axis is the energy loss (in eV). It can be seen that the valence distribution of the elements on the heterointerface underwent a transformation after voltage programming. For layer-by-layer comparison, EELS spectra were extracted from each atomic plane. In the initial state (Fig. 3b), the Ce-M edge exhibited a shift toward the lower energy direction with an offset of one atomic layer thickness at the interface, indicating that the pristine CeO2 film has a precise stoichiometry. Based on the decrease in t2g peak intensity at the Ti-L edge near the interface, it is evident that there are a large number of oxygen vacancies in the STO substrate (Fig. 3c). After negative programming (1000 voltage pulses of − 100 V, each with a pulse width of 0.5 s), the peak position of the multilayer Ce decreased significantly (Fig. 3d). Notably, larger peak shift amplitudes occurred closer to the interface, where the maximum energy shift reached 1.5 eV. This result indicates that the negative voltage drives the oxygen vacancies from the STO to the CeO2. In the positive programming sample, however, a completely opposite direction of migration was observed, since all the Ce in the spectra had a + 4 valence state (Fig. 3f). The above conclusion is supported by the variation of t2g peak intensity in the Ti L-edge under different modulation states (Fig. 3c, e, and g). To obtain more quantitative comparison results, the intensity ratios of the Ce M5/M4 peaks42 and the Ti eg(L3)/t2g(L3) peaks were extracted. It is worth noting that to obtain more accurate calculation results, the peak intensity ratio of Ti was defined as \(\frac{I({e}_{g}({L}_{3}))-{I}_{0}}{I({t}_{2g}({L}_{3}))-{I}_{0}}\), where I0 is the valley point between the eg(L3) and t2g(L3) peaks for each spectrum43,44. The increase in peak intensity ratio represents the decrease in the valence state of Ce and Ti, proving that the movement of oxygen vacancies occurred after applying the voltage.

We further examined the EELS spectra of the O-K edge at the interface to validate the above conclusion (Supplementary Fig. 20). It is evident that there are three featured peaks in both the CeO2 and the STO layers (denoted as A, B, C for CeO2, and a, b, c for STO), but in regions with a higher oxygen vacancies content, the featured peaks cannot be identified unambiguously45,46. The distinguishability of the O K-edge spectra varies in different states, which also confirms the migration of oxygen vacancies at the CeO2/STO interface. Moreover, EELS measurements were conducted on the Pt/CeO2 interface (Supplementary Fig. 21) where the Ce remains in the + 4 state, suggesting that this interface has very little effect on the modulation process.

We next analyzed the band structure changes modulated by oxygen vacancies. The electron affinities and bandgap widths of CeO2 and STO in their intrinsic states are shown in Supplementary Fig. 2247. In the absence of any doping, their Fermi levels are located in the middle of the bandgap. As the band filling is controlled by the number of oxygen vacancies, the Fermi level of the STO in the initial state is close to the bottom of the conduction band due to the presence of a large number of oxygen vacancies. When the two materials come into contact, the band bends at the interface and forms a n-n+ junction (Supplementary Fig. 23a). When a positive voltage is applied, most of the oxygen vacancies will be pushed from the interface to deeper layers of the STO substrate. Therefore, the Fermi level of STO is closer to its intrinsic position, which leads to a reduction in the barrier height of the n-n+ junction (Supplementary Fig. 23b). Under negative voltage modulation, positively charged oxygen vacancies are attracted and mostly transferred from STO to CeO2 (Supplementary Fig. 23c), forming an n+-n junction. Under the influence of a horizontal electric field, oxygen vacancies also undergo lateral migration in the STO (Supplementary Fig. 24). The non-uniform distribution of oxygen vacancies under the electric field breaks the symmetric potential barrier of the device and promotes the separation of photo-generated electron-hole pairs under illumination, resulting in large positive and negative Isc (Fig. 3h). In addition, since the oxygen vacancies are trapped in the CeO2 film, the modulation effect is retained when the voltage is removed.

Edge detection and image stylization

With the ability to perform multiply-accumulate operations, a photodetector array with 3 × 3 pixels was fabricated to implement real-time image processing functions. During the device fabrication, we designed our devices to be independent to avoid complications caused by crossed electrode wires. For array testing, a breadboard was used to connect the device electrodes, thereby a practical test of the device array could be performed. The optical image of this fabricated device array and the schematic of the corresponding circuit connections are shown in Fig. 4a. First, the array was programmed to act as a convolution kernel to demonstrate the vertical edge detection of a moving light spot. As shown in Fig. 4b, the weights of the left and right columns were set to − 1 and + 1, respectively, which was achieved by applying 50 voltage pulses of + 20 V and − 20 V (each with a pulse width of 0.5 s). The light spot scanned the convolution kernel along the x-axis at a constant speed controlled by an electric translation stage. Based on the characteristics of the device, negative and positive Isc should be observed when the convolution kernel is at the left and right edges of the light spot, respectively, while the photocurrent should be negligible when the light spot covers the entire convolution array. The experimental results agreed well with this hypothesis (Fig. 4c). The edge detection ability of the convolution kernel was also evaluated at different motor frequencies. Even when a frequency of 45 kHz was reached (corresponding to a movement speed of 2.8 cm s−1), clear edge detection could still be achieved thanks to the ultrafast optical response of the device. In addition, edge detection was also performed in the visible band (Supplementary Fig. 25), but no significant photocurrent output from the array was observed, indicating its visible blindness. To demonstrate a practical application, the device array was then exposed to sunlight and blocked with a movable shutter (Supplementary Fig. 26). The results show that the device array can also work well in natural light.

Fig. 4: A 3 × 3 ISC-UVPD array for edge detection and image stylization.
figure 4

a Optical image of the device array and schematic circuit diagram for the network. b Schematic of the light spot movement direction and the weight distribution of the device array. The UV light intensity was 10 mW cm−2. The weights of the left and right columns of the device array were assigned via electrical pulses of ± 20 V (50 pulses, each with a pulse width of 0.5 s), respectively. c Measured output photocurrent of the convolution kernel with the light spot moving across the whole device array at different speeds. d Original gray-scale image and schematic illustration of image stylization operations. e Various filtered images obtained through ISC.

Full size image

To demonstrate the versatility and scalability of the device in processing larger and more detailed images, the above investigation was extended to implement ISC convolution operations on a 256 × 256 pixel image. The original gray-scale image (8 bits, 0–255) and the schematic illustration of the convolution processing flow are shown in Fig. 4d. The 3 × 3 kernel slid over the input image with a step of 1 pixel, dividing the original image into multiple sub-images. Through the optoelectronic convolution, these sub-images were converted into photocurrents, which were then integrated and remapped, finally forming the output image. By adjusting the weights of each pixel separately, different types of convolution kernels can be realized. The convolution kernels are configured using electrical pulse programming, and the photoresponses of the array are then extracted experimentally and processed in simulation to achieve image stylization. Eight kernels with different image processing functions were implemented using the device array, and the detailed operation of image stylization can be found in Supplementary Note 1. The weight assignment method and R configurations are shown in Supplementary Fig. 27. Different image processing operations resulted in different image features (Fig. 4e). The high consistency between the key image features extracted during the experiment and the simulation results (Supplementary Fig. 28) demonstrates that the proposed device has outstanding performance as a building block for a convolutional neural network.

Pattern recognition and target ___location

By horizontally expanding the device array, we demonstrate the parallel computing capability of the system. As schematically shown in Fig. 1c, the ISC-UVPD array consists of N pixels with each pixel divided into M subpixels, where N depends on the image size. The subpixels with the same index (m = 1, 2, …, M) in each pixel were connected in parallel to create a device network with a spatially staggered arrangement. Each subpixel corresponds to a UV photodetector, which is operated under short-circuit conditions and, under light irradiation delivers a photocurrent of Imn = Rmn·Pn. Rmn is the photoresponsivity of the device at the nth pixel and mth subpixel, and Pn is the optical power at the nth pixel. The matrix-vector multiplication operation can be performed by summing all the photocurrents from the mth detector element of each pixel, resulting in the total output photocurrent Im. Here, a 3 × 3 pixel image sensor array was constructed in actual, where each pixel contained 4 subpixels (4 detectors per pixel), as shown in Fig. 5a. Since the photoresponsivity of each subpixel can be manipulated individually, the projection of an image of 3 × 3 pixels onto the array resulted in the generation of four different output photocurrents. This enables the device array to operate as a single-layer artificial neural network (ANN) with 9 input ports and 4 output ports, efficiently extracting multiple image features in a single projection operation. In the array, the negative electrodes of each row of devices are connected together and used as ground, while the positive electrodes are left unconnected. For subsequent classifier construction, the corresponding electrodes are routed and connected on a circuit board. This method facilitates the reconfiguration of the device array as required. The similarity of the I–V characteristics and the well-separated tunable photoresponsivity states of 36 devices under UV illumination clearly demonstrate the high device-to-device uniformity of our ISC-UVPD array (Supplementary Fig. 29). The uniformity of the device array can be further improved by enhancing film quality through the precise control of CeO2 growth.

Fig. 5: Pattern recognition and target ___location.
figure 5

a Optical image of the 6 × 6 array. The array was divided into 3 × 3 pixels, with each pixel containing 4 subpixels. The array acted as a single-layer perceptron, where the output currents I1, I2, I3, and I4, were the total sum of the photocurrents generated by the corresponding subpixels under the respective incident light intensities. b Schematic illustration of the optical setup and circuit diagram of the custom printed circuit board for pattern recognition tasks. When different letters were projected onto the device array through masks, the corresponding output LEDs lit up to display the recognition results. c Schematic diagram of input letter images ‘D’, ‘L’, ‘K’, and ‘J’ with Gaussian noise standard deviations varying from 0.1 to 1. d Noise-level-dependent recognition accuracy histogram. e Target ___location and discharge mode classification of corona discharge in power transmission lines. f Partial results of the target ___location. The red dots represent the experimental output, and the black line represents the ideal ___location results. g Classification accuracy of different discharge modes.

Full size image

The device was first operated as a classifier capable of recognizing the letters ‘D’, ‘L’, ‘K’, and ‘J’. To conduct a practical evaluation of the device’s pattern classification capabilities, a custom-printed circuit board was fabricated to operate the device network. The patterns were projected onto the device array through masks so that only the ISC-UVPDs under the transparent areas of the masks would generate output photocurrent (Fig. 5b). An offline learning approach based on a backpropagation algorithm was adopted to obtain array weights. During the training process, the output photocurrent of the target pattern was set to be the maximum positive value compared to those for the other patterns, enabling the identification of the target image. Then, the corresponding weights were written into our array for pattern recognition. When the light corresponding to the respective pattern irradiated the array device, the photocurrent output from each channel was processed by transimpedance amplifiers (TIAs). The classification results of the patterns are indicated by the brightness or darkness of the LED lights connected to each output terminal. The experimental results indicate that our ISC-UVPD array is well capable of performing pattern classification tasks (Supplementary Fig. 30 and Supplementary Movie 1).

We further prove their ability to recognize complex patterns. The device weight distribution was calculated using offline learning, allowing the device to recognize patterns with varying degrees of introduced Gaussian noise (standard deviations varying from 0.1 to 1 with a step of 0.1). The dataset used for training and testing is shown in Supplementary Fig. 31. The theoretical distribution of the devices’ photoresponsivity and the actual amplitude measurements are shown in Supplementary Fig. 32, and the detailed description of the constructed neural network can be found in Supplementary Note 2. Then, the photocurrent of each device was tested at different light intensities and summed through subsequent processing. The results suggest that the device array can effectively perform complex image classification tasks until the input image becomes unrecognizable (Fig. 5c and Supplementary Fig. 33). When the standard deviation of the introduced Gaussian noise was less than 0.4, the device array achieved nearly 100% recognition accuracy (Fig. 5d).

In power transmission systems, aging of insulators can lead to localized electric field concentration, resulting in intense discharge events such as corona discharges, flashovers, or arcing48. Therefore, timely inspection of insulators is crucial for the safe operation of power transmission systems. Optical imaging is a non-contact detection technology that can locate discharges more intuitively and accurately compared to traditional ultrasonic detection technology49. However, no visible light is generated in the initial stage of insulator discharge, so visible imaging is ineffective. In addition, since the detection results of infrared imaging methods are greatly affected by weather, these methods are unreliable in field conditions. When insulators discharge, UV rays are emitted at an intensity that increases with that of the discharge. Therefore, UV imaging offers advantages in detecting insulator discharge compared to other imaging technologies.

The target ___location and discharge mode classification functions for insulator discharge were demonstrated based on the scalability of our proposed device array and its ISC characteristics (Fig. 5e). To create an insulator discharge database as well as the labels of ___location and discharge mode, 600 normal insulator images from the public Chinese Power Line Insulator Dataset (CPLID) were manually modified to represent the two types of discharge patterns that can be observed with ISC-UVPDs. The resulting insulator discharge dataset thus contained 1200 images. No effective information about the insulator discharge could be seen in the visible image, but the feature information of the discharge was highlighted in the UV image so that the proposed ISC-UVPD array could process discharge information in real-time (Supplementary Fig. 34). A device array of the same size as the dataset images was simulated, with each pixel containing 4 subpixels. Two of the subpixels were used for target ___location and the other two for discharge pattern classification. For both tasks, the sigmoid function was used as the activation function of the output layer, and the mean square error function was used as the loss function. The detailed simulation methodology can be found in Supplementary Note 3. After 2000 training epochs, the target position was realized (Fig. 5f, Supplementary Fig. 35, and Supplementary Movie 2, 3). In addition, as training progressed, the classification accuracy of discharge patterns exceeded 90% after 42 epochs and reached 100% after 211 epochs (Fig. 5g). These results show that the proposed device array is capable of real-time ___location and classification of insulator discharge (Supplementary Movie 4). The UV radiation intensity of corona discharge ranges from μW cm−2 to mW cm−250,51, and our device can detect weak UV light intensities as low as 5 μW cm−2, maintaining controllable positive and negative photoresponses (Supplementary Fig. 36). It is believed that this ISC-UVPD is capable of capturing discharge phenomena in the real world.

Discussion

In summary, a reconfigurable ISC-UVPD with high-performance UV sensing, weight memorization, and parallel computing capabilities was constructed based on CeO2/STO heterostructures. Continuous linear control of the photoresponsivity in both positive and negative directions was achieved via controlling oxygen stoichiometry. Through interface engineering, the diffusion of oxygen vacancies in STO was limited, thereby improving the nonvolatility of the device. The device arrays are capable of performing edge extraction and image stylization functions using various kernels. Furthermore, the expanded device network can be used to form an ANN, enabling high accuracy in pattern classification and target ___location tasks. It is believed that the large-scale integration of the proposed ISC-UVPDs is possible based on the fabrication of the inch-scale STO films52. This work could lay the foundation for practical ISC-UVPD device applications.

Methods

Sample preparation

CeO2 thin films with a thickness of ~ 28 nm were epitaxially grown on (001)-oriented STO single crystalline substrates (HF-Kejing) by pulsed laser deposition using a 308 nm XeCl excimer laser, with an energy density of about 1 J cm−2 and a repletion rate of 3 Hz. The CeO2 films were deposited at 750 °C in a flowing oxygen atmosphere with a pressure of 10 Pa and then cooled down to room temperature at 20 °C min−1. The deposition rate was calibrated using X-ray reflection.

Device fabrication

We spin-coated photoresist (AR-P 3540) onto the sample surface and used standard photolithography technique to define the interdigital device structure. Subsequently, 50 nm Pt electrodes were deposited via radio frequency sputtering. The total sensing area of the device was 1 × 10−3 cm2, with a channel length/width ratio of 680/5 µm.

Material characterization

X-ray diffraction patterns of the CeO2 film were obtained using a Rigaku SmartLab instrument with a 2θ range from 20 to 80o in steps of 0.05°. Optical transmittance spectra were taken in air at room temperature with spectrophotometers (Cary 5000 UV-Vis-NIR, Agilent and Excalibur 3100, Varian).

Electron microscopy

STEM imaging was performed using a Cs-corrected JEOL JEM-ARM200CF NEOARM operated at 200 kV with a CEOS Cs corrector (CEOS GmbH, Heidelberg, Germany). HAADF-STEM images were recorded with collection semi-angles of 90-370 mrad. The EELS data were collected in dual-EELS mode to obtain both zero-loss and core-loss spectra and recorded with a Gatan spectrometer, applying an energy dispersion of 0.25 eV per channel for Ce-M, Ti-L, and O-K edges with a convergence semi-angle of 24 mrad. Core-loss EELS spectra were calibrated via the corresponding zero-loss EELS before further analysis using the Gatan Microscopy Suite Software.

Device characterization

All the photoresponsivity characterizations were performed using a Lakeshore probe station with a Keithley 4200 semiconductor parameter analyzer at room temperature. A UV laser with a wavelength of 375 nm was used for experimental device switching. The ultrafast photoresponse time measurement was tested at zero bias via a 1 GHz bandwidth oscilloscope and a 355 nm UV laser with a pulse duration of 20 ps. The spectral response of the device was measured using the DSR 300 Micro-nano Device Spectral Response Measurement System, and the EQ-99X-QZ-S, ENERGETIQ, and HAMAMATSU were used as light sources.

Creation of insulator discharge database

The 600 normal insulator images and their corresponding position information can be obtained from the public Chinese Power Line Insulator Dataset (CPLID), available at https://github.com/InsulatorData/InsulatorDataSet.git. Subsequently, the two types of discharge patterns obtained by the UV photodetectors were synthesized with the original images, and their positions were aligned with the center position of the largest insulator in each image. In this way, 1200 UV images containing insulator discharge information and corresponding composite images could be obtained, and each image contains 1152 × 864 pixels.

Simulation of target ___location and discharge mode classification

An 1152 × 864 imaging array with 4 UVPDs per pixel was constructed to simulate a single-layer ANN with an input layer (995328) and an output layer (4). Since the number of nodes in the input layer is much larger than the number of nodes in the output layer, the input value was scaled from 0–255 to 0–0.05. As the activation function for the output layer, we chose the sigmoid function \({y}_{i}={\left(1+{e}^{-{x}_{i}}\right)}^{-1}\), where yi is the output value of the ith output node, and xi is the received value of the ith output node. The mean-square loss function \(L={\left[{\sum }_{i=1}^{4}{\left({y}_{i}-{{label}}_{i}\right)}^{2}/4\right]}^{1/2}\) was used as the loss function, where L represents the difference between the output value and the label value as well as labeli is the label value corresponding to the ith output node. The weights were updated after every epoch by the backpropagation function of the gradient of the loss function.