Simulating the human visual system to achieve efficient image information perception is a research hotspot in the field of artificial intelligence. However, constrained by the Von Neumann architecture, which separates memory and processors, traditional machine vision systems are facing issues such as high energy consumption, speed delay, and data redundancy due to the lack of neuromorphic computing capabilities and the separation of sense-compute hardware structures^[1−3]. To address this challenge, one can develop artificial visual recognition systems by imitating the advanced processing mechanisms of biological visual systems when dealing with visual information, achieving high-speed, efficient, and low-power data processing. The biological visual system serves as a natural model for artificial vision technology, utilizing the hierarchical structure of the retina and the brain's visual cortex to achieve visual perception. The retina is responsible for capturing light and converting it into neural signals, which are then transmitted to the brain for further processing. In this process, horizontal cells and amacrine cells in the retina play an important role, parallel processing visual information, helping to identify key features such as color, shape, and motion in the image, thereby enabling rapid recognition and classification of targets^{[4, 5]}. The visual cortex can extract and analyze high-dimensional feature information, enabling the system to achieve high-precision recognition of environmental targets^[6]. Therefore, by imitating the biological visual system and implementing hierarchical and coordinated processing strategies, it helps to build more intelligent and efficient computational models to cope with various machine vision recognition and scene understanding tasks.

Currently, simulating the human visual system to achieve high-energy-efficient image information perception has made unprecedented progress^[7]. Among them, tunable photoresistors show great advantages in simulating visual synapses. They have both neuromorphic computing functions and optical control characteristics, which can be used to achieve the integration of image perception, storage, and processing functions in a new type of neuromorphic visual system. Under the joint action of electrical and optical signals, they simulate the plasticity of biological synapses^{[8, 9]}. Additionally, visual synapse devices have the advantages of efficient processing, strong adaptability, and swarm intelligence, making them one of the promising candidates for future bio-inspired artificial vision. However, current visual synapse devices have relatively limited functions and lack the ability to process complex information, urgently requiring the development of more innovative synaptic components. These components will be based on new materials and innovative structural designs to achieve a more diverse range of synaptic functions.

Organic materials, due to their photosensitivity, tunability, flexibility, and ease of processing, show great potential in the field of optoelectronic memristors to simulate the behavior of biological synapses, especially in the application prospects of neuromorphic computing^[10]. Additionally, organic photovoltaic materials also show great advantages in the field of solar cells, such as the star molecule PM6 : Y6 in the field of traditional solar cells, which is mainly due to the excellent compatibility between PM6 polymer and Y6 small molecule materials, which helps to reduce internal defects of materials and improve charge transfer efficiency^{[11, 12]}. At the same time, the energy level matching between PM6 as donor and Y6 as acceptor also promotes the effective generation and separation of photogenerated carriers, which has charge transfer potential and memristor possibility. Therefore, the development of tunable photoresistors using organic photovoltaic materials not only promises to simulate visual synaptic behavior but also has the potential to open up more advanced application fields.

This article proposes a vertically structured photo-tunable organic memristor based on the PM6 : Y6 system, which achieves stable and smoothly varying resistance behavior, and exhibits positive enhancement and negative suppression characteristics, realizing voltage-modulated LTP/LTD, and analyzes its working mechanism as well as the charge transfer behavior induced by photo-kinetics. Additionally, the device simulates the synaptic plasticity co-regulated by photo and electricity, successfully achieving reversible control of 22 consecutive photoelectric conduction states. Finally, the synaptic plasticity of the device was modeled and a hierarchical coordinated processing SLP-CNN cascade neural network was constructed to simulate the efficient visual processing mechanism of biological systems for the recognition training of handwritten digit characters. Compared to a single-layer perceptron recognition network, the hierarchical coordinated recognition system achieved a recognition accuracy as high as 93.4%.

The synthesis route and characterization details of the materials are described in the supporting information. ¹HNMR hydrogen spectra were measured by a Bruker 400 MHz/AVANCE Ⅲ 400 spectrometer by taking 2−3 mg of the sample dissolved in deuterium reagent and making the height of the solution in the NMR tube higher than 4 cm, and using tetramethylsilane as the internal standard for the test. The relative molecular mass of the polymers was determined by gel permeation chromatography (GPC) using tetrahydrofuran as eluent and polystyrene as reference material. Ultraviolet−visible absorption spectra (UV−Vis) were measured by Shimadzu UV-2600 spectrophotometer (Shimadzu, Japan); steady-state fluorescence spectra (PL) were measured by a HORIBA JOBIN YVON Fluoromax-4 spectrophotometer (HORIBA Scientific, France) (solvents were chloroform); atomic force microscopy (AFM) images were measured by Dimension Icon & FastScan Bio, Nanonavi E-Sweep instrument (SII); device I−V curves were measured using a Keithley 4200A semiconductor tester; the light pulse test is performed using an external light source, where the maximum light intensity of the light pulse is 255 levels, which corresponds to 53.5 w·m⁻²; the I−t curves were all read by the READ program in the 4200A.

The structure of the two-terminal resistive variable memory device based on PM6 : Y6 = 1 : 1 behaves as a Au/PM6 : Y6/ITO sandwich structure. The process of device preparation is as follows: 1 cm × 1 cm of ITO/glass substrate is selected, and ultrasonicated in ultrapure water, ethanol, acetone, and isopropanol sequentially for 15 min, then blown dry with ultrapure N₂, and activated by ozone plasma treatment for 5 min. The PM6 : Y6 dissolved in chloroform (20 μL, 10 mg·mL⁻¹) was spin-coated on the pre-treated ITO surface at 3500 rpm for 30 s. In the AFM testing, the thin film exhibited good uniformity and smoothness, with an average roughness of 2.08 nm in the 5 μm × 5 μm scanning area (Fig. S2(a)), SEM showed that ITO and active layer were 195.31 and 84.82 nm, respectively (Fig. S2(b)). After spin-coating, the devices was placed in a vacuum oven at 80 °C overnight, until the solvents were completely removed. Then the devices were sputtered with metallic gold (Au) using a high-vacuum metrological sputtering system at 10⁻⁷ vacuum to obtain domed electrodes with a thickness of about 50 nm and covering a diameter of about 100 μm. All electrical tests were performed on a Keithley 4200A semiconductor parameter analyzer equationuipped with a pulse test module.

The specific structure diagram and test method of Au/PM6 : Y6/ITO device are shown in Fig. 1(a), in the process, a two-dimensional conjugated polymer PM6 was obtained through the Stille coupling polymerization reaction, as shown in Fig. 1(b). The device with three-layer structure has a typical nonvolatile resistive-variable characteristic, which exhibits forward-biased turn-on and reverse-biased turn-off. It has a turn-on voltage of 1.61 V and a turn-off voltage of −0.89 V, with a switching ratio of about 30 (Fig. 1(c)). As shown in Fig. 1(d), the turn-on voltage of the device is primarily distributed at 2.6 V, with a proportion of 35%, and a distribution range from 1.0 to 3.0 V. The turn-off voltage, on the other hand, is mainly concentrated at −1.6 V, with a narrower distribution range from −0.8 to −1.8 V. The Weibull plot (Fig. 1(e)) statistically analyzes the distribution of high and low resistance states, exhibits a narrow distribution for both states with the low resistance state around 60 Ω and the high resistance state around 1800 Ω.

The device’s high resistance state (HRS) and low resistance state (LRS) can be maintained stably within 4000 s at a read voltage of 0.1 V (Fig. 1(f)). Furthermore, under the stimulation of 5 × 10⁵ continuous pulses, the device also demonstrates excellent pulse stability, with resistance changes kept within a controllable range (Fig. 1(g)). The device also exhibits excellent cyclic stability of over 230 cycles during the repeated on-off cycles of 3→0.1→(−2)→0.1→3 V (Fig. 1(h)). This indicates that the device has good repeatability during the cycling process.

Resistive random-access memory (RRAM) regulates its internal conductance state through direct current voltage. In the D/A system, charge transfer (CT) between donor and acceptor molecules occurs through CT states, and under the influence of voltage, CT is achieved^[13], leading to a change in the resistance state, from the HRS to the LRS. During the linear scanning process of direct current voltage, the conduction mode of the device undergoes a transition from Ohmic conduction to the space charge limited (SCLC) model. This change occurs because when the active layer does not capture electrons or holes, the device exhibits the HRS with minimal current. As the capture process occurs, the capture of electrons or holes significantly enhances the conductivity of the device, leading to an exponential increase in current relative to voltage, which is consistent with the SCLC model^{[14, 15]}. When the capture reaches a saturated state, the conduction mode shifts to follow the Child's law, at which point the rate of current increase slows down. Eventually, when all the trapping sites are filled, the device once again enters the HRS, and the relationship between current and voltage returns to the linear increase of Ohm's law (as shown in Fig. 2(a)). This process reveals the internal mechanism of the device based on the PM6 : Y6 system, which involves the dynamic process of carrier capture and release.

In order to intuitively characterize the above phenomenon, we conducted a kelvin probe force microscopy (KPFM)^{[16, 17]} test. Compared to the surface potential map of the uncharged film (Fig. S3), the overall potential of the thin film surface increased after electrification, and the surface potential significantly decreased after a negative voltage was applied to the central area, forming a visually perceptible "depression", indicating an enhanced sensitivity of the device to holes (Fig. 2(b)). By measuring the cross-sectional potential maps at 0, 30, and 60 min after electrification (Fig. 2(c)), we found that the potential curve rose overall after electrification, and a concave area appeared, which is consistent with the surface potential morphology map. As time progresses, the surface potential of the thin film gradually decreases and tends to return to the initial state, reflecting the disappearance of captured holes and electrons recombining. The results of the KPFM test confirm the presence of a capture mechanism inside the device, and the main captured substance is holes.

Based on the SCLC model and the results of the KPFM test, we could explain the mechanism of turn-on and turn-off processes in device, as a model diagram shown in Figs. 2(d)−2(i). In the initial state the device is in the off state, the high Schottky barrier between the co-blended PM6 : Y6 layer and the Au electrode make it difficult for carriers to jump the barrier, so the device is in the HRS (Fig. 2(d)); when a forward bias voltage (+3 V) is applied to the device, a large number of holes are trapped by hole defect sites in the PM6 : Y6, resulting in a lower Schottky barrier between the material and the Au electrode, and a gradual transition of the device from the HRS to the LRS (Fig. 2(e)); when the capture is completed, the device internal the LRS, the device is in the open state (Fig. 2(f)); when the opposite voltage is applied, due to the Au electrode is in the negative pole, which leads to the original active layer captured holes and back to the Au electrode to move to capture the process occurs gradually (Fig. 2(g)), at this time, the Schottky barrier is gradually increased, which makes the device gradually by the LRS into the HRS (Fig. 2(h)); when all the holes to capture is completed, the device is back to the closed state (HRS) (Fig. 2(i)).

Based on the light response behavior above, we also explored the potential for integrating photodetection functionality within PM6 : Y6 memristors by studying the variation of the photovoltaic electric field through voltage stimulation. In the experiment, the device was applied by negative pulse with varying voltages (from −1.0 to −5.0 V, lasting for 30 s), and then tested the I−V curves within −0.5 to 0.5 V, all tests were conducted under the maximum intensity of red light. As shown in Fig. 3(a), with the increase in initial negative bias, the I−V scan curve shifts gradually towards the negative field direction, also the photocurrent increases, which indicates an enhancement of the photovoltaic electric field. While the bias has no effect on the photovoltaic electric field in the dark. Conversely, the application of positive pulse voltages (from 0.5 to 4.0 V, for 30 s) causes the I−V curve gradually revert to initial state, effectively reducing the photovoltaic electric field, while the I−V curve in the dark state remains unchanged (Fig. 3(b)). By modulating the bias voltage applied to the photovoltaic electric field, the photocurrent response of the memristor at a specific read voltage could be controlled.

The photodetection characteristics of the PM6 : Y6 memristor can be explained through the photo-induced changes in the energy levels of PM6 : Y6 (Fig. 3(c)). In the initial state, the PM6 : Y6 blend forms as a bulk heterojunction solar cell with Au and ITO electrodes. Under light exposure, the photogenerated excitons generated in PM6 separate into electrons and holes at the interface of PM6 and Y6, then flowing towards the ITO and Au electrodes respectively, to form a photovoltaic electric field. When a negative bias is applied, holes transfer from PM6 to the Au electrode, causing a reduction in the LUMO level of PM6 and decreasing the LUMO level difference with Y6. This facilitates the transfer of more LUMO electrons from PM6 to Y6 and promotes the dissociation of new excitons, thereby enhancing the photovoltaic electric field. While the positive bias raises the energy levels of PM6, reducing the photogenerated carriers, and decreasing the photovoltaic electric field, which is reflected in the I−V curve as the voltage at which the current minimum approaches 0 V.

To further explore the application potential of PM6 : Y6 memristor in the field of in-sensor computing and brain-like bionics, we studied the photoelectric modulation ability and synaptic plasticity of PM6 : Y6 devices. As shown in Fig. 4(a), the memristor exhibits wavelength-dependent optical properties under electrical excitation. Particularly, the response to blue light is the most significant among white, UV, red, green, and blue light, with a photocurrent response of up to 70.6 nA at a voltage of −0.1 V, showing a high on−off ratio of over 4000 times compared to the dark current of 17.2 pA. By testing with white light, as the light intensity increases/diminishes, the device's response current shows an increasing/decreasing trend (Fig. 4(b)). This trend approaches saturation after 30 s of pulse voltage stimulation (Fig. S4). Moreover, under red light pulses, as the bias voltage increases/decreases, a linear trend of photocurrent change with bias voltage was obtained, manifesting 11 rising or falling tunable photoconductivity states (Figs. 4(c) and 4(d)). This result is also universally applicable to both green and blue light (Fig. S5), preliminarily demonstrating the device's capability for linear modulation in response to visible light.

Synapses are the connections between neurons and are responsible for the transmission of information. Presynaptic neurons affect the potentials of postsynaptic neurons by releasing neurotransmitters, generating postsynaptic current (PSC), which in turn stimulates or inhibits the latter. Artificial visual processing systems mimic this mechanism and use parallel computation to process and store information, achieving synaptic plasticity. Therefore, by integrating the photonic modulation capabilities with the non-volatile resistive state changes, we have further emulated the biological synaptic behaviors of LTP and LTD^[18]. As shown in Fig. 4(e), under the stimulation of a red light pulse (pulse width: 5 s, pulse interval: 5 s) at a light intensity level of 255, a positive voltage pulse signal (0.1→5→0.1 V) is set. The initial current level of the device is approximately in the nanoampere range. The electrical stimulation from 0.1 to 5 V instantly increases the photocurrent by four orders of magnitude, indicating that the device enters LRS under the stimulation of positive voltage. After the bias stimulation is completed, the device current does not decrease with the stimulation of the light pulse, and it needs external electrical stimulation factors to recover, indicating that the device forms a non-volatile conductivity state after the stimulation. The non-volatile photocurrent change simulates the electrical signal characteristics of LTP. Under the same stimulation conditions, after the input of a negative voltage stimulus from 0.1 to −5 V, the device transfer back to HRS, exhibiting the characteristics of LTD (Fig. 4(f)). This demonstrates that the device has characteristics of long-term memory (LTM), through repeated training, the connections between synapses are strengthened, thus forming LTM that can be sustained over time^[19].

Unlike traditional CMOS photodetectors, organic photoelectric memristors can adjust their photoresponsivity to achieve adaptive imaging under various lighting conditions. During neural network training and prediction, synaptic weights can also be updated through adjustable photoconductive state performance. Based on the excellent photoresponsivity capability of the Au/PM6 : Y6/ITO memristor, we designed a 5 × 5 device pixel array to recognize and read the pattern of the uppercase letter "H" by selecting light pulse signals of different intensities. As shown in Fig. 5(a), the dark current and photocurrent of the device are read at −0.1 V, with an initial average value of approximately 10.6 × 10⁻¹² A in the dark state. On the basis of electrical stimulation, when a beam of light with an intensity of 10.49 W/m² in the shape of an H is projected onto 11 devices of the memristor array, these devices can produce a blurry pattern of the uppercase letter "H" and exhibit a small photocurrent of 7.71 × 10⁻⁹ A. Due to the stimulation of different light pulse intensities, the devices exhibit photocurrents with varying degrees of growth, thus forming a clearer perceptual image. When the irradiation time is maintained at 3 min and the light intensity is increased from 10.49 to 53.5 W/m², the photocurrent of the device increases to 50.1 × 10⁻⁹ A. The photocurrent of the specific pattern "H" increases with the increase of light intensity, and the dark current of the background pixels gradually strengthens, forming a clearer image, simulating the adaptive image perception behavior in the human visual nervous system.

Based on the aforementioned adaptive image perception behavior, we constructed an SLP-CNN composite cascaded neural network (Fig. 5) to simulating the efficient visual processing mechanism of hierarchical coordination in biological systems. In this network, the SLP network simulates the function of the retina in the biological visual system, providing preliminary classification and recognition of visual targets by analyzing their contour features, while the CNN simulates the function of the visual cortex, achieving more refined recognition of the preliminarily classified targets by extracting and deeply analyzing the high-dimensional feature information of the visual targets. Utilizing the 22 photoconductive states of the PM6 : Y6 memristor to update the synaptic weights in a cascading network achieves the purpose of weight mapping to conductance. The characteristic of weight updates for enhancement and inhibition will be modeled through the following equation. The formulas are as the follows^[20]：

1GP=B(1−e(−P/A))+Gmin,

2GD=−B(1−e((P−Pmax)/A))+Gmax,

3B=(Gmax−Gmin)/(1−e−Pmax/A).

Here G_P and G_D represent the conductance values corresponding to enhancement and inhibition, respectively. G_max, G_min, and P_max represent the maximum, minimum conductance values, and the maximum pulse number, respectively. A is used to determine the nonlinear behavior of weight updates and the value of B is a function of A.

Taking handwritten digit character recognition as an example, an SLP-CNN cascaded neural network was built in the Python environment and trained using the MNIST dataset for recognition. The input size of each image is 28 × 28, and noise removal and pixel normalization were performed on the images before training using Python packages. As shown in Fig. 5(b), after the SLP neural network classifies unknown visual targets into completely closed class characters ("0, 6, 8, 9"), partially closed class characters ("2, 3, 4, 5"), and non-closed class characters ("1, 7") according to closed features, process the data using a 5-layer CNN network. Based on the detailed features of completely closed class characters, conduct fine recognition of the initially classified character target images according to the four categories of "0", "6", "8", and "9". As shown in the confusion matrix of Fig. 5(c), the classification and recognition accuracy of the 3-category SLP and 4-category CNN neural networks are 95.3% and 98.0%, respectively. Therefore, the overall recognition accuracy of the SLP-CNN cascaded visual recognition system is 93.4% (95.3% × 98.0%), which is 19.2% higher than the accuracy obtained by directly identifying unknown visual targets with SLP network. (as shown in Fig. 5(d) at 74.2%).

This article conducts an in-depth performance analysis of the Au/PM6 : Y6/ITO structured organic memristor by leveraging the outstanding photoresponsivity properties of the PM6 : Y6 system. The study found that the device not only has excellent retention and endurance but also exhibits stable and continuous resistance variation characteristics, with the properties of positive enhancement and negative suppression. These features enable the device to achieve voltage-controlled LTP and LTD functions. The research further explores its working mechanism, including the charge transfer process induced by photo-kinetics. Additionally, the device successfully simulates the photoelectric synergistic regulation of synaptic plasticity, achieving 22 consecutive reversible photoelectric conduction state regulations. Finally, by using the Python language, a cascaded neural network of SLP-CNN was constructed to simulate the hierarchical coordinated visual processing mechanism, and a recognition training simulation was performed on handwritten digital characters. When identifying the detailed information of visual targets, the overall accuracy rate was 93.4%. In comparison, the single-layer visual target recognition pattern implemented with the same organic memristors only showed a low recognition rate of 74.2%. Our research provides new ideas and methods for developing machine vision systems with integrated sensing and storage functions and adaptive adjustment and recognition systems, and promotes the development of intelligent optoelectronic device technology.