Introduction

The growing demand for artificial intelligence systems have exposed critical limitations in conventional computing paradigms1,2. While achieving unprecedented performance in complex cognitive tasks3,4, these systems grapple with unsustainable energy demands—modern GPU clusters consume megawatt-scale power during inference, exceeding biological neural efficiency by nine orders of magnitude5. Biological nervous systems excel in their energy-efficient operation, particularly the human brain, which orchestrates trillions of synaptic operations per second at power budgets of only 20 watts. This is achieved through highly parallelized, adaptive, and analog processing via neurons and synapses6,7. These biological mechanisms, including pulse coding, synaptic plasticity (Fig. 1a), and dynamic reconfiguration, serve as inspiration for neuromorphic computing, which aims to achieve similar efficiency through artificial synapses (Fig. 1b)8,9. Compared to conventional computing systems, this physical neuromorphic architecture can achieve biologically-inspired perception and learning functions in an efficient and low-energy manner10,11.

Fig. 1: Owl-inspired Neuromorphic Near-sensor Computing.
Fig. 1: Owl-inspired Neuromorphic Near-sensor Computing.
Full size image

a Owl optic neurobiological synaptic transmission. Spatiotemporal integration of spiking signals (postsynaptic currents, PSC) via neurotransmitter release and synaptic weight regulation. b Artificial neural system implementation. Near-sensor computing architecture (left) simulates biological signal perception and processing, and the artificial neural network (right) simulates synaptic plasticity through programmable weights. c Schematic of light adaptation thresholds. Comparison of light sensitivity ranges of complementary metal-oxide semiconductor (CMOS) photoreceptor devices (~1 lux, corresponding to a light intensity of ~10−7 W cm−2 @ 555 nm), human nocturnal vision (10−2 to 10−1 lux, corresponding to a light density of approximately 10−9–10−8 W cm−2 @ 555 nm), and barn owl nocturnal vision (~10−3 lux, corresponding to a light intensity of about 10−10 W cm−2 @ 555 nm). d Schematic of the adaptive mechanism of nighttime photoreception in owls. As time-dependent photosensitivity (Pt) increases with increasing light duration, dark adaptation of retinal optic rod cells leads to progressively clearer target imaging. e Schematic of autonomous unmanned aerial vehicle applications: integration of optical sensing, visual adaptation and neuromorphic computing enables target detection in photon-starved environments without the auxiliary lighting and post-processing techniques.

Among the various neuromorphic computing approaches, near-sensor computing, which integrates sensing and processing at the point of signal capture, has garnered significant attention12,13,14,15. Given that visual information accounts for more than 80% of human perceptual input, optoelectronic synapses—devices that combine signal sensing and processing—are central to enabling energy-efficient computation. Compared with the “sensing-processing separation” architecture of traditional machine vision (Supplementary Fig. 1), the near-sensor computing system shows an order of magnitude advantage in terms of energy efficiency (<1 pJ/operation) and delay16. Despite progress in this field, there remain substantial challenges, particularly in the context of low-light environments, where current technologies struggle to match the remarkable capabilities of biological systems17,18,19. Compared to living creatures (Fig. 1c), image sensors based on complementary metal-oxide semiconductor (CMOS) technology (which is now mature) can only perceive scenes with an illumination of about 1.0 lux20. This corresponds to a light density of approximately 10−7 to 10−8 W cm−2. Researchers have tried various post-processing techniques, such as microlenses, anti-reflective coatings, additional near-infrared sensing, noise filters, and image enhancement algorithms, to overcome ultra-low light environments. However, these techniques often have compatibility issues with mainstream CMOS technology and complex software systems20,21. Figure 1d elucidate the owl’s exceptional visual acuity in dim lighting from a biological standpoint. Its retina contains a greater density of rod cells compared to humans, and its photosensitivity (Pt) can be enhanced through the accumulation of photon input, a phenomenon referred to as dark adaptation of vision18,22. Upon completion of the adaptation, the imaging of objects in low illumination on the owl’s retina becomes more distinct23. Therefore, the development of an artificial synaptic device capable of mimicking the visual neurons of owls and being electronically controlled to participate in artificial neural network (ANN) computations paves the way for a monolithic, integrated near-sensing vision computing system. As shown in Fig. 1e, detection equipment equipped with this system, such as unmanned aerial vehicle (UAV), will be able to perform low-energy passive detection tasks in dark environments, such as night sky search and rescue, deep space exploration, and so on.

In this study, we present an owl-inspired dual-mode adaptive synapse (ODAS) that has the adaptive capability of the visual nerve. The device is designed as a three-terminal transistor structure with an insulating polymer poly-coating having an ester base used as a light-absorbing layer to reduce the density of charge-trapping states. The photogenerated carriers are transferred to the channel and accumulated using gate voltage control to dynamically enhance the photosensitivity of the device. ODAS has a high activity adaptive index of ~331, which demonstrates photon adaptive functionality for super-visual sensing in photon-starved environments (0.146 nW cm−2), an improvement of about three orders of magnitude compared to CMOS photosensors. In addition, the device has wide spectral sensing enabling pulse weight modulation from the near infrared to the ultraviolet range (365–940 nm). In the electrical pulse input mode, we achieved periodic stable synaptic weight modulation of long-term potentiation (LTP) with long-term depression (LTD) and verified the feasibility and generalization ability of deploying the weights in artificial neural networks at three different spatial scales. Using UAV, we developed an air-to-ground target detection neuromorphic computing system that mimics the owl’s visual nervous system, with over 95% simulations recognition accuracy on multi-dimensional datasets through adaptive improvement of the simulated root mean square contrast. Simulating neuromorphic computing plays a crucial role in demonstrating the superior efficiency of ODAS devices over conventional architectures for neural network tasks. More importantly, these simulations provide essential guidance for the development of next-generation devices, identifying key parameters that require optimization. This paradigm demonstrates the potential applications for which ODAS can be used in advanced optoelectronic neuromorphic devices, paving the way for hardware implementation of near-sensor computing systems in astronomical dark environments.

Results

Design and photoadaptation

The structural design of this optical synapse device is shown in Fig. 2a. The ODAS adopts a three-terminal transistor structure, with the gate (G) located at the bottom of the device. Conductive indium tin oxide (ITO) with high transmittance is selected to achieve light-unobstructed input. For the source and drain (S/D), silver (Ag) is initially deposited and subsequently modified with pentafluorobenzothiol (PFBT) to improve the work function. This approach allows a polyvinyl alcohol (PVA, Supplementary Fig. 2) film is used as the dielectric layer. Because PVA has the property of trapping electrons, the photogenerated electrons in the hybrid layer can be trapped on the PVA film under the driving of the gate electric field, thereby achieving non-volatile regulation of the current. The design of the light-absorbing layer is the key to ODAS. Lead sulfide nanocrystals (PbS NCs) were selected as the photosensitive material, which can effectively respond to light input in the infrared and visible light bands. In addition, an insulating polymer poly(vinyl-cinnamate) (PVCn, Supplementary Fig. 2) with an ester group was introduced to form a conjugated mixture with PbS NCs to passivate the charge capture state. This approach prevents the formation of carrier trap states within the transport channel or at the channel/dielectric interface. For the channel material, this results in an increase in the recombination a high-mobility 2,7-dioctyl[1]-benzothiopheno[3,2-b][1]benzothiophene (C8-BTBT, Supplementary Fig. 2) crystalline film as the channel material, ensuring high signal amplification efficiency of the transistor. Its light absorption characteristics in the ultraviolet band can extend the device’s light detection range. ODAS has an optical absorption layer, a charge trapping layer and multiple heterojunctions, which can maximize the retention of the intrinsic amplification characteristics of the transistor, thereby optimizing the grating effect.

Fig. 2: ODAS structure characterization and light adaptation performance.
Fig. 2: ODAS structure characterization and light adaptation performance.
Full size image

a Schematic diagram of the three-dimensional structure of ODAS: from bottom to top, ITO gate, PVA dielectric layer (120 nm), PVCn: PbS NC light absorption layer (100 nm), Ag source and drain electrode (100 nm), and C8-BTBT channel layer (50 nm). b Cross-section TEM image (scale bar: 25 nm) and corresponding EDS mapping images (scale bar: 25 nm). (c) Absorption spectra of PVCn: PbS NC blended films and C8-BTBT channel. d Dynamic process of light adaptation mechanism: light stimulation triggers the generation of electron-hole pairs in the PVCn: PbS layer (left); electron-hole pairs are separated under gate voltage, the holes are trapped in the interface trap state and accumulate at the channel interface (middle); after light removal, the carrier is recaptured and recombined to complete the photoadaptation cycle similar to the rod cell (right). e Drain current (IDS) with the increase of the adaptation time under different illuminances (0.146–11.7 nW cm−2), showing the adaptive ability to light intensity. f Relationship between the photosensitivity (Pt) and the illuminance under different adaptation times (1–30 s). g Correlation between active adaptation index (AAI) and light illuminance.

In order to demonstrate that the ODAS has the expected structure and properties, a number of characterization tools were employed to reveal the material properties. Figure 2b displays the transmission electron microscopy (TEM) image of the ODAS cross-section, accompanied by the corresponding energy dispersive system (EDS) spectra. The clean contact interface of the channel layer with the source/drain electrode is facilitated by the direct growth of C8-BTBT on the pre-deposited electrode. This approach circumvents the damage to the semiconductor/metal interface that is inherent to conventional evaporation and sputtering processes. The aforementioned method enables the C8-BTBT single crystal domain size to reach the millimeter scale, thereby facilitating coverage of the channel area of the device (Supplementary Fig. 3). Moreover, in-plane X-ray diffraction phi-scan measurement on the (020) plane of C8-BTBT observed two sharp peaks at an interval of 180°, further confirming the single-crystal domain properties of C8-BTBT (Supplementary Fig. 4). The surface of PbS NCs, which are a photosensitive material, has a high density of electron-well states. This results in an increase in the recombination probability of photogenerated carriers, which in turn affects the current response of the material under low-light conditions. The combination of PbS NCs with PVCn serves to passivate the intra-layer trapping states of carriers, thereby ensuring the generation of effective carriers under conditions of weak light. To verify the suppression effect of PVCn on PbS NC trap states, PbS NC films were prepared as a control group and compared with PVCn: PbS NC blend films. X-ray photoelectron spectroscopy (XPS) in Supplementary Fig. 5a demonstrates that the binding energy peaks of Pb 4 f in the PVCn: PbS blend film are shifted to lower energies, indicating that the introduction of PVCn results in a reduction of the binding energy of Pb. Furthermore, the absorption spectra in Supplementary Fig. 5b reveal significantly enhanced absorption for PVCn: PbS compared to pure PbS across key spectral bands, particularly in the near-infrared region. This absorption enhancement reflects improved photogenerated carrier efficiency within the material, typically associated with reduced defect states. As non-radiative recombination centers, trap states quench photogenerated carriers, thereby diminishing effective absorption. The above results do not contradict the hypothesis that the introduction of PVCn may contribute to the passivation of trap states in PbS NCs, yet direct evidence supporting this specific passivation effect remains currently missing. A visual inspection of the STEM image in Supplementary Fig. 6 indicates that the Pb and C elements are distributed in a uniform manner across the PVCn: PbS blend film. Atomic force microscopy (AFM, Supplementary Fig. 7) provided further confirmation of the film’s homogeneity. In addition, spectral absorption tests were performed on PVCn: PbS NC blend films and C8-BTBT crystal films to illustrate that they have desirable photosensitive properties. The absorption spectra (Fig. 2c) show that the PVCn: PbS NC blend film has effective absorption in the visible and infrared bands, and the C8-BTBT crystalline film has a strong absorption phenomenon for UV light between 300–400 nm. The trend of the absorption spectra is in agreement with the results of the previous report24. Although C8-BTBT exhibits photo responsive properties in the ultraviolet band, in the actual operation of ODAS, light stimulation is input via the gate. Under weak light conditions, short-wavelength light struggles to penetrate the PVCn: PbS layer and participate in the excitation of charge carriers in C8-BTBT. To validate this hypothesis, we conducted comparative experiments (supplementary Fig. 8). Two sets of devices with the same PVCn thickness were prepared. One group was doped with PbS quantum dots as the experimental group, while the other group consisted of pure PVCn thin films as the control group. The UV light response of the devices was tested under both conditions. The results showed that the devices exhibited significant photocurrent response only under the PVCn: PbS thin film, further confirming that the PVCn: PbS layer is the primary source of light absorption and photo-generated carriers.

ODAS benefits from the accumulation of carriers in the channel during the light adaptation process, which gradually increases the photosensitivity. Figure 2d illustrates this mechanism. In absolute darkness, the device is in a non-adapted state, with almost no electron–hole pairs in the PbS layer, and the conductivity of the C8-BTBT layer is mainly determined by the concentration of intrinsic carriers. When the device is in a photon-starved environments, photons in the PbS layer capture energy and generate electron–hole pairs. As the irradiation continues to extend and an external gate voltage is added, the electron-hole pairs separate, with the electrons being captured by the active functional groups in the lower dielectric layer and the holes accumulating in the device channel through the gating effect, resulting in an enhancement of the channel current, an improvement in the photosensitivity and ultimately reaching an equilibrium. After the light is removed, some of the trapped electrons with shallow energy levels are released to recombine with the holes, but most of the electrons remain trapped due to the energy level barrier, maintaining the photosensitivity of the device. This is the reason for the non-volatile nature of persistent photocurrent (PPC). The energy band structure of this mechanism is illustrated in Supplementary Fig. 9. To verify the light adaptation characteristics of ODAS, the IDS was tested at a specific low light intensity as a function of the adaptation time, and the results are demonstrated in Fig. 2e. Even at an ultra-low light intensity of 0.146 nW cm−2, the adaptation current of ODAS still increases and accumulates within a time scale of dozens of seconds (consistent with biological organisms), demonstrating significant adaptation effects. Furthermore, based on the light sensitivity characteristics of the owl retina, which are characterized by an increase in the postsynaptic potential of rod cells, the light sensitivity of this device can be quantified as Pt, calculated as follows:

$${P}_{t}=\frac{{I}_{t}-{I}_{{{\rm{dark}}}}}{{I}_{{{\rm{dark}}}}}$$
(1)

where Idark is the dark current and It is the channel current of the device after t seconds of adaptation22. Figure 2f shows the proportional relationship between Pt and illuminance for different adaptation times. In the biological visual adaptation system, active adaptation is usually described by the change in sensitivity during exposure. Here, the time-dependent active adaptation index (AAI) is defined as the change in Pt as a function of the order of magnitude change in illuminance, defined as follows:

$${{\rm{AAI}}}=\frac{d{P}_{t}}{d\log L}$$
(2)

where L is the illuminance value. It is generally accepted that Pt is independent of L when AAI is less than 1.0[22]. AAI of ODAS with luminance for different adaptation times is given in Fig. 2g. When the luminance is of ≤0.73 nW cm-2 and the adaptation time is 5 s, the AAI stays below 1.0, and the sensitivity growth is not activated. When the adaptation time was more than 10 s, the AAI remained high level at all illuminances, and Pt was positively correlated with L. When the exposure time was 25 s, AAI reached a maximum value of 331 at a light intensity of 0.146 nW cm-2. This further proves that ODAS still has positive photoadaptation in low-light environments. Since light signals in the real world have dynamic complexity, Supplementary Fig. 10 shows additional tests for dynamic adaptation under conditions where light input changes over time, verifying the response capability of ODAS in non-ideal dynamic environments.

Reliability of synapses array

To further simulate the dynamic adaptability of the owl’s retina to low illumination levels and to verify the reliability of ODAS in future practical applications, we constructed a 19 × 17 ODAS active array. The arraying of synaptic devices is necessary for the weight calculation of the neural network in the inference stage. As shown in Fig. 3a, the array architecture consisting of synaptic transistors can be used for weight deployment in a variety of inference models such as multilayer perceptron (MLP), convolutional neural network (CNN), and deep neural network (DNN). Optical photographs of the ODAS array are demonstrated in Fig. 3b (the manufacturing method of the synapses array refers to the methods section and Supplementary Fig. 11). The array is optically transparent due to the substrate of glass with ITO parallel electrode strips. The equivalent circuit of the active array is shown in Fig. 3c. Bit line, word line and drain voltage construct the cross-array architecture. Synaptic conductance modulation in the ODAS is achieved by applying voltage signals to the array through the word line. As shown in Fig. 3d and Supplementary Fig. 12, we selected the 17 × 17 portion of the artificial synapse array (excluding the boundaries) for device-to-device bidirectional IDS-VG scanning tests and IDS-VD tests. ODAS exhibits typical p-type transistor characteristics, with a VTH of approximately 1.5 V. The curve demonstrates a counterclockwise hysteresis under bidirectional scanning, indicating that “the device possesses storage capability. Subthreshold swing (SS) of the ODAS is ~200 mV dec−1, exhibiting efficient electrostatic tuning and excellent amplification characteristics. Figure 3e shows the spatial distribution of VTH in the array with a root mean squared error of ~0.09 V, proving high uniformity of the device array. This is further supported by the statistical distribution of the on/off ratio (Fig. 3f). Additionally, the conductivity retention characteristics and cyclic stability test results of ODAS are shown in Supplementary Fig. 13. The conductivity of the device did not show significant decay within 104 s after light stimulation, and no significant degradation of ODAS performance was observed during 10⁴ IDS-VG cyclic scans. More notably, within the operating range of 293.15–353.15 K and 30-80% humidity, the device’s transmission characteristics remained stable, with no observed degradation of the conductive state or systematic threshold drift (Supplementary Fig. 14). This provides direct evidence of the device’s engineering feasibility under multi-environmental conditions. After confirming the consistency of the synapse array, in order to generate a realistic adaptation and imaging scene under low light conditions, an optical pattern consisting of the five letters, “X”, “J”, “T”, “L”, and “U” was projected onto the device array (light intensity = 0.73 nW cm−2). Figure 3g visualizes the change in the normalized photocurrent of the projected device during the 20-second adaptation time. As the adaptation time increases, image details are gradually distinguished due to the differences in illumination on the pattern surface. This process reveals the visual adaptation and imaging of targets by the owl’s retina under ultra-low illuminance conditions, and verifies the feasibility of the ODAS array in future machine vision.

Fig. 3: Circuit architecture and performance characterization of ODAS array.
Fig. 3: Circuit architecture and performance characterization of ODAS array.
Full size image

a Schematic diagram of the ODAS array. b Optical image of the 19 × 17 ODAS array (scale bar: 2 mm) and magnified image of a single device (scale bar: 50 μm). c Equivalent circuit topology: the cross-array architecture is constructed by the bit line, word line and drain voltage. d IDS-VG bidirectional sweep curves between devices. e Spatial distribution of device threshold voltages on the array. f Analysis of the uniformity of the switching ratio of the array. The statistical distribution shows a Gaussian character (μ = 3.12, σ = 0.05). g Visualization of the dynamic optical adaptation process on the ODAS array. Light stimuli were applied to the array in a directional manner to form the letters “X”, “J”, “T”, “L” and “U”. The photocurrents of the activated devices gradually increased during the adaptation process.

Photo-controlled synaptic properties

In the owl’s visual system, light passes through the cornea and lens of the eye and focuses on the retina, where the synaptic structures of the photoreceptor cells are responsible for converting and initially encoding the light signal. When photons reach the photoreceptor cells, the opsin protein triggers a series of biochemical reactions that cause the synaptic sodium channels to close, thereby reducing the entry of sodium ions into the cell. This causes the cell membrane to depolarize, which in turn generates an action potential25. The highly sensitive rod cells in the owl’s retina give it excellent night vision, enabling it to accurately identify targets even in <0.01 lux19,26. What’s more, owls with special cone cells can perceive a wider range of light (from near-infrared to ultraviolet) than humans. ODAS’ photoreceptor characterization method uses light pulses as input. Figure 4a displays the specific test strategy. A light pulse of a specific energy is input into the transparent gate of the device using a light-emitting diode connected to a signal generator as the light source. At the same time, a voltage VDS is applied to the source-drain to monitor the current and observe the non-volatile response of the ODAS to the optical input. Figure 4b describes the expected input/output schematic. The synapse is continuously stimulated with light in the form of identical pulses, and the current output shows a step change. The pulse information is memorized. In the actual light pulse tests, five wavelengths of light-emitting diodes were employed based on the owl’s wavelength perception range: 365 nm, 470 nm, 590 nm, and 940 nm. The output level of the signal generator was adjusted using a light intensity meter placed in the test environment so that the five diodes emitted light pulses of the same intensity. Figure 4c shows the ODAS response to four wavelengths of light stimuli (1.46 nW cm−2) at VDS = −1 V, with a positive bias voltage of 0.5 V (Supplementary Fig. 15 illustrates that ODAS has the highest Pt at a bias voltage of 0.5 V). The test results indicate that the device has the same non-volatile persistent photocurrent (NV-PPC) trend for the four wavelengths of 365 nm-940 nm, and the current gain increases with the decrease of the light wavelength. This result is consistent with the expected stepwise current change. This indicates that the ODAS is capable of receiving a wide range of optical signals from ultraviolet to visible to infrared to support its synaptic performance. In addition, the intensity dependence of the ODAS as an optical synapse for pulsed signals was also tested (Fig. 4d). The device showed a positively correlated PPC change for different intensities (0.73 nW cm−2, 1.46 nW cm−2, and 4.38 nW cm−2) of optical pulse inputs. The attenuation rates (≤1.3%) after stimulation with light pulses of three intensities were calculated in Fig. 4e, demonstrating the long-term optical information memory property. Supplementary Fig. 16 also supplements the ODAS pulse width dependence and number dependence tests, which proves that ODAS has a rich ability to modulate optical signals.

Fig. 4: Photo-controlled synaptic properties of ODAS.
Fig. 4: Photo-controlled synaptic properties of ODAS.
Full size image

a Schematic of the ODAS photo synaptic test, with optical pulses input from the transparent gate at the bottom of the device. b Optoelectronic synapse has a non-volatile response to optical pulses input. The change in current output shows the ability to memorize information. c ODAS non-volatile photocurrent in response to light pulses of different wavelengths (365 nm, 470 nm, 590 nm, and 940 nm). d ODAS non-volatile photocurrent in response to light pulses under different illuminances (0.73 nW cm−2, 1.46 nW cm−2, and 4.38 nW cm−2). e Decay rate of the photocurrent after the light stimulus. f PPC of ODAS under ultra-low light intensity of 0.146 nW cm−2 at 940 nm. g Comparison of ODAS with the reported optoelectronic synaptic devices in terms of energy consumption and weak light detection ability. h ODAS’s photonic information re-memory ability test.

Furthermore, the limiting optical intensity for ODAS to perceive light signals was investigated. We used a 940 nm light source to stimulate the device. Figure 4f shows that the ODAS retains a significant non-volatile current gain even at a light intensity of 0.146 nW cm−2. This result confirms that the device can capture weak photons. Figure 4g provides a comparison of ODAS with other optoelectronic synaptic devices that have been reported in recent years17,25,27,28,29,30,31,32,33,34,35,36,37. The vertical axis represents the energy consumption of the device for detecting an optical pulse, while the horizontal axis indicates the light intensity for optical detection. ODAS achieves a low light detection capability at 0.146 nW cm−2 with a low energy consumption of 0.4 nJ per operation. Supplementary Table 1 provides detailed parameters of the cited literature. Finally, the “re-memory” ability of ODAS was demonstrated (Fig. 4h). During the test, 19 nA was specified as the target for synaptic weight regulation (normalized to 1.0). To reach the target during the initial perception, 21 input light pulses were required. During the second perception, the number of input pulses was reduced, and only 9 pulses were needed to reach the target. In the third trial, just 6 pulses were necessary. The ability to “re-memory” is a key attribute of synaptic devices for artificial intelligence applications, as it emphasizes the biologically dynamic optimization of synaptic weights38,39.

Electrically controlled synaptic properties

Synapses located in the central nervous system of living organisms enable the exchange of information between nerve cells mainly through the transmission of electrical signals. Similar to the nervous system, electrical pulses are used as input signals in the working of the computational modules in the physical neural computing architecture40,41. The plastic modulation of the postsynaptic current by the pulse input is one of the necessary functions of an artificial synapse. In the prepared ODAS device, the gate electrode and the semiconductor channel layer represent the pre-synapses and post-synapse, respectively. The excitatory postsynaptic current (EPSC) responds to the channel conductance, which represents the modulation of the post-synaptic current by the pre-synaptic signal. Figure 5a shows a schematic diagram of the synaptic test with a gate voltage pulse as the input. According to band engineering analysis, when a positive gate voltage is applied, the gate electric field induces electron tunneling from the PVCn: PbS layer to the deep defect states of the PVA electron capture layer. This process leads to two synergistic effects: first, the concentration of holes in the C8-BTBT channel rises, and the conductivity increases; second, the electrons captured by PVA form localized potential traps, inhibiting carrier recombination and achieving long-lasting retention of the conductive state. When the negative gate voltage is applied, the electrons trapped in the PVA layer escape to the PVCn: PbS layer through emission, while the gate electric field repels the holes in the C8-BTBT, resulting in a decrease in the non-volatile conductivity.

Fig. 5: Synaptic characteristics test using electrical pulses.
Fig. 5: Synaptic characteristics test using electrical pulses.
Full size image

a Schematic diagram of the ODAS electrical pulse test. The pulse input is regulated by the gate voltage. b Dynamic response of the EPSC. The EPSC shows a step-like increase in the 0-16 s range under the action of alternating pulses with a baseline voltage (0.2 V) and a peak voltage (1.2 V) (pulse width 500 ms), which is consistent with the linear dynamics of LTP. c Amplitude-dependent electrical pulse conflict synaptic current test. Comparison of the EPSC amplitude changes under different pulse amplitudes (1.0 V, 1.5 V, and 2.0 V) at a fixed pulse width of 500 ms and a frequency of 1 Hz. d LTP and LTD cycle characteristics. The peak voltage switches between 1.2 V and −0.2 V, showing 30 cycles of stability. e Boxplots for high PSC value and low PSC value of the LTP-LTP cycles test (n = 30). For box plots, the horizontal central line represents the median value, the lower and upper quartiles represent the 25th and 75th percentile, and the whiskers show the maximum and minimum values. f Schematic diagram of hardware mapping of floating-point weights and 4-bit discrete weights. The weight matrix of the fully connected layer of the neural network is physically encoded by the synaptic conductance values. g 4-bit simulation performance of the MLP on the MNIST dataset based on ODAS: the test accuracy reaches 90% (blue curve), and the training loss value converges to 0.12 with the increase of epoch (orange curve).

As shown in Fig. 5b, when a positive electrical pulse (Vbase = 0.2 V, Vpeak = 1.2 V, width = 500 ms) is applied to ODAS, there is a non-volatile increase in EPSC, which is consistent with the behavioral trend of synaptic LTP. In addition, when voltage pulses of different amplitudes (1.0 V, 1.5 V, 2.0 V) are applied, the EPSC increase is different (Fig. 5c), which indicates that ODAS has amplitude-dependent electrical pulse conflict plasticity. Supplementary Fig. 17 further demonstrates that ODAS electrical pulse conflict plasticity is also related to pulse width, frequency, and number. In contrast to LTP, LTD is regulated by applying negative gate voltage to the device. Figure 5d shows that after 16 positive pulses (Vpeak = 1.2 V, Vbase = 0.2 V), the same number of negative pulses (Vpeak = −0.2 V, Vbase = 0.2 V) are applied immediately afterwards. The PSC level returns to the initial state, completing an LTP-LTD cycle. After 30 consecutive cycles of this test, no significant degradation in device conductance is observed. Figure 5e gives a statistical analysis of the PSCs corresponding to the highest and lowest conductance states during these 30 cycles. The median of high PSC value remains at 32–34 nA, and the median of low PSC value remains at 13–16 nA, revealing the cyclic stability of ODAS in the regulation of synaptic weight by electrical control. Additionally, the saturation capacity and maximum operating frequency of the ODAS intermediate conductive state were tested. After applying 128 positive and negative pulses (7 bits) to the ODAS in sequence (Supplementary Fig. 18), the device exhibited a richer intermediate state, enabling precise neural network regulation. Furthermore, the operation frequency of up to 10 Hz supports the update rate of the ODAS in actual weight deployment (Supplementary Fig. 19). In addition to the validation of the pulse plasticity of the ODAS, the energy consumption of the device in practical operation was also tested, which provides confirming its low energy consumption characteristics of the ODAS for future use in neuromorphic computing architectures. As shown in Supplementary Fig. 20, at very low VDS and electrical pulse stimulation, the ODAS was still able to trigger an EPSC response with the energy consumption calculated as:

$$E={V}_{{{\rm{DS}}}}\times {I}_{{{\rm{peak}}}}\times t$$
(3)

where VDS = −0.1 mV is the voltage applied to the drain terminal of the synaptic transistor, Ipeak = −9.2 nA is the spike current flowing from the drain to the source terminal, and t = 10 ms is the width of the electrical pulse. Thus, the minimum consumption of ODAS per synaptic event is 9.2 fJ, which is lower than the energy consumption of biological synapses (~10 fJ per synaptic event)25.

To verify the feasibility of deploying ODAS in electronic control mode for artificial neural network tasks, an actual network model was built to perform the image classification task. Figure 5f shows the mechanism of network weight deployment at first. As the most important part of the ANN calculation, the weight matrix (Wm × n) stores the weight values between network nodes. During the forward propagation process, the weight matrix is multiplied by the input matrix to obtain the inference result, and during the backpropagation process, the network performance is dynamically updated and optimized. In traditional computer architecture, a weight value usually consists of a floating-point number, occupies a 32-bit storage space, and can be regarded as a continuous value. However, the weight value represented by an ODAS artificial synapse is 4 bits (16 conductance states), which can be regarded as discrete. Although discrete weights will inevitably lead to a decrease in accuracy in neural network calculations, they have absolute advantages in terms of saving physical space, reducing computational energy consumption, and improving calculation speed. Figure 5g demonstrates the classification results of the MNIST handwritten digit dataset after deploying the 4-bit conductive state space of ODAS in MLP (Supplementary Table 2) network. The classification accuracy is >90% on both the training set and the test set, and the network loss value tends to converge. This demonstrates the feasibility of integrating ODAS into basic neural networks. To further verify the generalization ability of ODAS weight tuning, we trained three datasets (MNIST, Fashion-MNIST, and CIFAR-10, Supplementary Fig. 21) in MLP, CNN (Supplementary Table 3), and a deep learning network (VGG16), respectively. Supplementary Table 4 and Supplementary Figs. 22–24 summarize the classification accuracy comparison of ODAS 4-bit conductive state and 7-bit conductive state participating in the above task simulations. This validates the high network generalization ability of the synaptic device, which can integrate network models of multiple complexities and perform diverse image classification tasks.

Target recognition in photon-starved environments

This study constructs an air-to-ground target recognition system based on the bionic light adaptation mechanism. The near-sensor computing system composed of ODAS is simulated by a drone, which simulates the photoadaptation of owl vision and the central nervous system’s image processing capabilities in a night environment (Fig. 6a). The experiment was conducted in a five-level gradient brightness scene, with a light intensity range of 0.146–11.7 nW cm−2. A low-resolution garment was used as the recognition target (Fig. 6b). Supplementary Fig. 25 and Fig. 26 show the UAV platform used for dataset collection with part of the obtained air-to-ground images. The dataset features multiple dimensions: varying angles, backgrounds (turf, woods, riverside, and mudflat), and heights (30-100 m). The images in different brightness environments are preprocessed by simulating the relationship between root mean square (RMS) contrast, defined at Supplementary Note 1, and adaptation time shown in Fig. 6c. When the ambient light intensity is 0.146 nW cm−2, the system increases the image RMS contrast from 0 to 21 within 20 s. This process constructs a nonlinear conversion model of photocurrent-contrast by mapping grayscale values to normalized photocurrents (0–255 range), providing a training dataset for the neural network model to adapt to low signal-to-noise ratio conditions (refer to Supplementary Fig. 27 for feature distribution). For the computational module in the near-infrared receptor computing architecture, YOLOv5_s is used as the object to be deployed for the synaptic matrix42 (Supplementary Fig. 28), and the readout time for this architecture’s single-layer synapse was investigated (Supplementary Note 2). The results of network training (Fig. 6d) show that during training, the loss function converges faster with increasing brightness. For the 11.7 nW cm-2 light adaptation dataset, it converges to below 0.2 after 80 epochs during training, while for the 0.146 nW cm-2 light adaptation dataset, it takes 180 epochs to converge to the same level. The performance evaluation of the accuracy is displayed in Fig. 6e and Supplementary Fig. 29. As the training cycle increases, mAP_0.5 improves from the initial value of 0.62 to 0.98, mAP_0.5:0.95 is optimized from 0.31 to 0.89, and the maximum precision greater than 95%, which verifies the enhancement of the light adaptation mechanism on multi-scale object detection. Further analysis shows that the system’s recognition confidence is related to the adaptation time of the dataset (Fig. 6f). When the light intensity is 0.73 nW cm−2, the network’s confidence in target grasping increases from 0 to 0.85 during an adaptation time of 20 s. Figure 6g and Supplementary Table 5 compare the performance of ODAS with that of the recently reported artificial optical synapses in a neural network recognition simulation43,44,45,46,47,48,49,50. The higher level of recognition accuracy (95.2%) of our device under a much weaker light stimulus (0.146 nW cm−2) demonstrates that the ODAS-based simulation system has the ability to recognize targets in dark environments. This result provides a hardware-algorithm collaborative verification framework for the application of biomimetic machine vision in fields such as night search and rescue and deep space exploration.

Fig. 6: Schematic of intelligent target recognition applications.
Fig. 6: Schematic of intelligent target recognition applications.
Full size image

a Schematic of nocturnal air-to-ground recognition using UAVs to simulate the barn owl. b Schematic of the UAV’s optical capture of a target in five brightness levels. c Relationship between RMS contrast and adaptation time. The RMS contrast of an image can be defined based on the gray value of the image mapped to the normalized photocurrent under different light intensity. d Training process of the YOLOv5 network on the dataset emulated with different illuminance levels, and the network loss value converges with the increase of the training cycle. e Trend of precision, mAP_0.5 and mAP_0.5:0.95 under different training epochs. f Effect of changes in contrast during adaptation on confidence level in target recognition. g Comparison of ODAS with reported optical synaptic devices for recognition simulation.

Discussion

In conclusion, to address the significant challenges in the perception and integration of neuromorphic physical vision systems in ultra-low light environments, we have proposed ODAS that possesses nocturnal adaptive capability and bimodal stable modulation for near sensor computing. By separating photocarrier generation and the electrical amplification process, ODAS devices exhibit owl-like dark adaptation capabilities with a high AAI of 331 and thus achieves a low perceptible light intensity of 0.146 nW cm−2. Furthermore, the ODAS array verifies the high homogeneity and pattern adaptation homogeneity of the device. Under electric pulse modulation, the ODAS exhibits long-term potentiation and depression characteristics, and with an energy consumption of only 9.2 fJ per synaptic event, which lays the foundation for biologically efficient computation. Validated by multiple underlying neural networks (MLP, CNN, and VGG16), ODAS achieves stable >90% accuracy in image dataset classification tasks. Additionally, a UAV-based air-to-ground target recognition system demonstrates that a near sensor computing system based on ODAS delivers target recognition accuracy greater than 95% in 0.146–11.7 nW cm−2 environments through simulated adaptive RMS contrast optimization, providing a perceptual-computational fusion solution for nighttime machine vision without the need for auxiliary illumination. This work extends the detection limit of neuromorphic vision sensors to scotopic-level illuminance and verifies the feasibility of future near-sensor computing physical architectures for integration, holding great promise for broad applications such as deep space exploration and emergency rescue.

Methods

Material

Anhydrous toluene was acquired from Shanghai Adamas Reagent Co., Ltd. Polyvinyl alcohol (with an average molecular weight of about 31 kDa) was purchased from Shanghai Aladdin Biochemical Technology Co., Ltd. Chlorobenzene (99.5%) was sourced from Macklin, while butylamine (99%) came from TCI Chemicals. Polyvinyl chloride (PVCn) with an average molecular weight of approximately 40 kDa, polystyrene (PS) at 200 kDa, and PFBT were sourced from Sigma-Aldrich Co. Additionally, the compound C8-BTBT was provided by Luminescence Technology Corp. (Lumtec). All materials and solvents were used as received without further purification.

Fabrication of the films

To prepare the C8-BTBT crystalline film, 10 mg of C8-BTBT and 10 mg of polystyrene (PS) were dissolved in 1 mL of toluene. The resulting solutions were combined at a 1:1 volume ratio and subjected to ultrasonication for about 2 min to create the C8-BTBT solution. For the electrode modification solution, 25 μL of PFBT was mixed with 25 mL of isopropanol to produce the PFBT solution. C8-BTBT crystalline films were fabricated on the surface of a substrate covered with source/drain electrodes using blade coating in ambient air at room temperature. A blended solution of C8-BTBT: PS was injected into the gap between the substrate and the blade. It was then moved at a consistent speed of 200 μm s-1 using a stepper motor. As the solvent gradually evaporated at the front end of the meniscus, the crystals grew continuously, resulting in a large and uniform C8-BTBT crystalline film. The crystallinity of the thin films was analyzed using a crossed-Nicols polarized microscope (Olympus BX53). In-plane XRD measurements were performed with a thin-film diffractometer (Bruker D8 DISCOVER) utilizing monochromated CuK radiation. The diffraction intensity was recorded with a scintillation counter under normal conditions, with a grazing incident angle of 0.20° set for the in-plane XRD measurement.

To prepare the PVCn: PbS NC blend film, 25 mg mL−1 of PbS NC solution was added to 60 mg mL−1 of PVCn solution and stirred for one minute to obtain a mixed PVCn: PbS NC solution. The obtained solution was spin-coated at 500 rpm for 5 s and then at 3500 rpm for 60 s; the second step of spin-coating for 5 s was followed by vacuum annealing at 100 °C for 1 h to prepare a 400 nm thick PVCn: PbS NC blend film. The films were then exposed to UV light for 1 h for UV crosslinking. Film surface morphology and thickness were measured by an atomic force microscope (Asylum Research Cypher S). For XPS measurements, the grown hybrid films were sealed in a vial, placed in a glove box filled with N2, and then quickly transferred to an XPS vacuum chamber. The XPS studies were performed using a Kratos Axis Ultra DLD spectrometer equipped with a panchromatic Al Ka X-ray source. The spectrometer was equipped with a panchromatic Al Ka X-ray source. operating at 150 W and a vacuum pressure of 10−9 mbar.

Fabrication of the device

The substrate was first prepared by cleaning the ITO-coated glass using a UV ozone cleaner for 20 min. 15 mg of PVA was added to a mixture of water-ethanol ratio of 1:1 for the preparation of PVA solution, which was then spin-coated onto the substrate (at a speed of 5000 r min−1), at which point the thickness of the PVA film was equal to approximately 50 nm. The film was then annealed for one hour at 100 °C on a heating table. PVCn: PbS NC blend films were prepared on the PVA surface as described. After drying, Ag electrodes were prepared on their surface by thermal evaporation (deposition rate was about 0.4 Å s−1) and Ag vapors were deposited as S/D electrodes under the mask plate. Then, C8-BTBT crystal films were prepared on the substrate as described above. In order to improve the work function of the Ag electrode and enhance the hole injection efficiency, the substrate was immersed in PFBT solution for 5 min in advance and the Ag surface was modified with PFBT at room temperature. Cross-sectional TEM samples of the device were prepared by a standard focused ion beam boosting process on a FEI quantum 3D dual-beam microscope. TEM images of the device were obtained on a FEI Talos 200X microscope with an acceleration voltage of 200 kV. Further, regarding the preparation method of a 19 × 17 ODAS array, ITO parallel electrode strips were firstly employed as gate electrodes (spacing = 200 μm). Then PVA and PVCn: PbS NC blend film was sequentially deposited onto the gate electrode array as described above. In this process, a portion of ITO was exposed by placing a metal bar layer on the edge of the ITO electrode strips. Then S/D electrodes were precisely deposited on the top of the gate electrodes by a patterned mask plate. After performing the PFBT treatment process, a thin film of C8-BTBT is deposited on the S/D electrode pattern while the mask is removed. The ODAS device array fabrication is completed.

Electrical measurements

A source meter (Keysight B1500A) coupled to a probe table setup was used to evaluate the performance of ODAS single devices versus synaptic arrays in terms of photoelectrical synaptic properties. The measurements were performed in a closed chamber at room temperature. The chamber was shielded from external light or radiation interference under normal atmospheric conditions and grounded to avoid electro-magnetic interference (EMI). Monochromatic light emitting diodes (LEDs) of four wavelengths (365 nm, 470 nm, 590 nm, and 940 nm) are utilized to generate the light stimulus. The pulse input was controlled by a signal pulse generator unit (SPGU, Keysight B1525) mounted on the light source instrument. Light intensity was measured by using a calibrated silicon photodetector (Newport 843 R with an 818UV/DB optical power detector).

Simulation of image recognition

To simulate the visual adaptation in the owl retina, we first defined the function for the contrast of imaging (Ct) obtained from the light intensity-dependent and time-dependent properties of the ODAS23:

$${C}_{t}=255\times \frac{{I}_{t}-{I}_{{{\rm{dark}}}}}{{I}_{{{\rm{s}}}}-{I}_{{{\rm{dark}}}}}$$
(4)

where It is the current after adaptation for time t. Idark is the dark current, defined here as 10 pA, corresponding to gray level 0. Is is the saturation current, defined as 2 nA, corresponding to gray level 255. Ct obtained under different light intensities is then used as the RMS contrast index of the preprocessed image to obtain the simulated image of ODAS perception and adaptation. The neural network simulation is realized through a well-established synaptic computing framework51,52. Nonlinearity of LTP and LTD, defined at Supplementary Note 3, was first evaluated using the NeuroSim+ framework at a fixed number of applied stimuli53. Subsequent deployment of weight matrices in each model was performed using normalized conductance values (G), which was calculated as follows: G = (Gn - Gmin)/(Gmax - Gmin), where Gn is the conductance value of the nth pulse, Gmax is the maximum conductance value, and Gmin is the minimum conductance value. Due to the need to dynamically track model performance changes, the deployment of weights and evaluation of the simulated physical model will be performed after each epoch training.