Bio‐Inspired Full‐Spectrum Bidirectional Optoelectronic Synaptic Transistor for In‐Sensor Neuromorphic Vision

ABSTRACT

Inspired by avian visual systems, wide‐spectrum bidirectional optoelectronic synaptic transistors (OSTs) have emerged as a transformative platform for neuromorphic vision processing. However, current artificial synaptic technologies remain constrained by fragmented control paradigms that rely on wavelength‐selective or electrical polarity‐dependent modulation, restricting their capacity for complex visual information processing. In this study, an organic 2D molecular crystal (2DMC) heterojunction is developed to achieve high‐performance full‐spectrum bidirectional OSTs. The high‐quality 2DMC heterostructure facilitates efficient exciton separation and diffusion, yielding remarkable ambipolar characteristics with an exceptional responsivity of 6.3 × 10⁴ A W⁻¹ under weak light. The bipolar heterostructure transistor design enables dynamic reconfiguration between excitatory and inhibitory synaptic behaviors through dual modulation mechanisms of spectral tuning and gate polarity control, successfully emulating bidirectional neuroplasticity features. The practical viability of this technology is demonstrated through implementation in a broad‐spectrum intelligent‐vehicle‐mounted vision system that achieves a classification accuracy of >90% for traffic objects across complex scenarios. This work not only provides an effective strategy for developing high‐performance bidirectional optoelectronic synapses but also paves the way for advanced neuromorphic computing and intelligent machine vision technologies.

The persistent challenge of integrating full‐spectrum sensing with bidirectional neuromorphic plasticity in bioinspired optoelectronics is addressed through heterostructural engineering of organic 2D molecular crystals, enabling efficient exciton separation and dynamic reconfiguration between excitatory and inhibitory behaviors. These devices exhibit exceptional responsivity (6.3×10⁴ A W⁻¹) and achieve over 90% classification accuracy in a broad‐spectrum neuromorphic vision system for intelligent vehicles.

1. Introduction

Neuromorphic computing devices with in‐memory computing capabilities have advanced rapidly to address the memory wall bottleneck of von Neumann architectures [1, 2, 3, 4]. The parallel processing and integrated sensing‐memory‐computing architecture of these devices offer significant potential for edge computing applications in machine vision and deep learning‐based technologies such as intelligent assisted driving and artificial vision chips [5, 6]. Conventional vision sensors rely on a “sense–store–compute” separated architecture, which necessitates frequent data transfers between graphic processing units and memory for multispectral environmental data (e.g., visible/near‐infrared (NIR)/ultraviolet (UV)). This requirement not only incurs significant power consumption [7, 8, 9] but also introduces processing latency [10, 11, 12], making it unsuitable for resource‐constrained applications such as intelligent assisted driving. To resolve this problem, neuromorphic vision sensors based on artificial synapses have evolved from simple two‐terminal memristors to more biologically plausible three‐terminal transistor synapses for providing a new hardware foundation for real‐time and “sensing–computing integrated” vision processing [3, 4, 7, 8, 13, 14].

However, existing devices in the field of optoelectronic neuromorphic computing face a fundamental limitation: their operational modes are overly simplistic. Most devices rely solely on wavelength selection [11, 12, 15, 16, 17, 18, 19] or electrical polarity reversal for unidirectional synaptic weight modulation [20, 21]. This fragmented control strategy cannot satisfy the demands of complex visual information processing. For example, intelligent assisted driving systems require multispectral sensing channels to cooperatively process and switch between spectral responses in real time. Current wavelength‐modulated optoelectronic synapses depend heavily on light‐induced current enhancement or suppression [22, 23, 24, 25], severely limiting switching speed and precision and often falling short of brain‐like millisecond‐level responses. Conversely, purely voltage‐modulated synaptic devices offer fast switching, while sacrificing the multidimensional sensing and control advantages of optical signals [7, 26, 27, 28]. Moreover, conventional organic semiconductor‐based thin‐film optoelectronic devices suffer from low charge‐carrier mobility and high interfacial defect density, which results in photoresponsivity and sensitivity performances far below the requirements for intelligent assisted driving [29, 30]. In contrast, avian visual systems employ a sophisticated spectral‐polarity cooperative encoding mechanism (Figure 1a). Specialized photoreceptors such as double cones and oil droplet filters enable full‐spectrum perception from the UV to IR (300–900 nm) wavelengths, while adaptive complementary interactions among visual cells facilitate bidirectional synaptic plasticity and visual adaptation. This enables birds to dynamically adjust their visual sensitivity across the UV–NIR range. However, current optoelectronic synapses demonstrate a fragmented control mechanism that relies on wavelength or polarity, and consequently, they cannot adequately handle complex visual scenarios. Thus, full‐spectrum, bidirectionally tunable optoelectronic artificial synapses with high photosensitivity and responsivity are urgently required for machine vision and intelligent assisted driving applications.

FIGURE 1.

Full‐spectrum bidirectional artificial synapse inspired by avian vision. (a) Schematic of the retina and tetrachromatic vision system of a bird. (b) Schematic of the designed bidirectional optoelectronic synaptic transistor based on TFT‐CN/C8‐BTBT 2DMC heterojunction. (c) Vehicle‐mounted vision system comparison: conventional RGB vision (left) versus an intelligent broadband system (right).

In this study, we propose a novel full‐spectrum bidirectional optoelectronic synaptic transistor (OST) based on organic 2D molecular crystal (2DMC) p‐n heterojunctions that integrates wavelength‐selective and electrical‐polarity‐dependent tunable synaptic bidirectional modulation in a single platform (Figure 1b). This breakthrough addresses the critical limitations of current neuromorphic technologies, opening new avenues for developing advanced neuromorphic devices tailored for intelligent machine vision applications. The exceptional crystallinity and molecular‐scale thickness of the 2DMC heterostructure enable efficient exciton separation and diffusion, resulting in pronounced ambipolar characteristics with outstanding photoresponses. Remarkably, the proposed device achieves photosensitivity (P) and photoresponsivity (R) values of 4969.3 and 6.3 × 10⁴ A W⁻¹, demonstrating superior light detection capabilities. Further, the device exhibits comprehensive bidirectional synaptic plasticity that can be dynamically reconfigured through wavelength‐specific optical inputs or gate‐voltage (V _GS) polarity reversal, successfully reproducing fundamental synaptic functions that include short‐term plasticity (STP), long‐term plasticity (LTP), and paired‐pulse facilitation (PPF). Leveraging these unique switching properties, we successfully implement a broad‐spectrum vehicle‐mounted intelligent vision system incorporating our 2DMC heterojunction bidirectional OSTs (Figure 1c). This system achieves a classification accuracy >90% for traffic objects in complex scenarios through advanced multimodal sensory fusion, demonstrating its potential for next‐generation machine vision applications and intelligent transportation systems.

2. Results and Discussion

2.1. Morphology and Ambipolar Properties of 2DMC Heterojunctions

We prepared an organic 2DMC p‐n heterojunction with exceptional optoelectronic characteristics to design a high‐performance full‐spectrum bidirectionally tunable artificial synapse. For this heterostructure, we selected two archetypal semiconductors, namely, 2,7‐dioctyl[1]benzothieno[3,2‐b][1]benzothiophene (C8‐BTBT) and 2,2’‐((5Z,5’Z)‐5,5’‐(furan‐2,5‐diylidene)bis(4‐octylthiophene‐5,2(5H)‐diylidene))dimalononitrile (TFT‐CN) as the p‐ and n‐type components, respectively. These materials were selected based on their outstanding charge transport capabilities, superior optoelectronic performance, and remarkable atmospheric stability. The fabrication process involved a streamlined liquid‐phase substrate self‐assembly approach (Figure S1) for the controlled formation of isolated C8‐BTBT and TFT‐CN 2DMC layers. These well‐defined molecular crystalline layers were precisely stacked layer‐by‐layer to construct the high‐quality 2DMC heterojunction. The van der Waals interactions between adjacent molecular layers ensured exceptional structural integrity and optimal interfacial contact, yielding a defect‐minimized high‐performance heterostructure.

Optical microscopy (OM) images of C8‐BTBT and TFT‐CN 2DMCs (Figure S2) reveal uniform morphology over a substantial area. A distinct stacking boundary is clearly visible in the OM image of the 2DMC heterojunction (Figure S3). As illustrated in Figure 2a, atomic force microscopy (AFM) characterization reveals exceptionally smooth surfaces for both C8‐BTBT and TFT‐CN 2DMCs, with a sharply defined interfacial boundary in the resulting p‐n heterojunction. The root‐mean‐square roughness values are measured to be 0.22 nm for TFT‐CN/C8‐BTBT hetero‐stacked domains, unequivocally confirming the near‐atomic‐level flatness of both the constituent 2DMCs and resulting heterojunction. Altitude profiling reveals the well‐defined thicknesses of 11.6 and 1.6 nm for C8‐BTBT 2DMC and TFT‐CN 2DMC, respectively. They correspond to exactly 4 and 1 molecular layers (Note S1), respectively, demonstrating the successful fabrication of high‐quality ultrathin 2DMC heterojunctions. In highly crystalline few‐layer 2DMC, charge transport is strongly confined within the in‐plane 2D ordered π–π stacking networks, while interlayer electronic coupling remains relatively weak because of the insulating alkyl chains and van der Waals interactions. Therefore, even at a thickness of 11.6 nm, the C8‐BTBT layer largely exhibits 2D charge transport behavior instead of the conventional isotropic bulk‐like properties. Further, the thicknesses of the 2DMCs can be further fine‐tuned by modulating the concentration of the growth solution (Figures S4 and S5), offering additional flexibility in material design. Transmission electron microscopy (TEM) and selected area electron diffraction (SAED) shows a smooth crystal surface and a consistent single set of diffraction spots recorded from various regions of the same crystal, reinforcing the single‐crystallinity nature of both C8‐BTBT and TFT‐CN 2DMCs (Figure 2b and Figures S6 and S7). The phase purities of C8‐BTBT and TFT‐CN 2DMC are confirmed through X‐ray diffraction (XRD) patterns, which display sharp characteristic diffraction peaks (Figure 2c). These sharp diffraction peaks can be indexed to the (00l) facets for both C8‐BTBT and TFT‐CN crystals, suggesting an oriented packing of C8‐BTBT and TFT‐CN with their crystallographic ab‐plane parallel to the substrates. The XRD pattern of the 2DMC heterojunction presents all characteristic peaks of both C8‐BTBT and TFT‐CN, indicating a successful integration of both crystal types. Grazing‐incidence wide‐angle X‐ray scattering (GIWAXS) characterization is performed to characterize the crystallographic properties of the 2DMCs to gain deeper structural insights (Figures S8 and S9). The appearance of sharp, well‐defined Bragg diffraction spots unambiguously demonstrates the single‐crystalline nature and (001) out‐of‐plane direction of both C8‐BTBT and TFT‐CN 2DMCs. UV–Vis‒NIR spectroscopy is employed to characterize the optical properties of C8‐BTBT 2DMC, TFT‐CN 2DMC, and their 2DMC heterojunctions (Figure 2d). The C8‐BTBT and TFT‐CN 2DMCs exhibit absorption peaks at 357 and 830 nm, respectively, whereas their heterojunctions present overlapping absorption features of both C8‐BTBT and TFT‐CN 2DMCs. These results underscore the combined characteristics of the C8‐BTBT and TFT‐CN 2DMCs within their heterojunctions, thereby confirming the high quality of the as‐fabricated 2DMC heterojunctions.

FIGURE 2.

Structural and optoelectronic characterizations of the TFT‐CN/C8‐BTBT 2DMC heterojunction. (a) AFM topography of the 2DMC heterojunction. Scale bar: 1 µm. (b) TEM images and corresponding SAED patterns of the heterojunction. (c) XRD patterns of C8‐BTBT, TFT‐CN, and their heterojunction. (d) UV‒Vis‒NIR absorption spectra of C8‐BTBT, TFT‐CN, and their heterojunction. (e) Transfer characteristics of the heterojunction OST under 365 and 810 nm illumination at varying intensities. Illuminated transfer curves were recorded within 1 s after light onset. (f) Photosensitivity and (g) photoresponsivity as functions of V _GS and illumination intensity under 365 nm excitation. (h) Comparison of photosensitivity and photoresponsivity between our device and previously reported artificial synapses. (i) Photosensitivity and (j) photoresponsivity as functions of V _GS and illumination intensity under 810 nm excitation.

The inset in Figure 2e illustrates the optoelectronic synaptic device architecture. This architecture employs a three‐terminal organic field‐effect transistor platform with a conventional bottom‐gate top‐contact configuration, wherein the 2DMC p‐n heterojunction serves as an optically active channel layer. The electronic transport properties are systematically characterized under dark conditions; Figure 2e (gray curve) and Figure S10 present the representative transfer characteristics. The heterojunction optoelectronic synaptic transistor demonstrates pronounced ambipolar charge transport properties, with hole and electron mobilities determined to be 1.88 and 1.40 cm² V⁻¹ s⁻¹, respectively. The light intensity‐dependent transfer characteristics are examined under 365 nm (blue) and 810 nm (orange) light illumination to explore the photoresponse properties of the heterojunction OSTs. As depicted in Figure 2e and Figure S11, a positive shift in the threshold voltage (V _th) is observed under UV light exposure, with the magnitude of this shift increasing in tandem with illumination intensity, which indicates a significant photoresponse. In the hole‐dominant region, a positive photocurrent (PPC) is observed with positive illumination intensity, whereas in the electron‐dominant region, a negative photocurrent (NPC) is observed under the same conditions, emphasizing the bipolar characteristics of the device. The PPC and NPC behaviors exhibit selectivity toward specific wavelengths of light. Under the 810 nm excitation, the heterojunction displays a negative response in the hole‐dominant region, while it shows a positive response in the electron‐dominant region with an increase in illumination intensity (Figure 2e and Figure S12). Given the significant difference in the absorption coefficients at 365 and 810 nm, the illumination intensities for the two wavelengths are separately optimized to obtain clear photoresponse characteristics. Further, control experiments are performed under identical illumination conditions (Figure S13). Consistent with the previous observations, negative and positive shifts in the threshold voltage are observed for the 810 and 365 nm illumination, respectively. These results demonstrate that the bidirectional photoresponse, i.e., the opposing direction of the threshold voltage shifting under 365 nm versus that under 810 nm illumination, is a wavelength‐selective phenomenon that arises from the distinct carrier dynamics in the heterojunction rather than from an intensity‐dependent effect. Furthermore, single‐component devices (pure C8‐BTBT 2DMC or TFT‐CN 2DMC) exclusively exhibit unidirectional positive photocurrent modulation behaviors (Figure S14), underscoring the importance of the unique structural and electronic properties of the heterojunction in enabling bidirectional photoresponse functionality. This finding suggests that the observed positive and negative photoconductive effects can be effectively regulated by wavelength and gate voltage, underscoring the promise of 2DMC p–n heterojunctions for advancing bidirectional optoelectronic synapse simulations.

As pivotal metrics for evaluating the performance of phototransistors, photosensitivity P and photoresponsivity R are quantitatively defined as P = (I _light − I _dark)/I _dark and R = (I _light − I _dark)/P _inc S. In these equations, I _light and I _dark represent the source‐drain currents measured under illuminated and dark conditions, respectively; P _inc represents the incident light intensity; and S represents the illuminated channel area. To demonstrate the best achievable performance, the evolutions in P and R across different incident wavelengths as functions of light intensity under V _GS of 0, −60, and +60 V are illustrated in Figure 2f–j and in Figures S15 and S16. Under the irradiation of UV light, the p values attain the maxima of 4969.3 at a light intensity of 73.5 µW cm⁻² (V _GS = 0 V), whereas the highest R value of 6.3 × 10⁴ A W⁻¹ is observed under the weak illumination of 5.2 µW cm⁻² (V _GS = 60 V). For comparison, P and R reach peak values of 9.8 (V _GS = 0 V, 2757.8 µW cm⁻²) and 448.3 A W⁻¹ (V _GS = −60 V, 213.6 µW cm⁻²), respectively, under 810 nm light illumination. Figure 2h presents a comparative analysis of the performance of our device against those of previously documented high‐performance phototransistors [31, 32, 33, 34, 35]. Our p–n heterojunction OSTs demonstrated exceptional photosensitivity and responsivity, and this can be attributed to the high‐quality 2DMC heterojunction.

We statistically analyzed the electrical and optoelectronic performance across 25 independently fabricated devices to evaluate the reproducibility of the heterojunction devices. As summarized in Figure S17 and Table S1, both mobility and responsivity show slight fluctuations. The consistently small standard deviations (SD) and relative variation (RV, defined as SD/mean) across these key performance metrics confirm the highly reproducible device performance, demonstrating the reliability of the 2DMC heterojunction approach.

2.2. Wavelength‐Selective Bidirectional Optoelectronic Synapses Based on 2DMC Heterojunctions

In biological neural networks, the release of excitatory or inhibitory neurotransmitters from the presynaptic terminal to the postsynaptic receptor results in observable behaviors characterized by excitatory postsynaptic currents (EPSC) or inhibitory postsynaptic currents (IPSC) (Figure 3a). Leveraging the aforementioned bipolar nature, the 2DMC p–n heterojunction is adeptly utilized for emulating analogous bidirectional synaptic behaviors. Applying a light pulse with a tunable wavelength to the synaptic device enables control over the PPC and NPC channels between the source and drain electrodes, corresponding to EPSC and IPSC, respectively. The channel conductance serves as a reflection of fluctuations in synaptic weights that signify the connection strength between two synaptic neurons, whereas the dynamic modulation of synaptic weights, i.e., synaptic plasticity, is intricately linked to learning and memory within the human brain. The transfer curves (Figure 2e) show that pronounced PPC and NPC effects are observed at V _GS = +60 V, which produce the expected intriguing EPSC and IPSC behaviors elicited by optical stimuli with wavelengths of 810 and 365 nm, respectively (Figure 3b). The observed persistent photocurrent poststimulation relevant to memory functions can be understood through the sluggish recombination kinetics of photoinduced carriers at the interface between the 2DMC and dielectric layer. Hysteresis loops observed under double‐sweep measurements further corroborate the charge‐trapping mechanism at the semiconductor/dielectric interface (Figure S18), wherein the enhanced hysteresis window under illumination reflects the effective capture of photogenerated carriers, underpinning the photocurrent relaxation dynamics essential for synaptic plasticity [35]. C8‐BTBT or TFT‐CN OSTs exhibit exclusively excitatory synaptic behavior [35, 36], underscoring the significance of the unique structural and electronic properties inherent to 2DMC p–n heterojunction configurations in enabling bidirectional synaptic functionality, which helps enhance the potential for advanced neuromorphic applications.

FIGURE 3.

Bidirectional optoelectronic synaptic behaviors of 2DMC heterojunction‐based OST device. (a) Schematic of multiwavelength synaptic stimulation in neural systems. (b) EPSC and IPSC responses under 810 and 365 nm illumination. (V _GS: 60 V; V _DS: 0.1 V; pulse duration: 1 s; light power: 1.37 mW/cm² (810 nm) and 6.5 µW/cm² (365 nm)). (c) PPF index versus time interval for optical spikes at 810 nm (top) and 365 nm (bottom) wavelengths. (V _GS: 60 V; V _DS: 0.1 V; pulse duration: 1 s; light power: 1.37 mW/cm² (810 nm) and 6.5 µW/cm² (365 nm)). (d) Pulse duration‐dependent and (f) pulse number‐dependent EPSC behaviors under 810 nm illumination. (V _GS: 60 V; V _DS: 0.1 V; light power: (d) 1.37 mW/cm²and (f) 2.66 mW/cm²) (h) Gate‐voltage‐regulated EPSC and IPSC behaviors under 810 nm illumination with different light intensities. (V _DS: 0.1 V; pulse duration: 1 s). (e) Pulse duration‐dependent and (g) pulse number‐dependent EPSC behaviors under 365 nm illumination. (V _GS: 60 V; V _DS: 0.1 V; light power: (e) 6.5 µW/cm² and (g) 9.9 µW/cm²) (i) Gate‐voltage‐regulated EPSC and IPSC behaviors under 365 nm illumination with different light intensities. (V _DS: 0.1 V; pulse duration: 1 s).

We comprehensively investigated the synaptic plasticity and dynamic modulation of 2DMC p–n heterojunction OST devices under photonic control (PPC mode with 810 nm excitation and NPC mode with 365 nm excitation). PPF is pivotal for the temporal encoding and decoding of visual stimuli, manifesting through successive presynaptic activations that can amplify postsynaptic responses. This phenomenon reflects the excitatory and inhibitory effects of dual stimulation. Our heterojunction OST device replicates PPF characteristics by delivering two sequential light pulses. Analogous to a biological synapse, a shorter interval (Δt) between the two consecutive light pulses results in a more pronounced augmentation of EPSC or a greater suppression of IPSC, as indicated in the inset of Figure 3c. We define the PPF index as PPF index = (A₂/A₁) × 100% to quantify PPF behavior. As illustrated in Figure 3c, the PPF indices for EPSC or IPSC behaviors demonstrate an exponential decay with increasing Δt, with the decay profile effectively fitted by a general learning model of neuron behaviors ($\mathrm{PPF}\mathrm{index}=C_1{e}^{\frac{-\mathrm{\Delta }t}{\tau 1}}+C_2{e}^{\frac{-\mathrm{\Delta }t}{\tau 2}}+C$). In addition, the stimuli strength‐dependent synaptic response can be effectively emulated by modulating either the intensity or duration of light pulses within the proposed 2DMC p–n heterojunction OST devices. As illustrated in Figure 3d and Figures S19 and S20, the enhanced photoresponse elicited by increasing the intensity or duration of the 810 nm light pulse augments synaptic weight, manifesting as pronounced EPSC peaks and elevated persistent EPSC levels poststimulation. In contrast, 365 nm light stimulation produces an inverse relationship, where increased light intensities or extended pulse durations lead to the progressive attenuation of both IPSC peak magnitudes and poststimulation retention currents (Figure 3e and Figures S21 and S22). These phenomena align with the STP to LTP transition functions observed in biological systems. The STP–LTP transition is achieved through iterative training in biological synapses, which can be substantiated by modulating the number of light pulses applied to the synaptic device. An increase in the number of light pulses augments the ΔPSC and longer decay time of EPSC or IPSC; this can be attributed to the memory effect inherent in the synaptic device, as demonstrated in Figure 3f,g. Moreover, the bidirectional synaptic plasticity of our devices can be precisely controlled via gate voltage modulation. The EPSC and IPSC responses at V _GS of +60 and −20 V are presented as illustrative examples because the transfer curves (Figure 2e) indicate that the device exhibits pronounced bidirectional photoresponses under these two biases. Under 810 nm illumination (Figure 3h and Figure S23), the OST device exhibits EPSC behavior at a V _GS of 60 V, with EPSC amplitudes increasing with an increase in light intensity. Conversely, at a V _GS of −20 V, the device switches to IPSC behavior, suppressing PSC peaks with an increase in light intensity. This trend is reversed under 365 nm illumination (Figure 3i and Figure S24), wherein a V _GS of 60 V induces IPSC behavior, while −20 V elicits EPSC behavior, thereby highlighting the gate‐tunable, wavelength‐dependent synaptic response. Bidirectional synaptic responses are not limited to the specific gate voltages of +60 or −20 V. As shown in the transfer curves (Figure 2e), stable positive or negative responses can be achieved over continuous ranges of both positive and negative gate voltages. For example, Figure S25 illustrates the IPSC and EPSC responses to 810 and 365 nm light under a series of negative gate voltages. Thus, bidirectional responses can be realized over a relatively wide voltage window rather than only at discrete voltage points. These findings demonstrate a high‐performance full‐spectrum optoelectronic artificial synapse with dual‐mode bidirectional tunability, which offers promising potential for advanced neuromorphic computing applications.

A total of 60 consecutive excitatory (810 nm) and 60 inhibitory (365 nm) light stimuli are applied to demonstrate the potential of the device for application to artificial neural networks (Figure S26). According to the formula in Note S2, the nonlinearity values for long‐term‐potentiation and long‐term‐depression weight updates were calculated as 0.035 and −0.061, respectively, and the dynamic range (G _max/G _min) was calculated to be 9.24 dB, demonstrating good linearity and a large dynamic range. Further, the energy consumption of the heterostructured synaptic device was evaluated under low operating voltages. As shown in Figure S27, typical EPSC behavior could be completely imitated even at a low V _DS of 0.001 V and pulse duration of 0.1 s. Based on the formula E  = V_DS   × I_DS × t, the energy consumption was determined to be as low as 18.4 fJ per event, reflecting the extremely low energy operation. Furthermore, we evaluated the cycling stability by applying consecutive alternating 810 and 365 nm optical pulses (Figure S28). The device exhibited consistent synaptic responses across multiple cycles, demonstrating reliable synaptic modulation under repeated operation and prolonged light exposure. This stability suggests the potential of this device for neuromorphic applications. We summarize the key metrics of our synaptic device and compared them with those of bipolar synaptic transistors reported in the literature to provide a comprehensive performance assessment (Table S2). The results revealed that our 2DMC heterojunction device exhibited competitive performance, highlighting the advantages of our 2DMC heterojunction strategy for full‐spectrum bidirectional optoelectronic synaptic applications.

Further, we extended our characterization to the visible range to evaluate the full‐spectrum bidirectional optoelectronic synaptic behaviors of the device comprehensively. Figures S29 and S30 present the transfer characteristics of the device under visible wavelengths (565 and 660 nm) along with the corresponding evolution curves of P and R as functions of V _GS and illumination intensity. Under these two visible wavelengths, an increase in illumination intensity results in negative and positive photocurrent responses in the hole‐ and electron‐dominant regions, respectively. Furthermore, bidirectional optoelectronic synaptic behavior is also demonstrated under visible illumination, where both EPSC and IPSC behaviors can be triggered by modulating the gate bias polarity (Figures S31 and S32). These results demonstrate bidirectional optoelectronic behavior across the UV–Vis–NIR range, confirming the full‐spectrum discrimination and bidirectional synaptic functionality of the device.

2.3. Working Principle of Heterojunction OSTs

Figure 4 illustrates the operational mechanism of the device. Finite‐element semiconductor physics simulations of the device were performed using COMSOL software. In this simulation, V _GS = ±60 V were selected as the primary bias conditions for the analysis to clarify the wavelength‐switching mechanism and demonstrate that simply reversing gate bias polarity can enable bidirectional responses. A model was constructed based on the physical geometry of the device, and the internal potential distribution was calculated under the conditions of V _GS = −60 V and V _DS = 0.1 V, as indicated in Figure 4a. Under these conditions and driven by the electric field, holes in the p‐type region migrated toward the SiO₂/p‐type interface, whereas minority carrier holes in the n‐type region moved toward the n‐type/p‐type interface. Similarly, the internal potential distribution is calculated under V _GS = +60 V and V _DS = 0.1 V, as depicted in Figure 4b. In this scenario, the electrons in the p‐type region moved toward the SiO₂/p‐type interface, whereas the majority of carrier electrons in the n‐type region drifted toward the n‐type/p‐type interface. Figure 4c presents the 1D potential distribution along the white dashed line in Figure 4b, clearly illustrating the influence of V _GS on the potential distribution of the device. In the V _GS<0 regime, holes in the n‐type region drifted toward the p‐n interface, whereas in the V _GS>0 regime, electrons in the n‐type region drifted toward the p‐n interface. This behavior was closely related to the positive and negative photoresponse mechanisms of the device.

FIGURE 4.

Mechanism analysis of the device. (a,b) Finite‐element semiconductor physics simulations of the device were performed using COMSOL software under V _GS = −60 V and V _GS = +60 V. (c) 1D potential distribution along the white dashed line in (b). (d) Relative electron concentration distribution under 810 nm (above) and 365 nm (below) illumination. (e) Thin‐film surface potential under different illumination conditions measured via KPFM. (f) Statistical distribution of surface potential under varying illumination conditions. (g) Schematic of the positive photoresponse mechanism under 810 nm illumination. (h) Schematic of the negative photoresponse mechanism under 365 nm illumination. The black, white, and yellow arrows indicate the direction of carrier movement caused by gate voltage, concentration, and built‐in electric field, respectively.

The electron density distribution is calculated under V _GS>0 with 810 nm and 365 nm illumination to quantitatively analyze the relationship between carrier transport and the positive/negative photoresponse of the device; the results are shown in Figure 4d. The n‐type semiconductor (upper region) exhibited a relatively higher absorption at 810 nm, and therefore, most of the 810 nm optical pulse was absorbed in this region, generating a large number of photogenerated electron–hole pairs. This resulted in a significantly higher electron concentration in the n‐type region, leading to enhanced conductivity in the n‐type conduction channel. Conversely, the p‐type semiconductor (below this region) exhibited stronger absorption at 365 nm. When illuminated with a 365 nm pulse, most of the light was absorbed in the p‐type region. Under the influence of the V _GS electric field, the photogenerated holes drifted toward the p‐n interface and recombined with the majority carriers (electrons) in the n‐type region, reducing the electron concentration (indicated by the red dashed box). This process resulted in a decreased conductivity in the n‐type conduction channel. Kelvin probe force microscopy (KPFM) was employed to analyze the carrier transport mechanism by measuring the contact potential difference (CPD) between the tip and sample. The recorded surface potential follows CPD = (ϕ _tip − ϕ _sample)/e, where ϕ _tip and ϕ _sample represent the fixed work function of the tip and work function of sample, respectively. As indicated in Figure 4e,f, the dark‐state CPD was ∼100 mV. Under 810 nm illumination, CPD increases to ∼400 mV, suggesting a decrease in the ϕ _sample. A decreased work function indicates that the Fermi level shifts toward the conduction band minimum, thereby corresponding to an increased surface electron concentration, i.e., electron accumulation in the top n‐type semiconductor layer. Conversely, under 365 nm illumination, CPD drops to ∼–600 mV, suggesting an increase in the ϕ _sample. This implies that the Fermi level shifts toward the valence band maximum, which corresponds to a decreased surface electron concentration, i.e., electron depletion in the top n‐type semiconductor layer. These observations are consistent with the finite element simulation results.

The positive and negative photoresponse mechanisms of the device were elucidated based on the theoretical calculations and experimental analyses. The transfer characteristic curve revealed that, under dark conditions, the device exhibited minimal current at V _GS = 0 V, indicating a balanced electron–hole concentration in the dual‐conduction channels (electron‐dominated and hole‐dominated) of the organic heterojunction‐based device. As illustrated in Figure 4g, the n‐type organic semiconductor TFT‐CN absorbed most of the incident light because of its higher absorption coefficient upon illumination with an 810 nm optical pulse, thereby generating a substantial number of photogenerated electrons and holes. Under the influence of V _GS, photogenerated electrons drift toward the p‐n interface, while holes drift toward the top of the n‐type layer. At the interface, the increased photogenerated electron concentration facilitated the diffusion of electrons from the n‐type layer into the p‐type layer. However, the built‐in‐field‐driven drift of electrons toward the n‐side opposes this diffusion, retaining a portion of photogenerated electrons within the n‐type region. This increases the electron concentration in the conduction channel and yields a positive photoresponse. Under 365 nm illumination (Figure 4h), the p‐type organic semiconductor C8‐BTBT becomes the primary light absorber, generating a high density of photogenerated electron–hole pairs. The gate electric field drives photogenerated holes toward the p‐n interface and electrons away from the interface. Although the built‐in field simultaneously drove electrons and holes toward the n and p sides within the depletion region, respectively, the substantial hole concentration accumulated at the interface, enabling a considerable number of holes to diffuse into the n‐type region. These diffused holes then recombined with electrons in the TFT‐CN conduction channel. This recombination reduced electron density and output current, resulting in a negative photoresponse. This comprehensive analysis confirmed that the bidirectional photoresponse mechanism of the device was governed by carrier transport dynamics under different illumination wavelengths and bias conditions.

2.4. Broad‐Spectrum Vehicle‐Mounted Intelligent Vision System Based on Heterojunction OSTs

With the advancement of artificial intelligence technology, intelligent driving systems have significantly enhanced autonomous vehicle decision‐making capabilities through advanced optoelectronic sensing devices and deep‐learning algorithms. This has substantially improved traffic safety and made it a research and commercial focus in the digital information era. However, traditional vehicle‐mounted‐assisted driving systems that rely on RGB visible light imaging exhibit degraded recognition performance or target misjudgment under complex environmental conditions such as low‐light nighttime scenarios, rain, fog, or dust, potentially leading to severe traffic accidents.

In this study, we designed a broad‐spectrum vehicle‐mounted intelligent vision system based on our OST devices with UV‐to‐IR sensing capability, which enabled switching between different spectral responses through V _GS polarity and spectrum. As shown in Figure 5a, urban road asphalt containing aromatic hydrocarbons demonstrated significantly higher UVA (315–400 nm) absorption compared to that of surrounding environment (e.g., soil and vegetation). The sensitivity of the device to ultraviolet light enables capturing reflection differences between road surfaces and backgrounds, thereby generating high‐contrast boundary features for lane‐marking identification. Meanwhile, near‐infrared light with superior penetration capability facilitates the detection of moving objects such as pedestrians and vehicles. The device exhibits short‐term photoresponse memory characteristics (Figure 5b), functioning as an event‐driven optoelectronic sensing unit that activates event addresses based on photocurrent decay to generate event streams for object trajectory marking. Figure 5c illustrates the time‐dependent conductance states under single‐pulse illumination, wherein spatiotemporal object information is encoded through interframe timestamp coupling. When an object moves from pixel address 1→4→13 (Figure 5d), the frame‐difference algorithm triggers event address generation upon detecting current variation exceeding the threshold (O₁₃(F_n‐1) − O₁₃(F_n) < 0). Infrared imaging (Figure 5e) distinguishes between moving (white) and static (black) objects through event‐address labeling. The device serves as a weighting matrix for preprocessing input images by utilizing V _GS‐polarity‐tunable spectral responsivity as weighting coefficients (Figure 5f), and it extracts three 224 × 224 feature channels: ultraviolet, infrared, and dynamic event channels. After channel‐wise convolution and feature fusion, the convolutional neural network achieves >85% training accuracy (Figure 5g,h) with dynamic vision enhancement, demonstrating a classification accuracy of >90% for five typical traffic objects (Figure 5i).

FIGURE 5.

Device‐based broadband in‐vehicle intelligent vision system. (a) Schematic of device‐based broadband in‐vehicle intelligent vision system. (b) Short‐term optoelectronic memory capability reflecting the temporal and spatial states of the object through time stamps. (c) Conductance states of the device at different time stamps after receiving a single optical pulse input. (d) Event address of the motion trajectory of an object can be obtained from the decay direction of the short‐term memory photocurrent. Moving objects can be marked in an event‐driven manner by combining the frame‐difference method. (e) Under the infrared feature channel, a moving electric motorcycle and stationary car are distinguished as event and nonevent, respectively, with the event address of the moving object marked successfully. (f) Wavelength‐selective dependency enables the spectral feature channel extraction of input images without optical filters via the adjustment of the V _GS. These extracted features serve as inputs to our designed multimodal convolutional neural network (CNN) for traffic object classification. (g) The training and (h) inference accuracies of the CNN. (i) Classification accuracy for five typical traffic objects.

3. Conclusions

We developed a full‐spectrum bidirectional optoelectronic synaptic device based on a 2DMC p‐n heterojunction. Unlike conventional neuromorphic devices that rely on either wavelength‐selective or electrical‐polarity‐dependent control, our device integrates both strategies to achieve bidirectional synaptic plasticity. The precisely engineered 2DMC heterojunction featuring highly ordered C8‐BTBT and TFT‐CN crystalline structures demonstrated exceptional photoresponsivity reaching 6.3 × 10⁴ A W⁻¹. This bipolar transistor architecture successfully emulated fundamental neurosynaptic behaviors, including PPF, STP, LTP, and bidirectional synaptic weight modulation. A broad‐spectrum vehicle‐mounted intelligent vision system was developed based on this device. This system performed real‐time traffic object recognition with an accuracy of >90% under complex environmental conditions. These results not only advance our fundamental understanding of artificial synaptic devices but also pave the way for next‐generation intelligent systems in autonomous vehicles, adaptive robotics, and edge computing applications.

While prior demonstrations of bidirectional optoelectronic synapses relied on either wavelength‐dependent photoresponse in carefully engineered multiabsorption materials or gate‐voltage polarity switching in transistor structures, this work uniquely integrated both control mechanisms (optical wavelength and gate voltage polarity) within a single field‐effect transistor based on high‐quality 2DMC heterojunctions, extending the operable range to UV–NIR while maintaining high responsivity. This dual‐mode design enables the device to flexibly switch between optical and electrical pathways and perform synergistic programming, offering energy efficiency in dynamic visual processing tasks and providing a versatile, reconfigurable hardware foundation for future neuromorphic visual processing. Thus, our work distinguishes itself from existing studies in terms of the single‐crystal heterojunction design, dual‐mode bidirectional functionality, and broad‐spectrum high‐responsivity performance.

Of course, we must acknowledge that the devices used in this work are still in the laboratory stage. Several limitations need to be addressed before practical applications can be realized:

1. The relatively high operating voltages pose a challenge for low‐power integration. Although bidirectional modulation has already been achieved at V _GS as low as −3 V, further voltage reduction is necessary. Possible strategies include employing high‐κ dielectrics (e.g., HfO_x or AlO_x) to increase the gate capacitance per unit area and optimizing the electrode–semiconductor interface to minimize contact resistance and trap states.

2. This study focuses on single‐device characterization, whereas practical systems require large‐scale device arrays. Achieving a uniform performance across an array presents additional challenges in terms of device fabrication, yield, and crosstalk suppression. Future studies should explore scalable fabrication techniques and array‐compatible device designs.

3. The combination of hardware experiments with simulation applications inevitably involves a degree of idealization. In practical applications, it is necessary to consider automotive‐grade standards for readout digital circuits, design specifications of ASIC architectures, and requirements of a complete circuit system. Addressing these challenges constitutes an important direction for the subsequent in‐depth exploration of organic integrated circuit systems.

4. Experimental Section

4.1. Materials

C8‐BTBT and TFT‐CN were purchased from Sigma‐Aldrich and used without further purification.

4.2. Preparation of 2DMC Heterojunction Device

C8‐BTBT and TFT‐CN 2DMCs were independently prepared using a controlled liquid‐phase substrate self‐assembly method. Droplets of C8‐BTBT (1 mg mL⁻¹ in chlorobenzene) or TFT‐CN (0.5 mg mL⁻¹ in chlorobenzene) solutions were deposited onto the glycerol surface within clean weighing bottles. Subsequently, crystal growth apparatuses were transferred to an oven maintained at 25°C with 30% humidity, facilitating controlled solvent evaporation. This optimized process yielded large‐area, ultrathin 2DMCs of self‐assembled C8‐BTBT or TFT‐CN crystals formed at the air–glycerol interface upon complete solvent evaporation. Prior to heterojunction fabrication, the SiO₂ (300 nm)/Si substrate was successively ultrasonically cleaned in acetone, isopropanol, ethanol, and deionized water and dried under nitrogen flow. The as‐grown C8‐BTBT crystal was initially transferred onto an SiO₂/Si substrate via an inverted dipping technique, wherein the substrate was carefully immersed and withdrawn to capture the floating crystal, followed by thorough rinsing with deionized water to ensure surface purity. Subsequently, the TFT‐CN crystal was transferred onto a C8‐BTBT‐coated substrate using the same methodology to form a well‐defined p‐n heterojunction. This layer‐by‐layer transfer process ensured crystal integrity and optimal interfacial contact between the organic semiconductor layers. Finally, a pair of source‐drain electrodes composed of Ag (80 nm)/Au (80 nm) bilayers (with Au deposited as the top layer to prevent the oxidation of the underlying Ag) were sequentially transferred onto the surface of the TFT‐CN 2DMC, thereby completing the fabrication of the 2DMC heterojunction device.

4.3. Instrumentation and Characterization

The OM images were obtained using an optical microscope integrated with a charge‐coupled device camera (Vision Engineering Co., UK). AFM and KPFM measurements were performed using a Bruker Dimension Icon system. For structural analysis, TEM and SAED were conducted on a Tecnai G2 F20 S‐TWIN microscope. The XRD patterns were acquired using a DX‐2700B diffractometer with Cu‐K𝛼 radiation (λ = 0.15406 nm). GIWAXS studies were carried out on a XEUSS SAXS/WAXS system at an incidence angle of 0.2°. The optical absorption spectra were measured with an Agilent Cary 7000 UV–Vis–NIR spectrophotometer. For optoelectronic characterization, the device performance was evaluated using a Keithley 4200 semiconductor parameter analyzer integrated with a Lake Shore manual probe station under ambient conditions. The illumination system employed 365 nm and 810 nm light‐emitting diodes, with irradiance power precisely controlled by a Thorlabs DC2200 LED driver.

4.4. Performance Calculation

Field‐effect mobility µ was extracted from the saturation regime and calculated using the transistor equation I _DS = (W/2L) × C _i × µ × (V _GS − V _th) [2], where I _DS, C _i, V _GS, V _th, W, and L represent the drain current, dielectric capacitance (10 nF cm–²), gate voltage, threshold voltage, channel width, and channel length, respectively.

Conflicts of Interest

The authors declare no conflict of interest.

Supporting information

Supporting File: adma73698‐sup‐0001‐SuppMat.docx.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request.