High-linearity all-2D heterojunction-based neuromorphic image-sensing transistor array with a sandwiched tunneling layer for dynamic vision Authors: Yiming Yuan, Yinxing Zhang, Jie Wang, Tianci Wang, Ruotao Yu, Zhipeng Xu, Jialiang Xu, Wentao Xu Published: 2026 ABSTRACT AbstractNeuromorphic visual systems have attracted tremendous attention for their in-sensor image sensing and information preprocessing paradigm that replicates human visual system. However, current neuromorphic visual devices employed in these systems still confront challenges on lateral conductivity, air-stability, and photo-response linearity. Here, we demonstrate a neuromorphic image-sensing transistor array based on a tri-layer van der Waals heterostructure composed of MoS2, (BA)2PbI4 and graphdiyne (GDY). The top MoS2 layer with high intrinsic mobility serves as carrier-transport channel, while encapsulating the 2D perovskite to shield it from ambient air. The sandwiched 2D-(BA)2PbI4 nanosheet converts light into electricity, and functions as a photo-tunable tunneling junction, facilitating unidirectional tunneling of photo-generated carriers toward the underlying GDY layer. GDY, featuring well-distributed alkyne bonds, enables uniform interaction and storage of charges. The device demonstrates long-term air-stability (>10 weeks), highly competitive linearity in synaptic transistor weight modulation (αP = 0.006, αD = 0.025), and remarkably low energy consumption (~41.8 aJ per synaptic event) among 2D heterojunction synaptic transistors. Furthermore, a hardware-level neuromorphic kernel constructed using a synaptic array of these heterostructure devices realizes spatiotemporal neuromorphic in-sensor dynamic image processing and trajectory detection for object movement. This work provides hardware architecture for the next-generation neuromorphic vision.Subject terms: Two-dimensional materials, Electronic devices Introduction Neuromorphic visual systems integrate image sensing, preprocessing, and memory functionalities1–4, thus replicating the energy-efficient operational mechanism of the human visual systems5–8. A wide range of material systems have been explored to build high-performance neuromorphic visual sensors9–12. Among them, two-dimensional (2D) Ruddlesden-Popper (RP) perovskites, such as (BA)2PbI4 (BA = n-butylamine), possess a stable layered structure with densely packed organic components13,14. Nevertheless, neuromorphic image sensors based on 2D perovskite materials still desire improvements in lateral conductivity, long-term air-stability and light-sensitive linearity. Transition metal dichalcogenides (TMDCs), a class of materials with high carrier mobility (e.g. MoS2), can enhance channel conduction. When stacked with 2D perovskite nanosheets, they form van der Waals (vdW) heterojunctions that are less affected by lattice-matching constraints15,16. Specifically, MoS2 nanosheets exhibit an optical absorption bandgap matched to that of 2D-(BA)2PbI4 perovskites. Even with these improvements, such heterojunction devices still struggle to achieve precise control of synaptic weight updates. Most designs lack uniform sites to regulate carriers. This leads to inconsistent synaptic behavior and unwanted interfacial carrier recombination that drives up energy use. Meanwhile, graphdiyne (GDY) is an emerging 2D carbon allotrope, featuring sp and sp2 hybridization and a highly conjugated π-extended structure17–19. The triangular nanopores in GDY can serve as efficient migration pathways for ions, while its abundant alkyne bonds provide numerous active sites for interaction with charges20–23, benefiting to controllable synaptic weight modulation for reliable neuromorphic in-sensor computing. 2D-layered nanosheets enable high interfacial quality via defect-free stacking in vdW heterojunctions24–26. This clean interface minimizes unwanted trap states and carrier scattering at the material boundary. Consequently, carrier migration within vdW heterostructures is weakly affected by interfacial scattering, thereby enhancing energy efficiency. Furthermore, photo-tunable carriers in vdW heterojunctions undergo unidirectional tunneling along the electric field, which suppresses carrier recombination and reduces energy consumption27–29. Herein, we demonstrate a neuromorphic image-sensing transistor based on a tri-layer vdW heterostructure, comprising a carrier transport layer (MoS2), a light-sensitive layer ((BA)2PbI4), and a charge reservoir layer (GDY). The sandwiched 2D-(BA)2PbI4 nanosheet converts light into electricity and facilitates hole tunneling from the conductive channel to the GDY layer. The overlying MoS2 layer enables lateral transport of the generated carriers while encapsulates the 2D perovskite nanosheet to protect it from ambient air. The underlying GDY layer stores and interacts with injected charges on the abundant alkyne bonds to form controllable memory. Such a synergistic integration of GDY’s charge storage capability with the photosensitivity and tunneling effects of the 2D perovskite enables precise regulation of charge storage, thereby realizing high-linearity synaptic behavior. The device exhibits long-term air-stability (>10 weeks), the highly competitive linearity of synaptic transistor weight modulation (αP = 0.006, αD = 0.025), and the ultra-low energy consumption (~41.8 aJ per synaptic event) among 2D heterojunction synaptic transistors. These superior characteristics enable efficient neuromorphic spatiotemporal dynamic image sensing and processing in a device array. The human action classification and object movement direction could be successfully detected and recognized by the GDY hardware-level neuromorphic kernel. This work thus paves a promising pathway for the development of neuromorphic vision systems with reliable in-sensor computing. Results Heterostructure structural characterization The proposed van der Waals heterojunction transistor mimics the human visual pathway for image perception and preprocessing. Benefiting from GDY’s high-linearity synaptic modulation, the GDY-based tri-layer vdW heterojunction neuromorphic image-sensing transistor (GDY-NIST) adopts a bio-inspired in-sensor sensing and processing architecture to build a neuromorphic synaptic convolution kernel, achieving high-precision dynamic image processing and motion direction detection30–32 (Fig. 1a and Fig. 1b). The tri-layer synaptic transistor is constituted of the 2D MoS2/(BA)2PbI4/GDY vdW heterostructures (Fig. 1c). MoS2 serves as the channel layer, with strong covalent bonds connecting within the layer and weak vdW forces linking between layers (Supplementary Fig. 1a). (BA)2PbI4 is employed as a photosensitive tunneling layer to reduce the device’s dark current and lower its energy consumption. With hydrophobic organic long chains, (BA)2PbI4 belongs to the 2D RP-type perovskite (Supplementary Fig. 1b), which can be synthesized via a cooling crystallization method33 (Supplementary Fig. 2). Typical (00l) planes are detected via X-ray diffraction (XRD) characterization of (BA)2PbI4 single-crystal powders, consistent with its layered structure34 (Supplementary Fig. 3). GDY is formed by a 2D planar network structure with benzene rings conjugated by 1,3-diyne bonds, and the porous structure provides abundant storage sites for carrier transport35–37 (Supplementary Fig. 1c). Conception and characterization of the GDY-based tri-layer vdW heterojunction neuromorphic image-sensing transistor. a Schematic of neuromorphic photoelectric transistor structure. b Schematic of bio-inspired visual mechanism, and high-linearity synaptic convolution kernel-enabled dynamic motion direction detection. c Optical microscope image and structure diagram of the transistor. The scale bar is 10 μm. d Thickness characterization by AFM. The scale bar is 2 μm. Raman spectra of e MoS2 and f GDY. g XPS spectrum of C1s of GDY. For detailed structural and interfacial characterization of the heterostructure, atomic force microscopy (AFM) has been used to measure the thicknesses of the individual layers, to be 25.9 nm (MoS2), 29.2 nm ((BA)2PbI4), and 5.3 nm (GDY), respectively (Fig. 1d). The characteristic peaks at 407 cm–1 and 383 cm–1, corresponding to the A1g and E2g1 modes of MoS2, have been clearly demonstrated in the Raman spectra38 (Fig. 1e). Meanwhile, the intensity of G band (1565 cm–1) of GDY is significantly higher than that of D band (1386 cm–1), indicating the high crystallinity and low defect density of the as-synthesized GDY sample (Fig. 1f). The remaining peaks (1939 and 2179 cm–1) are assigned to stretching vibrations of the diyne linkage in GDY18. Furthermore, the prepared GDY film is verified to have high thickness uniformity, a well-defined sp/sp2 hybridized single-phase structure, and low defect density via AFM and Raman mapping characterizations (Supplementary Fig. 4 and Fig. 5). In parallel, the peak of C1s orbital is observed in X-ray photoelectron spectroscopy (XPS), which could be fitted into four subpeaks corresponding to the C = C bond (284.6 eV), C ≡ C bond (285.3 eV), C − O bond (286.7 eV) and C = O bond (288.7 eV), respectively (Fig. 1g)39. The high polymerization degree of GDY is further verified by Fourier transform infrared spectrometer (FT-IR) results, where in-plane conjugated alkyne linkages and negligible interlayer reactive sites ensure dominant vdW interlayer interactions (Supplementary Fig. 6). Finally, the band alignment at the interface of MoS2, (BA)2PbI4, and GDY has been calculated by the ultraviolet photoelectron spectroscopy (UPS) measurement and the absorption spectra (Supplementary Fig. 7 and Fig. 8). Electrical properties and graphdiyne charge modulation To study the electrical properties of the GDY-NIST device, transfer curves have been demonstrated at different drain-source voltage (Vds). The device exhibits typical n-type semiconductor characteristics, achieving an on-state current (Ion)/off-state current (Ioff) ratio of approximately 105 at a bias of 1.0 V (Fig. 2a, b). The obvious hysteresis observed in the transfer curves is confirmed to reflect the potential for modeling biological synaptic plasticity. Additionally, it is verified by the transfer curves obtained at different light intensities that the device can harvest photo-generated carriers under illumination, leading to more electrons produced in the channel region40 (Supplementary Fig. 9). Electrical characterization and mechanism of the photoelectric synaptic transistor. a Linear-scale and b log-scale transfer characteristics of the GDY-NIST device. c Cycling stability of device’s electrical properties. d Output characteristics of the device. e Transfer characteristics of the device controlled by different Vgs sweeping ranges. f Extracted memory window as a function of maximum Vgs sweeping ranges. g Mechanism of light/electric modulation in the GDY-based device. Notably, the cycling stability of the device transfer characteristics has been evaluated under Vds = 1.0 V, and stable hysteresis loops could be maintained during 60 consecutive sweeps (Fig. 2c). It is revealed by quantitative analysis that the small cycle-to-cycle variation arises from reversible charge attraction and release in the heterojunction, with negligible impact on device stability (Supplementary Fig. 10). The field-effect mobility extracted from the transfer curve is determined to be 5.98 cm2 · V−1 · s−1, confirming its efficient carrier-transport capability (Supplementary Fig. 11). Meanwhile, the predictable linear transport behavior between MoS2 and Au electrodes at Vds of 0-0.9 V is displayed by the output characteristics, ensuring efficient carrier transport (Fig. 2d and Supplementary Fig. 12). Current saturation at higher voltages arises from carrier velocity saturation in 2D vdW devices41. By changing the forward sweeping range of gate-source voltage (Vgs) (10-50 V), tunable memory windows are observed in the transfer curves (Fig. 2e). The memory window (ΔVT) is defined as the difference of threshold voltage (Vth) between on and off states, increased linearly with maximum Vgs42 (Fig. 2f). Consequently, these results reveal the effective and uniform attraction of charges by the GDY layer. More specifically, the kink effect in the transfer curves also reveals the GDY modulation layer-dominated, trap filling-induced carrier transport transition. The extracted transconductance (gm) curves indicate that, as Vgs approaches the kink voltage, photo-generated carriers rapidly attracted by GDY’s intrinsic alkyne bonds immediately modulate the local potential, triggering a steep rise of gm (Supplementary Fig. 13). Further increasing Vgs gradually fills residual trap states, boosting channel conductivity and driving continuous rise of gm until saturation. This kink phenomenon, originating from the rapid trap filling and hysteretic detrapping process in the GDY layer, directly governs the synaptic weight modulation. Similar kink effects have been observed in other 2D synaptic transistors43,44, organic electrochemical transistors (OECTs)45, and standard bulk-silicon metal-oxide-semiconductor field-effect transistors (MOSFETs)46. For comparison, the GDY-free control device has been fabricated under the same process (Supplementary Fig. 14a and Fig. 14b). Although the synaptic transistor application potential is demonstrated by the transfer curve hysteresis (Supplementary Fig. 14c), the memory window of 10.79 V is lower than that of the GDY-based tri-layer device (14.29 V) under the same 1.0 V bias, confirming the strong charge storage capability of the GDY layer (Supplementary Fig. 14d). Specifically, the GDY layer plays a crucial role in regulating carrier dynamics in the synaptic device. Under light stimulation, the photocurrent increases in response to the separation of photo-generated electron-hole pairs at the MoS2 and (BA)2PbI4 interface. After removal of illumination, the photocurrent begins to decay due to carrier recombination, but the decay rate slows down as positive charges are attracted by abundant alkyne bonds in the GDY layer22,47. Furthermore, the gate voltage stimulation induces interfacial charge accumulation and accelerate the recombination holes. Consequently, channel carriers are thus rapidly depleted, leading to a gradual decrease in the drain current (Fig. 2g). In contrast, such efficient charge-storage sites are absent in the GDY-free control device, resulting in the rapid generation and recombination of photo-generated carriers. Photoelectric synaptic response and ultra-low energy consumption The modulation of carrier transport by the GDY layer is also verified in light pulse measurements. Under a bias voltage of 0.1 V, the GDY-based and GDY-free devices have been stimulated with a single light pulse (wavelength (λ) = 405 nm, pulse width (T) = 6.25 s, intensity = 0.6 mW∙cm−2) (Fig. 3a). The GDY-based device exhibits quasi-linear growth of the photocurrent and slow relaxation after removal of the light field (Fig. 3b). Conversely, for the GDY-free control device, the photocurrent rapidly saturates and undergoes fast recombination once light is turned off. To make a quantitative comparison, a linear time-dependent photocurrent generation can be analyzed using the following equation48,49: where I1 and I2 are the photocurrent coefficients, t1 and t2 are the time constants, and t is the time for light irradiation. The GDY-based device has been demonstrated a larger photocurrent coefficient ratio (I2/I1) value of 13.77 and a more prolonged time constant (t2 = 67.11 s) (Fig. 3c). In the fitting model of the GDY-based device, Iph,1(t) increases rapidly but quickly reaches saturation. In contrast, Iph,2(t) increases quasi-linearly without saturation. For the GDY-free control device, both the photocurrent coefficient ratio and time constant are relatively low (I2/I1 = 0.95, t2 = 2.19 s). Accordingly, the dominant contribution of Iph,2(t) accounts for the quasi-linearly increasing photocurrent in the GDY-based device. A similar phenomenon occurs under consecutive light pulse stimulation (λ = 405 nm, pulse interval (Δt) = 0.15 s, T = 0.15 s, intensity = 0.6 mW∙cm−2) (Fig. 3d). Synaptic weight is defined as the ratio of the photocurrent (An) generated by the nth light pulse to the photocurrent (A1) generated by the first pulse. A significantly higher synaptic weight (maximum: A10/A1 = 11.65) has been exhibited in the GDY-based device (Fig. 3e). Quasi-linear time-dependent photocurrent behavior and ultra-low power consumption performance. a Post-synaptic currents of GDY-based and GDY-free devices under a single light pulse stimulus. b Modeling of time-dependent photocurrent generation. c Fitted photocurrent ratio (I2/I1) value and time constant (t2). d Synaptic photocurrent behavior of devices under 10 consecutive light pulses. e The ratio of the extracted An/A1. f Schematic illustration of the light-regulated tunneling mechanism. g Photocurrent triggered by a single light spike at Vds = 0.1 mV, with a light intensity of 1.0 μW·cm−2. h Energy consumption as a function of the light pulse duration. i Comparison of energy consumption versus applied Vds of reported 2D heterojunction synaptic transistors. These distinctive photo-response features could be attributed to the uniform attraction and release of photo-generated carriers by the alkyne bonds in GDY, which enable a quasi-linear time-dependent photocurrent generation and delayed photocurrent. Steady-state and time-resolved photoluminescence (PL/TRPL) characterizations have been employed to validate the dominant driving force of conjugated alkyne bonds in GDY for efficient interfacial charge transfer (Supplementary Fig. 15). Spatially uniform structural variations of alkyne bonds, accompanied by homogeneous charge attraction and storage across the GDY film, are revealed by in-situ Raman spectroscopy (Supplementary Fig. 16). Complementarily, the reversible charge interaction and storage enabled by alkyne bonds in GDY under varied bias voltages are verified by the conductive atomic force microscopy (C-AFM) (Supplementary Fig. 17). Such a reversible and uniform carrier modulation is crucial for enhancing the contrast of dynamic image preprocessing in neuromorphic visual systems. The 2D-(BA)2PbI4 nanosheets serve as an important photo-tunable tunneling layer for the regulation of carrier transport. Under illumination, the effective barrier height (φlight) in the valence band is reduced compared to dark conditions (φdark), facilitating quantum tunneling of photo-generated hole carriers (Fig. 3f). To determine the dominant mechanism, the current-voltage characteristics under bias have been analyzed using quantum tunneling models, including the direct tunneling (DT) model and the Fowler-Nordheim (FN) tunneling model28 (Supplementary Fig. 18 and Fig. 19): where d, m*, φ, q, and h are the barrier width, effective mass, barrier height, elementary charge, and Planck constant, respectively. For positive sweeping, the fitted ln(I/V2) vs. 1/V curve reveals linear FN tunneling at high biases and direct tunneling at low biases. The higher barrier height enables the (BA)2PbI4 layer to act as an effective charge blocking barrier, significantly suppressing carrier transport while allowing easier attraction of hole tunneling by the alkyne bonds in the GDY layer. Under illumination, photo-generated carriers are drifted to and accumulated at the MoS2/(BA)2PbI4 interface, inducing a photogate effect that increases channel hole concentration and facilitates hole tunneling into GDY. Accordingly, the (BA)2PbI4 layer blocks minority carrier leakage and suppresses dark current, which endows the GDY-NIST device with remarkable potential for ultra-low energy consumption operation. Minimizing the energy consumption per synaptic event is critical to realize energy-efficient device-level neuromorphic perception50,51. The energy consumption of photoelectric synaptic devices can be calculated by the following formula based on the time integral of instantaneous power52: where Ids is the peak value of photocurrent, Vds is the drain-source bias, and T is the pulse width of light stimulus. The energy consumption of the GDY-NIST device has been demonstrated to be ultra-low, about 41.8 aJ per photonic synaptic event (Fig. 3g and Supplementary Fig. 20). Besides, the energy consumption of (BA)2PbI4-free control device is significantly elevated, which verifies that the tunneling layer effectively suppresses the device off-state current and reduces energy consumption during weight modulation (Supplementary Fig. 21). The ultra-low energy consumption performance could be attributed to the synergistic regulation of carrier transport by the (BA)2PbI4 tunneling layer and GDY layer. Furthermore, the energy consumption is decreased as the light pulse width decreases (Fig. 3h and Supplementary Fig. 22), reaching only 4.41 fJ at a pulse width of 0.85 s. This performance is comparable to the energy consumption of biological neural systems (1-10 fJ per single synaptic event)53. In contrast to previously reported 2D heterojunction synaptic transistors with various functional layers, the GDY-NIST device markedly reduced lower energy consumption, representing a highly competitive performance among its peers (Fig. 3i and Supplementary Table 1). Synaptic plasticity emulation and high-linearity visual processing Paired-pulse facilitation (PPF) is a typical behavior of biological synapses for decoding spike-coded information54,55. When two successive light pulses (Δt = 2.25 s) are applied to the GDY-NIST device, the second pulse induces a higher excitatory postsynaptic current (EPSC) (A2) than the first (A1) (Fig. 4a). The PPF index gradually decreases as the time interval Δt increases, fitted with a decay function56 (Fig. 4b): where C1 and C2 are the initial facilitation amplitudes of the fast and slow phases, τ1 and τ2 are the relaxation time constants, respectively. Fitted values for the GDY-based device (C1 = 53.86 and C2 = 38.83) are far higher than those of the GDY-free control device (C1 = 3.01 and C2 = 2.74) (Supplementary Fig. 23a and Fig. 23b), verifying the strong charge attraction effect of GDY for enhanced charge accumulation. On this basis, biological learning-forgetting-relearning behavior is emulated. The initial learning photocurrent peak is reached with only ten pulses, whereas 28 pulses are required in the first learning cycle (Supplementary Fig. 24). Neuromorphic visual simulation of the photoelectric synaptic transistor. a PPF response under light pulse stimulation. b PPF index. c Spike-frequency dependent plasticity. d Spike-number dependent plasticity. e Spike-duration dependent plasticity. f Spike-intensity dependent plasticity. Weight regulation of potentiation and depression via alternating light and electrical pulses in the g GDY-based and h GDY-free devices. i Comparison of non-linearity factor with previously reported synaptic transistors. Spike-frequency dependent plasticity (SFDP), which underpins biological visual filtering, is mimicked via 10 consecutive light pulses at 0.25-5.0 Hz (Fig. 4c and Supplementary Fig. 2, 3c). Higher sensitivity to frequency variations is exhibited by the GDY-based device relative to GDY-free control devices, with a cutoff frequency of 0.66 Hz derived from SFDP fitting (versus 1.11 Hz for the control). It exhibits biological synapse-like high-pass filtering properties (Supplementary Fig. 25), enabling selective high-frequency information extraction from dynamic environments and highlighting its potential in neuromorphic image processing57. The synaptic plasticity is also emulated by the GDY-NIST device via the tuning of presynaptic light pulse parameters (Fig. 4d–f). EPSC amplitude and decay time increase with rising pulse number (3-30, spike number-dependent plasticity, SNDP), duration (0.15-0.9 s, spike duration-dependent plasticity, SDDP), and intensity (0.42-0.9 mW∙cm−2, spike intensity-dependent plasticity, SIDP), consistent with biological synaptic rules58,59. Analogous trends have been observed in the GDY-free control device (Supplementary Fig. 23d–f). To clarify the respective contributions of each functional layer, a (BA)2PbI4-free control device has been fabricated and fully characterized under identical conditions. Its slightly narrower memory window confirms that charge attraction and release are dominated by the GDY layer, with negligible contribution from the (BA)2PbI4 layer. Further characterization of typical synaptic behavior has been revealed that larger synaptic currents and higher energy consumption are observed for the control device under identical test conditions (Supplementary Fig. 26). For neuromorphic computing applications, the linearity of long-term potentiation (LTP) and long-term depression (LTD) is a critical metric for high-precision weight update (Fig. 4g, h). Carriers tunneling from the (BA)2PbI4 layer are uniformly attracted and released by the sp-hybridized alkyne bonds in the GDY layer, whereby highly linear modulation of the synaptic LTP/LTD processes (from 0.280/0.288 to 0.006/0.025) is realized. The channel conductance gradually increases with 50 consecutive light pulses (λ = 405 nm, Δt = 0.15 s, T = 0.05 s, intensity = 0.42 mW∙cm−2) and then gradually decreases with 50 consecutive back-gate voltage pulses (Δt = 0.15 s, T = 0.05 s, Vgs = −1 V) stimuli. The non-linearity factors of the LTP and LTD processes are calculated using the following equations60: where Gn+1 and Gn are the conductance updated by the (n + 1)th and nth pulse, α and β are the non-linearity factor and step size, respectively. The GDY-NIST device has been exhibited low non-linearity factors of αP = 0.006 and αD = 0.025, which are markedly lower than those of previously reported synaptic transistors (Fig. 4i and Supplementary Table 2), owing to the uniform carrier attraction and linear weight update enabled by GDY. Such an ultra-low non-linearity enables higher inference precision for neuromorphic vision devices in artificial intelligence (AI) image processing tasks13. Additionally, highly stable conductance modulation without significant performance degradation of the synaptic device is verified by 20 consecutive cycles (2000 pulses) of LTP and LTD endurance tests (Supplementary Fig. 27). To ensure the precise and controllable implementation of such weight update in practical operation, the decoupling of synaptic responses between each layer is realized via wavelength-selective optical excitation and dark-state electrical regulation, enabling independent, crosstalk-free modulation of each layer’s synaptic performance (Supplementary Fig. 28). The environmental and operational stability, a key prerequisite for practical applications, is systematically evaluated. Stable EPSC and PPF characteristics without degradation are exhibited by the unencapsulated GDY-NIST device after 10 weeks of storage in ambient conditions (22-26 °C, 40-60% relative humidity, light-shielded) (Supplementary Fig. 29). Over 88% performance retention after 200 h operation at 85 °C (Supplementary Fig. 30), corresponding to an estimated room-temperature (25 °C) lifetime of over 104 h (≈14 months), is verified by accelerated aging tests. Such a robust stability arises from the hydrophobic long-chain A-site cations of the RP perovskite and the protective MoS2 layer. To further optimize the device performance, the scaling relationship between the thickness of each functional layer (MoS2, (BA)2PbI4 and GDY) and the core synaptic performance metrics is clarified by fabricating and characterizing control devices with gradient thicknesses. The optimized thickness of each layer is determined to balance device stability, energy consumption and synaptic linearity (Supplementary Fig. 31). Finally, a photoelectric artificial neural network (ANN) is constructed to verify the device’s image perception capability (Supplementary Fig. 32a). Benefiting from the GDY-NIST device’s ultra-low LTP/LTD non-linearity and highly balanced symmetry (γ = ||αP | − |αD | |), the network reaches 88.5% recognition accuracy after 30 epochs, versus only 80.9% for the GDY-free control device network due to poorer linearity (Supplementary Fig. 32b). Heterojunction array and dynamic neuromorphic vision system A 3×3 heterojunction array for the GDY hardware neuromorphic convolution kernel has been fabricated on 2D monolithic integrated nanosheets with a lateral dimension of about 40 μm (Fig. 5a and Supplementary Fig. 33). The channel length of each transistor cell is 2 μm. Based on the uniform photocurrent accumulation effect of the GDY heterojunction photoelectric array, dynamic image sensing and processing can be achieved simultaneously. To verify the reliability of the as-fabricated array for practical neuromorphic applications, the device-to-device (D2D) variations of the synaptic array have been analyzed (Supplementary Fig. 34a), which is the basis for the array’s subsequent neuromorphic processing applications. The EPSC variation (the ratio of A10 to A1) of the array has been evaluated using the ratio of the standard deviation to the average value, with a relatively low fluctuation of 3.68% (Supplementary Fig. 34b). Based on such a high degree of array uniformity, the array holds potential for neuromorphic image logic operations (Supplementary Fig. 35). Furthermore, due to the strong synaptic spike accumulation effect of the GDY layer, the preprocessed images can be extracted from a series of noisy optical inputs (Supplementary Fig. 36 - Fig. 38). Dynamic image processing performance of the GDY hardware-level neuromorphic kernel. a Schematic diagram of the GDY hardware-software co-processing system for in-sensor neuromorphic image processing. b The photocurrent distribution of various types in the GDY array. c Hardware-implemented image processing results using the obtained kernel weights. d Schematic diagram of the neural network for human action classification. e Recognition accuracy and loss of the neural network. f Confusion matrix results of the trained neural network. g Real-time object trajectory detection based on the GDY neuromorphic kernel system. h Predicted probability distribution of the eight movement directions. A customized multi-channel data acquisition (DAQ) circuit has been developed to support hardware-implemented convolution processing. This circuit achieves high-sensitivity nA-level photocurrent acquisition from the GDY array while preserving its intrinsic linearity, and establishes a hardware-software co-processing platform for real-time dynamic image processing (Supplementary Fig. 39). To construct the hardware convolution kernel, incident light pulse (λ = 405 nm, Δt = T = 0.15 s, intensity = 0.9 mW∙cm−2) is utilized to achieve uniform linear photocurrent stacking in each cell of the array. By individually adjusting the weight of each cell, various types of convolution kernels can be designed (Supplementary Fig. 40). The stable 3×3 photocurrent weights are obtained by simple addition and subtraction operations to construct the convolution kernel, and the uniform numerical values are nearly consistent with the standard kernel (Fig. 5b). Furthermore, the vertical edge detection result (F1) and horizontal edge detection result (F2) are combined by (F1)2+(F2)2 to obtain the reconstructed image feature map61–63. The image processing results (Fig. 5c) achieved using the GDY-NIST device array as an in-sensor convolutional kernel exhibit nearly identical performance to the standard convolutional kernel (Supplementary Fig. 41). Beyond static image edge detection, the GDY neuromorphic kernel can be further extended to dynamic neuromorphic vision tasks. Spatial feature extraction for human action recognition is enabled by the photocurrent distribution in the GDY neuromorphic kernel. The Weizmann dataset64 has been used to evaluate neuromorphic dynamic image recognition, with 10 human actions including bend, jack, jump, pjump, run, side, skip, walk, and waves1 and 2. Temporal attention features are clustered into global features, then the probabilities of ten action types are output through the fully connected layer of the neural network (Fig. 5d). After training for 79 epochs, the convolutional neural network achieved 100% recognition accuracy on the Weizmann dataset (Fig. 5e). Moreover, the results of the confusion matrix’s high diagonal accuracy confirm the GDY neuromorphic kernel achieves a high ranking in the human action recognition task for the Weizmann dataset (Fig. 5f). In addition to human action recognition, the bio-inspired GDY neuromorphic kernel vision system can also realize real-time moving target trajectory tracking and direction judgment (Fig. 5g). Real-time tracking of the moving target is performed and its motion paths are extracted by this system, with frame differences of the target region computed by combining convolutional features. The tennis movement trajectories accuracy of the eight directions (left, right, up, down, up-left, down-right, down-left, and up-right) are labeled and evaluated by the GDY neuromorphic kernel system (Fig. 5h and Supplementary Movie. 1). Furthermore, the link between the linearity of synaptic transistors and the performance of the neuromorphic vision system has been verified via GDY-free control experiments. It is validated that cumulative convolution errors, degraded low-precision image processing, and a dynamic recognition accuracy of merely 60–70% are caused by device-level nonlinearity (Supplementary Fig. 42). Tested in non-uniformly illuminated dynamic scenes, a tennis ball direction recognition accuracy of over 95% is achieved by the GDY neuromorphic kernel, thanks to the robust anti-interference capability enabled by its high linearity (Supplementary Fig. 43). It is demonstrated that the GDY neuromorphic kernel holds broad application prospects in sports analysis, robot navigation, and intelligent traffic management. Discussion In summary, we have proposed a neuromorphic image-sensing transistor by using a tri-layer 2D vdW heterostructure for neuromorphic vision. The heterojunction comprises the charges modulation layer of GDY, a photo-responsive and charge tunneling layer of (BA)2PbI4, and a carrier transport layer of MoS2. The GDY layer with abundant alkyne bonds can uniformly interact and modulate photo-generated carriers, providing homogeneous charge accumulation and inducing a quasi-linear time-dependent synaptic behavior. The photo-tunable perovskite (BA)2PbI4 nanosheet facilitates hole tunneling, preventing electrical leakage and reducing the dark current, allowing the synaptic device to work in high energy-efficiency. The MoS2 layer with high intrinsic mobility can match the optical absorption bandgap of 2D-(BA)2PbI4 perovskite and encapsulate perovskite from ambient air. Additionally, the GDY-NIST device array exhibits long-term air-stability (>10 weeks), along with highly competitive linearity in synaptic transistor weight modulation (αP = 0.006, αD = 0.025), and notably low energy consumption (~41.8 aJ per synaptic event) among 2D heterojunction synaptic transistors. Benefiting from the vdW heterojunction, the proposed 2D nanosheet synaptic array realizes photoelectric neuromorphic dynamic image processing as a hardware-level neuromorphic convolutional kernel. These results may contribute to the neuromorphic image perception, memory, and processing capabilities, paving the way for the next-generation neuromorphic vision. Methods Synthesis of (BA)2PbI4 PbO (1.12 g, 5.0 mmol, 99.9%) powders were dissolved in a mixture of hydriodic acid (6.0 mL, 57% w/w) and H3PO2 (1.0 mL, 50% w/w) under stirring to obtain a clear yellow solution and heated to 100 °C. The BAI solution was obtained by mixing n-butylamine (293 mg, 4.0 mmol, ≥99.5%) solution with hydriodic acid (4.0 mL, 57% w/w) and stirring continuously for 4 h in an ice bath. Then, the BAI solution was slowly injected to the hot PbI2 solution with continuous stirring. The stirring was stopped and the resultant solution was cooled naturally to room temperature and left overnight to complete the crystals growth. Finally, orange (BA)2PbI4 crystals were obtained by suction filtration and dried overnight in a vacuum oven. Synthesis of graphdiyne GDY was prepared on copper film through a modified wet-chemistry lithography method18. Hexakis[(trimethylsilyl)ethynyl]benzene (HEB-TMS) was synthesized according to the previous work65. HEB-TMS (5.0 mg, 7.6 μmol) was added to a vial charged with 10 mL of chloroform. Then, tetrabutylammonium fluoride (1.0 M in tetrahydrofuran, 50.0 μmol) was added and stirred for 10 min in the dark to obtain hexaethynylbenzene (HEB) solution. A freshly prepared Cu film (30.0 nm thickness, thermally evaporated onto a pre-cleaned rigid substrate) was immersed in argon de-aerated chloroform (80.0 mL, purged with argon for 15 min) under the protection of argon atmosphere. After adding 0.35 mL of pyridine, the previously prepared HEB solution was added dropwise within 10 min. The reaction system was kept undisturbed for 12 h, yielding the GDY nanomembrane on the Cu substrate. The GDY membrane was then rinsed with acetone and dried under an argon stream. Device fabrication After synthesizing GDY on the Cu film, the polymethyl methacrylate (PMMA) (46 mg·mL−1 in chlorobenzene) was spin-coated on the surface as a sacrificial layer and the copper film was etched in a 0.1 M ammonium persulfate ((NH4)2S2O8, ≥98%) solution. The PMMA/GDY membrane was then transferred onto the SiO2/Si substrate rinsed three times with deionized water. After removing the PMMA layer with acetone and isopropyl alcohol, the GDY/SiO2/Si sample was dried with N2 flow for further use. The bulk MoS2 crystals were purchased from Six Carbon Technology (Shenzhen). Multilayer (BA)2PbI4 and MoS2 were exfoliated from their crystals with Scotch tape and transferred onto transparent polydimethylsiloxane (PDMS) film. A transfer platform was employed to sequentially transfer MoS2 and (BA)2PbI4 nanosheets onto the GDY/SiO2/Si substrate. Source-drain electrodes (Cr/Au, 8 nm/60 nm) were patterned via standard electron-beam lithography (EBL), followed by thermal evaporation and lift-off processes. Characterizations AFM and C-AFM images were obtained using a Bruker Dimension Icon microscope. The XRD patterns were recorded at a scan speed of 5° min−1 by an X-ray diffractometer (Rigaku SmartLab SE) with Cu Kα radiation. Raman spectra were recorded with the Micro-Raman Spectroscopy System (Edinburgh Instruments RMS1000C). The FT-IR absorption spectra were recorded using a Fourier transform infrared spectrometer (Bruker INVENIO S). The absorption spectrum of the film was recorded by using a UV-Vis spectrophotometer (Agilent Cary 5000). The XPS spectra and UPS spectra were acquired using an ESCALAB 250Xi spectrometer (Thermo Scientific). PL and TRPL data were obtained using an Edinburgh FS5 fluorescence spectrometer. The electrical properties were measured at room temperature under the air atmosphere by a semiconductor analyzer (Keithley 4200-SCS). Supplementary information Supplementary Information Description of Additional Supplementary Files Supplementary Movie 1 Transparent Peer Review file Source data Source data Author contributions Y.Y., Y.Z., and J.W. contributed equally to this work. W.X. and J.X. conceived and supervised the work. Y.Y., Y.Z. and W.X. designed the experiments. Y.Y. and Y.Z. fabricated the devices and tested the performance. J.W. and Z.X. prepared the GDY film, and performed the characterization. T.W. and R.Y. built the circuit system, and assisted in device testing. Y.Y. wrote the manuscript, J.X. and W.X. revised the manuscript. All authors reviewed and commented on the manuscript. Peer review Peer review information Nature Communications thanks Yoonyoung Chung, Timon Rabczuk, Min Zhang and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding W. X. discloses support for the research of this work from National Key R&D Program of China (2022YFA1204500 and 2022YFA1204504), the National Science Fund for Distinguished Young Scholars of China (T2125005), and the Fundamental Research Funds for the Central Universities, Nankai University. J. X. discloses support for the research of this work from National Key R&D Program of China (2022YFA1204500 and 2022YFA1204504), the National Natural Science Foundation of China (22475106 and 52172045), the Natural Science Foundation of Tianjin (23JCJQJC00110), and the Fundamental Research Funds for the Central Universities, Nankai University. Data availability The data generated in this study are provided in the Source Data file. Source data are provided with this paper. Code availability The Weizmann dataset used in this study is sourced from https://www.wisdom.weizmann.ac.il/~vision/SpaceTimeActions.html. The code that used for the simulation for this study is available from the corresponding authors. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Yiming Yuan, Yinxing Zhang, Jie Wang. Contributor Information Jialiang Xu, Email: jialiang.xu@nankai.edu.cn. Wentao Xu, Email: wentao@nankai.edu.cn. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-75159-6. References Associated Data Supplementary Materials Supplementary Information Description of Additional Supplementary Files Supplementary Movie 1 Transparent Peer Review file Source data Data Availability Statement The data generated in this study are provided in the Source Data file. Source data are provided with this paper. The Weizmann dataset used in this study is sourced from https://www.wisdom.weizmann.ac.il/~vision/SpaceTimeActions.html. The code that used for the simulation for this study is available from the corresponding authors.