Interface Engineering for Scalable Optoelectronic Reservoir Computing

ABSTRACT

Two‐dimensional (2D) materials offer an attractive platform for optoelectronic reservoir computing (RC) and neuromorphic hardware, promising energy‐efficient in‐sensor processing for edge intelligence. However, scaling such systems to practical large‐scale arrays is hindered by substantial device‐to‐device variability due to the stochastic distribution of intrinsic defects in these materials. Here, we demonstrate a scalable and highly uniform reservoir array based on vertical p‐GaN/n‐MoS₂ heterojunctions via an interface engineering strategy. A controlled thermal pretreatment process produces a uniform GaO_X interlayer with a high density of statistically homogeneous defects, which serve as reproducible carrier trapping centers to generate reliable memory effects. This approach ensures highly consistent nodal responses across the array, overcoming a key bottleneck in 2D material‐based neuromorphic hardware. The system exhibits robust spatiotemporal processing capabilities, experimentally realizing dynamic trajectory reconstruction, an 87.24% accuracy in static digit classification, and a normalized mean squared error of 7.01 × 10⁻⁵ in predicting second‐order nonlinear dynamics. These results establish interface engineering as a decisive route to overcoming the uniformity bottlenecks of 2D materials, advancing the practical implementation toward wafer‐scale optoelectronic neuromorphic hardware.

Interface engineering via thermal pretreatment creates a uniform GaO_X interlayer in vertical p‐GaN/n‐MoS₂, delivering statistically homogeneous carrier traps and uniform synaptic dynamics across scalable arrays, enabling UV‐driven reservoir computing that achieves dynamic trajectory reconstruction, 87.24% handwritten digit classification, and second‐order nonlinear prediction with an NMSE of 7.01 × 10⁻⁵.

1. Introduction

The explosion of the internet of things (IoT) and edge intelligence has driven an urgent demand for the real‐time, energy‐efficient processing of massive unstructured spatiotemporal data [1, 2, 3, 4]. However, conventional von Neumann architectures are limited by the physical separation of processing and memory units, leading to severe latency and efficiency bottlenecks when handling such data‐intensive tasks [5, 6, 7]. Reservoir computing (RC) presents a compelling neuromorphic computing paradigm, employing a fixed, nonlinear dynamical system to map input signals into a high‐dimensional feature space (Figure 1A), followed by training only a simple linear readout layer [8, 9, 10, 11]. In particular, optoelectronic RC further combines optical sensing and neuromorphic computing directly at the sensor side (in‐sensor computing), offering an ideal hardware solution for edge vision tasks [12, 13, 14, 15, 16, 17, 18]. While progress has been made with single devices, practical hardware acceleration requires the deployment of large‐scale, high‐density physical reservoir arrays.

FIGURE 1.

Structure and optoelectronic characterization of p‐GaN/n‐MoS₂ heterojunctions for a UV‐driven reservoir computing system. (A) Schematic of the reservoir computing system comprising an input layer, a reservoir layer, and a readout layer. (B) I‐V characteristics of the p‐GaN/n‐MoS₂ heterojunction. Inset: optical micrograph of the array. (C) Cross‐sectional schematic of a single device, illustrating the multilayer stack of Si substrate, AlN buffer, composition‐graded AlGaN, GaN buffer, p‐type GaN, and monolayer MoS₂. (D) High‐resolution TEM (HRTEM) image of the device interface and the corresponding energy‐dispersive X‐ray spectroscopy (EDS) elemental maps, revealing the spatial distributions of Mo (green), S (pink), O (blue), and Ga (purple) in the heterojunction region. (E) 0 V bias band diagram of the heterojunction. (F) Current response of the device under ultraviolet (385 nm, 2.046 µW/cm²) and visible light (470 nm, 1.982 µW/cm², 530 nm, 2.168 µW/cm², 625 nm, 1.876 µW/cm²). (G) PPF magnitude (A ₂/A ₁) as a function of the inter‐pulse interval, well described by a double‐exponential model. (H) Current response of different pulse numbers at a frequency of 1 Hz for a duration of 0.1 s. (I) Statistics of single‐pulse responses across the 64 devices, reporting the peak current during illumination and the residual current 1 s after pulse removal.

Two‐dimensional materials (2D materials), with their atomic thinness, strong optoelectronic response, and gate tunability, have emerged as promising candidates for optoelectronic reservoir nodes [19, 20, 21, 22, 23]. Current approaches often rely on intrinsic lattice defects as trapping centers for photogenerated carriers to emulate synaptic plasticity. However, the stochasticity of defects originating from thermodynamic fluctuations during growth leads to significant spatial and density variations [24, 25, 26]. This variability, tolerable in single devices, becomes a critical limitation in large‐scale arrays. Inconsistent nodal responses distort the feature mapping space in the RC system, severely degrading recognition accuracy and increasing the training burden [27, 28, 29]. Therefore, controlling defect distribution and uniformity is not merely a fabrication challenge but a fundamental requirement for scalable 2D material‐based neuromorphic hardware.

In this work, we propose an interface engineering strategy to construct high‐performance, highly uniform reservoir arrays based on vertical p‐GaN/n‐MoS₂ heterojunctions. Instead of relying on random intrinsic defects, we engineer a uniform GaO_X interface layer on the p‐GaN surface via thermal pretreatment prior to device assembly. This engineered interface contains a high density of statistically homogeneous defects that act as dominant carrier trapping centers, providing reliable and consistent memory effects. The resulting array exhibits highly uniform nodal responses, providing a robust hardware foundation for scalable neuromorphic computing. Leveraging this uniformity of the array, we constructed a UV‐responsive RC system that exploits the wide bandgap of p‐GaN to effectively filter out ambient visible noise. The UV‐responsive RC system is capable of dynamic trajectory reconstruction, static handwritten digit classification (87.24% accuracy on MNIST), and accurate prediction of second‐order nonlinear dynamics (normalized mean squared error (NMSE) of 7.01 × 10⁻⁵). These results highlight interface engineering as a transformative path to overcome the scalability bottlenecks of 2D materials, paving the way for practical optoelectronic neuromorphic computing.

2. Results

2.1. Structural and Optoelectronic Characterization of p‐GaN/MoS₂ Heterojunction Array

To achieve high spatial uniformity across the array, we first deposited bottom electrodes on the p‐GaN substrate and then applied a controlled thermal pretreatment to engineer a uniform, thin GaO_X layer on the surface. Monolayer MoS₂ was subsequently transferred onto this prepared substrate via a wet transfer process. The device regions were then patterned by reactive ion etching (RIE), followed by the deposition of top metal electrodes to complete the vertical p‐n junction array (Figure S1). The structural integrity and composition of the heterostructure were confirmed through Raman spectroscopy. As shown in Figure S2A, the characteristic in‐plane (E2g) and out‐of‐plane (A1g) vibrational modes of MoS₂ exhibit a peak separation of 18.8 cm⁻¹, confirming the monolayer nature of the transferred MoS₂ film. In addition, the typical Raman peaks of GaN can also be observed in Figure S2B, verifying the successful formation of the heterostructure. The rectifying electrical behavior of the junction is evidenced by its nonlinear current‐voltage (I‐V) characteristics (Figure 1B). The inset is the optical micrograph of the partial array. A schematic cross‐section in Figure 1C illustrates the complete device stack. High‐angle annular dark‐field scanning transmission electron microscopy (HAADF‐STEM) and energy‐dispersive X‐ray spectroscopy (EDS) provide direct visualization of the critical interface (Figure 1D). A well‐defined van der Waals gap of approximately 6.2 Å was observed between the MoS₂ and p‐GaN layers. The EDS elemental mapping revealed the presence of oxygen at the interface, indicating slight GaO_X formed at the van der Waals gap. This interfacial oxide layer is formed during surface oxidation under thermal pretreatment in an oxygen‐containing environment. The resulting GaO_X contains a high density of unsaturated bonds that act as trapping centers for photogenerated carriers and delay their release, which provides the dynamical foundation for device‐level synaptic plasticity.

The energy band of the p‐GaN/n‐MoS₂ heterojunction at 0 V bias (Figure 1E) provided a clearer understanding of the carrier dynamics. Due to the difference of Fermi levels, electrons diffuse from n‐MoS₂ to p‐GaN while holes diffuse in the opposite direction, establishing an interfacial depletion region and a built‐in electric field. Systematic investigation of the bias‐dependent synaptic response (Figure S3) identified ‐1 V as the optimal operating point. At this reverse bias, the expanded depletion width and enhanced interfacial electric field facilitate the controlled capture and subsequent release of carriers by the GaO_X traps. This process engenders a stable, gradually decaying photocurrent after stimulus removal, ensuring highly repeatable synaptic dynamics.

To further elucidate the role of the engineered GaO_X interface layer in carrier dynamics and to decouple its contribution from intrinsic material variability, we evaluated a control group of heterojunctions without the thermal pretreatment. Under the identical ‐1 V reverse bias and optical stimulation, these untreated devices exhibited conventional photodetector behaviors. As shown in Figure S4, the transient photocurrent of the control devices decayed rapidly upon stimulus removal, lacking a stable and controllable memory effect. In stark contrast, the devices with the engineered GaO_X layer demonstrated a significantly prolonged current relaxation. To quantitatively assess the charge trapping and release dynamics, the transient recovery processes were fitted using a double‐exponential decay model [30, 31, 32, 33]:

$$Y=y_0+a_1{e}^{-t/{\tau }_1}+a_2{e}^{-t/{\tau }_2}$$ (1)

where τ₁ and τ₂ represent the characteristic relaxation times for the fast and slow phases, respectively. The fitting results reveal that the thermally treated devices exhibit a substantially longer slow‐relaxation time (τ₂ =  6.30541 s). This pronounced kinetic transition provides electrical evidence that the engineered GaO_X acts as a carrier trapping centers. These states effectively capture photogenerated carriers and restrict their recombination rate, thereby providing the essential physical mechanism for robust synaptic plasticity.

Spectral response measurements (Figure 1F) highlight the system's pronounced sensitivity to ultraviolet (UV) illumination (385 nm), which elicits a significantly larger photocurrent with prolonged decay compared to visible wavelengths. This result indicates that UV illumination more effectively generates and traps photocarriers at the heterointerface, making it ideal for driving synaptic plasticity. To evaluate short‐term synaptic plasticity, we subjected the device to paired UV pulses with a fixed width of 100 ms. As shown in Figure S5, the second current peak A ₂ significantly exceeds the first peak A ₁, clearly demonstrating paired‐pulse facilitation (PPF). This enhancement confirms that the device conductance is dynamically modulated by the temporal history of the input. The magnitude of this facilitation is quantified by the PPF index, defined as:

$$\mathrm{PPF}\mathrm{index}=A_2/A_1\times 100\%$$ (2)

The observed facilitation originates from the short‐term trapping of photogenerated carriers at the interface, where the built‐in electric field actively promotes charge retention. Consequently, residual carriers from the first pulse raise the conductance baseline, effectively potentiating the response to the second pulse. The PPF index decays as the inter‐pulse interval Δt increases, fitting well to a double exponential model:

$$Y=1+a_1{e}^{-\mathrm{\Delta }t/{\tau }_1}+a_2{e}^{-\mathrm{\Delta }t/{\tau }_2}$$ (3)

where a ₁ and a ₂ are the initial facilitation magnitudes of the rapid and slow phases, respectively, and τ₁ and τ₂ are the characteristic relaxation time of the rapid and slow phases, respectively. As shown in Figure 1G, this model captures the multiscale carrier release dynamics of interfacial traps, establishing the kinetic foundation for synaptic plasticity.

To systematically evaluate the device's temporal processing capabilities, we first investigated spike‐number‐dependent plasticity (SNDP). As illustrated in Figure 1H, applying a train of UV pulses (0.1 s duration, 0.9 s interval) with varying counts (11–51) results in a progressive enhancement of the postsynaptic current. Notably, the relaxation time extends significantly with increased pulse number, indicating a transition from short‐term toward long‐term memory behavior. This dynamic modulation is consistent across other stimulation dimensions. By increasing the pulse duration (SDDP, Figure S6), frequency (SFDP, Figure S7), or intensity (SIDP, Figure S8), the device exhibits analogous conductance enhancements and prolonged decay dynamics. Collectively, multidimensional optical modulation promotes the accumulation of captured photocarriers and leads to a prolonged relaxation time, marking a clear transition from short‐term to long‐term memory.

Attributed to the statistically homogeneous defect density of the engineered GaO_X layer, the array exhibits exceptional uniformity. Statistical analysis across 64 independent devices (Figure 1I) demonstrates highly consistent distributions for both the peak photocurrent under single‐pulse illumination and the residual memory current measured 1 s after light removal. The statistical histograms of both the peak and residual currents are provided in Figure S9, along with the explicitly calculated mean, standard deviation (SD), and low coefficient of variation (CV), demonstrating the exceptional spatial uniformity of the engineered device array. This low device‐to‐device variation is paramount for reliable large‐scale integration. By comparing the synaptic responses immediately after fabrication and 100 days later (Figure S10), we demonstrate the device's robust long‐term stability under ambient, unencapsulated conditions. We further verified operational reproducibility by applying 100 successive UV pulses (0.1 s duration), allowing for complete relaxation between stimuli. The resulting current responses exhibit exceptional consistency in both amplitude and decay kinetics across all cycles (Figure S11), and the statistical histogram of peak currents exhibits a low coefficient of variation of 5.09% (Figure S12), validating the device's repeatability for synaptic emulation. In addition, although the estimated lowest energy consumption per single synaptic event (approximately 25 pJ, detailed in Figure S13) is higher than purely electrical synapses, it falls well within the optimal and accepted range (typically pJ to nJ) [34, 35, 36, 37] for current state‐of‐the‐art optoelectronic synapses. The demonstrated combination of nonlinearity, repeatability, and uniformity establishes a robust hardware foundation for implementing physical reservoir nodes capable of advanced temporal information processing.

2.2. Dynamic Trajectory Reconstruction Under UV Light Path Input

We evaluated the spatiotemporal information processing capability of an 8 × 8 p‐GaN/n‐MoS₂ heterojunction array by testing its performance in reconstructing dynamic UV light trajectories. The system operates by leveraging the inherent, graded relaxation dynamics of each device node. As a focused UV spot scans across the array, it illuminates individual pixels, whereby the induced photocurrent in each pixel decays gradually, after which the light moves away. This decay effectively encodes both the spatial position and temporal sequence of a moving light spot into a distributed pattern of residual currents across the array (Figure 2A). Consequently, the array performs parallel, in‐sensor processing and retains a short‐term history of the light path without external memory.

FIGURE 2.

Temporal response characteristics of the device array under light trajectory input. (A) The p‐GaN/n‐MoS₂ heterojunction array exhibits a gradual decay in current intensity during ultraviolet (UV) light scanning along a predefined trajectory. (B) Schematic illustration (left) and optical image (right) of the UV light scanning along an “L”‐shaped path across the device array. (C) Input patterns of two representative trajectories (“Z” and “L”) (top) and the corresponding normalized current response heatmaps (bottom) recorded at four steps, visualizing the dynamic conductance evolution induced by sequential light exposure.

To demonstrate this capability, we programmed the UV spot to trace letter‐shaped paths, such as an “L” shape, across the device grid (Figure 2B). Owing to the synaptic decay dynamics, a correlation is established between the residual current level and the timing of stimulation, where pixels triggered earlier show lower residual currents than those stimulated more recently. This direct correlation enables the generation of trajectory maps that encode both spatial coordinates and temporal order. For visualization, we extracted the residual photocurrent distributions at successive time steps and mapped them as normalized heatmaps. Figure 2C presents two representative trajectories (“Z” and “L” shapes). The upper panels display the time‐resolved optical input patterns, while the lower panels present the corresponding residual current maps. The clear, evolving patterns in the heatmaps confirm that the array accurately records the spatial morphology and faithfully reconstructs the temporal order of the input trajectories, confirming its effectiveness for real‐time trajectory reconstruction.

2.3. Optoelectronic RC for Handwritten Digit Classification

To assess the system's performance in static image recognition, we employed an RC system based on a 16 × 4 p‐GaN/n‐MoS₂ device array for a classic handwritten digit classification task using the MNIST dataset. This dataset comprises 60 000 handwritten digit images with a resolution of 16 × 16 pixels. The chosen resolution retains essential digit features while enabling efficient mapping and encoding within the spatial architecture of the 16 × 4 reservoir array designed in this work. The encoding strategy is illustrated in Figure 3A. Each image is unfolded row‐wise and partitioned into four 4‐pixel segments per row. In this scheme, white pixels are encoded as “1” (1‐s UV pulse), while black pixels correspond to “0” (no illumination). These 4‐bit segments are converted into temporal pulse sequences and sequentially applied to the reservoir nodes, inducing nonlinear transient responses that enable dynamic state mapping. Figure 3A also displays an example row and its corresponding temporal pulse sequence.

FIGURE 3.

Handwritten‐digit recognition using the optoelectronic reservoir system. (A) Schematic of the pulse‐encoding strategy for handwritten inputs. (B) Current response mapping of 10 handwritten digits in the reservoir system. (C) Confusion matrix on an MNIST subset. (D) Evolution of classification accuracy and loss during model training.

Prior to full‐scale classification, we characterized the consistency and separability of nodal responses. As detailed in Figure S14, four representative units (Cells 1–4) were subjected to six distinct 4‐bit UV pulse streams. In these tests, each pulse had a duration of 1 s, and the state was read 1 s after stimulus removal. Although minor device‐to‐device variations were observed, all units exhibited consistent synaptic dynamics, where the current response increased with the pulse count. Crucially, the output currents corresponding to different pulse configurations are clearly distinguishable. This distinct separability validates the device's capability for dynamic state mapping and nonlinear modulation, providing a robust hardware basis for optoelectronic RC. Detailed characterization of both individual units and the full array under 16 distinct 4‐bit pulse streams is provided in the supplementary materials (Figures S15 and S16). Specifically, Figure S15 illustrates the response of a single device, while Figure S16 displays the readout current distribution across all 64 devices. These results collectively demonstrate the robust capability of the system for array‐level dynamic state mapping. Building on this foundation, we investigated the image mapping capability using a custom set of ten 5 × 4 binary digit patterns (0‐9), as illustrated in Figure S17A. In this encoding scheme, each pixel was temporally encoded as a 1‐s UV pulse sequence, where white pixels represented illumination and black pixels indicated the absence of light. This scheme transformed spatial image structures into temporally extended input sequences, enabling the reservoir to encode spatiotemporal features. Figure S17B shows the current response of device nodes corresponding to each row of the image in the reservoir array. The resulting current distributions exhibit clear differences across digit patterns, demonstrating the system's ability to perform nonlinear and separable mapping of spatiotemporal image features under temporal optical flow.

For the MNIST task, the reservoir's state was defined by the photocurrent read from each of the 64 nodes after presenting the temporally encoded image. Figure 3B visualizes the readout current maps (as heatmaps) for representative samples from each digit class (“0”–“9”). These heatmaps directly correspond to the physical 16 × 4 reservoir array, where the 16 rows represent the row indices of the original image and the 4 columns correspond to the four pixel segments within each row. The resulting matrices exhibit highly structured local distributions that align well with the spatial geometry of the input images, indicating that the reservoir array can effectively capture spatial patterns and local feature variations. These distinctive state representations provide a solid foundation for subsequent classification. A linear single‐layer softmax regression network without any hidden layers was trained via gradient descent to classify the current states (please see supporting information Note S1 for details). Figure 3C shows the classification confusion matrix for the entire “0”–“9” test set. Most predicted labels align well with the ground‐truth values along the main diagonal, indicating strong class separability across multiple digit categories. Figure 3D illustrates the changes in accuracy and cross‐entropy loss over 1000 epochs. As training progressed, the accuracy gradually improved and converged at 87.24%, while the loss decreased to 0.502, demonstrating the reservoir's learning capability and robustness in handwritten digit classification. By achieving a classification accuracy highly comparable to a standard software ESN baseline in supporting information Note S2, we solidly verify that the thermally engineered interface provides a powerful and reliable physical platform for complex neuromorphic classification tasks.

2.4. Dynamic Processing of Nonlinear Temporal Sequences

Beyond static handwritten digit recognition, we evaluated the reservoir's proficiency in processing complex temporal signals by challenging it to predict the output of a benchmark second‐order nonlinear dynamical system. Nonlinear dynamical systems are employed in a wide range of engineering fields, with second‐order cases receiving particular attention for their close correspondence to electrical circuit models (e.g., RLC networks). Figure 4A shows the structure of the second‐order nonlinear system prediction task in our experimental setup. A randomly generated discrete signal u(k) serves as the system input, generating an output y(k) through a nonlinear transfer function. In this experiment, the second‐order nonlinear transfer equation is defined as:

$$y\left(k\right)=\left[0.6\ast y\left(k-1\right)+0.2\ast y\left(k-2\right)+\frac{0.05\ast u\left(k\right)}{1+y^2\left(k-1\right)}+0.05\right]$$ (4)

FIGURE 4.

Processing of second‐order nonlinear dynamic tasks using the optoelectronic reservoir system. (A) Workflow for modeling nonlinear dynamic systems using physical reservoirs. (B) Comparison between the predicted output y′(k) and the ground‐truth output y(k). The upper panel shows the training phase using 300 samples, while the lower panel displays the testing phase with 100 samples. (C) Radar chart comparing the performances of p‐GaN/n‐MoS₂ heterojunction for reservoir computing with reported devices.

The output y(k) at time step k depends not only on the current input u(k) but also on the preceding outputs y(k − 1) and y(k − 2). Such temporal dependencies make the prediction of these systems challenging for conventional neural networks. In contrast, a physical reservoir leverages the intrinsic short‐term memory effect of the heterojunction to retain the influence of past inputs in the system state.

Constructing the PRC primarily involves embedding the underlying relationship between the input u(k) and the output y(k) within the reservoir dynamics. Herein, the input sequence u(k) is converted into an optical signal that is projected onto the device, whose current response forms the physical reservoir. Owing to its short‐term memory, the device's current is influenced not only by the present optical input but also by previous inputs. A random sequence u(k) of 400 points was used as the input, and the corresponding output y(k) was generated through the transfer equation, as shown in Figure S18.

The first 300 points of the sequence were used to train the readout layer of the reservoir, while the remaining 100 points were employed to evaluate the prediction performance. As shown in Figure 4B, the predicted output y′(k) obtained from the trained reservoir matches well with the ground‐truth target y′(k) for the remaining 100 test points. The reservoir achieves a normalized mean square error (NMSE) of 7.01 × 10⁻⁵, confirming its strong capability for nonlinear temporal sequence modeling.

We contextualized this performance by comparing our device against other reported physical reservoirs in a radar chart encompassing key metrics such as uniformity, recognition accuracy, NMSE, and operational flexibility (Figure 4C) [16, 17, 18, 38, 39]. Our p‐GaN/n‐MoS₂‐based reservoir demonstrates a balanced and overall superior profile, highlighting its potential for scalable, low‐power, and high‐performance neuromorphic processing of diverse spatiotemporal tasks.

3. Conclusion

Strategic interface engineering implemented via thermal pretreatment endows the p‐GaN/n‐MoS₂ heterojunction array with uniform synaptic dynamics, which is effectively leveraged in the reservoir system to process both static and dynamic tasks. Supported by such consistent nodal responses, the reservoir array demonstrates superior spatiotemporal information processing capabilities, successfully realizing dynamic trajectory reconstruction, handwritten digit classification, and nonlinear prediction. Our finding overcomes the critical limitations associated with relying on stochastic intrinsic defects or constructing complex device architectures to induce synaptic plasticity. Instead, by implementing a thermal pretreatment, we introduce a GaO_X interlayer populated with statistically homogeneous carrier‐trapping states, thereby achieving scalable and highly uniform synaptic plasticity. This controllable processing strategy can be considered a pivotal advancement for the practical deployment of neuromorphic computing hardware.

4. Experimental Section/Methods

4.1. Epitaxy of p‐GaN and Monolayer MoS₂

The p‐GaN epitaxial structure was grown on a 6‐inch (111)‐oriented silicon wafer with a thickness of 1000 µm using metal–organic chemical vapor deposition (MOCVD) in an AIXTRON G5 system. Trimethylaluminum, trimethylgallium, and ammonia served as the precursors, with hydrogen as the carrier gas. A 250 nm aluminum nitride (AlN) buffer layer was first deposited at 1090°C, followed by a 2.8 µm AlGaN compositional grading layer grown at 1010°C. Subsequently, a 2 µm GaN layer was deposited, upon which a 300 nm magnesium‐doped p‐GaN layer was epitaxially grown.

Wafer‐scale monolayer MoS₂ was synthesized on c‐plane sapphire substrates via a home‐built three‐zone chemical vapor deposition (CVD) system. Sulfur powder (20 g, Alfa Aesar, 99.9%) and molybdenum trioxide (240 mg, Alfa Aesar, 99.9%) were used as precursors and loaded separately into quartz boats in Zone 1 and Zone 2, respectively. Zone 1 was supplied with 40 sccm of pure Ar gas, while Zone 2 received a mixture of 240 sccm Ar and 8 sccm O₂ as the carrier gas. The temperatures for Zone 1 (sulfur sublimation), Zone 2 (oxide precursor), and Zone 3 (growth zone) were maintained at 160°C, 580°C, and 930°C, respectively. The chamber pressure was held at approximately 1 Torr, and the growth proceeded for 45 min, yielding a continuous monolayer MoS₂ film across the wafer.

4.2. Device Fabrication Process

The device fabrication began with the deposition of a Pd/Au bilayer onto the surface of the p‐GaN to serve as the contact electrode, using electron beam evaporation. Following the electrode deposition, a thermal pretreatment was performed on the p‐GaN substrate in an atmosphere with 10% oxygen content at 550°C for 1 h. This process induced a uniform, thin GaO_X interfacial layer on the p‐GaN surface, serving as the functional trapping layer. Monolayer MoS₂ was subsequently transferred onto the p‐GaN substrate via a PMMA‐assisted wet transfer technique. Specifically, a PMMA support layer was spin‐coated onto the MoS₂/sapphire structure, followed by chemical etching of the sapphire to release the PMMA/MoS₂ stack, which was then transferred onto the p‐GaN/Ni/Au surface. After transfer, PMMA was removed using acetone, and an optional post‐transfer annealing step was performed to improve interface quality. The MoS₂ film was patterned using RIE to define the functional device area. Finally, the n‐type contact electrode was formed by electron beam evaporation of an Au (10 nm)/Ti (3 nm)/Au (10 nm) tri‐layer stack, aligned to the MoS₂ region using a shadow mask or lithographic method. This resulted in the formation of a vertically stacked p‐GaN/monolayer MoS₂ heterojunction.

4.3. Materials Characterization and Device Measurement

Structural and compositional characterizations of the p‐GaN/n‐MoS₂ heterostructure were conducted using aberration‐corrected scanning transmission electron microscopy (STEM, Themis Z) operated at 300 kV in high‐angle annular dark‐field (HAADF) mode. Raman spectra were acquired using a WITec Alpha300R confocal microscopy system equipped with 325 and 532 nm excitation lasers. Optical illumination was provided by LED sources at ultraviolet (Thorlabs M385L3‐C1), blue (Thorlabs M470L5‐C1), green (Thorlabs M530L4‐C1), and red (Thorlabs M625L4‐C1) wavelengths. The incident optical power was measured using a Thorlabs PM100D power meter equipped with an S401C thermal sensor head. Electrical measurements, including both I‐V characteristics and transient current responses, were carried out using a Keysight 4155C Semiconductor Parameter Analyzer. Time‐resolved optical pulse inputs were modulated via an external waveform generator (RIGOL DG4162).

Following fabrication, the individual devices on the array were connected via wire bonding to a custom ceramic chip carrier. This carrier was then mounted onto a bespoke 4‐layer printed circuit board (PCB) designed specifically for this study. The PCB served as an interposer, fanning out the 65 required connections (64 individual node lines and one common ground) from the fine‐pitch bond pads to a standardized high‐density connector. This design incorporated a dedicated ground plane and guarding techniques to minimize noise and cross‐talk between channels, which is critical for resolving the low current levels (potentially down to picoamperes) involved in photocurrent and dark current measurements. The PCB was interfaced with a Keysight 4155C Semiconductor Parameter Analyzer via two Keysight B2201A multi‐channel switch matrix modules. This configuration provided the necessary 64 channels to address each device independently. The common ground of the array was connected to the 4155C's ground unit (GND), while the output of a single source measurement unit (SMU) was routed through the switch matrices to each device's node in a sequential manner. The entire setup was controlled by a PC running custom automation scripts, which synchronized the switching and measurement procedures.

The measurement of the array's static response was conducted in a two‐step process. First, the I‐V characteristics of all 64 devices were measured in complete darkness to determine the dark current and diode quality factors. Subsequently, the array was illuminated under a UV light source, and the I‐V‐t sweep was repeated to acquire the photocurrent. For both conditions, a voltage bias of ‐1 V was applied by the SMU, while the current was measured with high precision.

Funding

National Natural Science Foundation of China (Grant Nos. 92477102, 62122084) National Key Research and Development Program of China (Grant No. 2022YFB4400100) Songshan Lake Materials Laboratory 2023SLABFK09 111 Center (Grant No. B25033)

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supporting File: smll74516‐sup‐0001‐SuppMat.docx.

Data Availability Statement

All data are available in the main text or the supplementary materials. The data that support the plots within this paper and other findings of this study are available from the corresponding authors upon reasonable request.