All‐Metal‐Oxide Heterojunction Optoelectronic Synapses with Multilevel Memory for Artificial Visual Perception Applications

Abstract

Metal‐oxide semiconductor‐based optoelectronic synaptic transistors have attracted considerable attention due to their high energy efficiency and stability. This work proposes a novel WO₃/InWZnO heterojunction optoelectronic synaptic transistor, demonstrating the strong potential for emulating the human visual system. The fabricated heterojunction synaptic transistor achieves an impressive optical responsivity of 58.37 A W⁻¹ when exposed to 650 nm light. Furthermore, it successfully emulates transitions from short‐term memory (STM) to long‐term memory (LTM) by modulating pulse duration, illumination intensity, and pulse number. The heterojunction transistor also exhibits an optimal paired‐pulse facilitation (PPF) index of 176% and post‐tetanic potentiation (PTP) index of 890% under 460 nm light illumination. It also demonstrates long‐term multilevel storage capability via the photogating effect. A multilayer perceptron (MLP) model, employing synaptic weights modulated by 650, 525, and 460 nm light in the long‐term potentiation (LTP) region and by electrical stimuli in the long‐term depression (LTD) region, achieves a high recognition accuracy of 87.4% for severely distorted handwritten digits. Finally, the U‐Net architecture is adopted to evaluate the image segmentation performance through RGB channels, revealing an optimal accuracy of 74.5%, demonstrating the feasibility of the proposed WO₃/InWZnO heterojunction synaptic transistor for advanced neuromorphic vision system applications.

This study demonstrates that the WO₃/InWZnO synaptic transistor exhibits excellent photoresponsivity, long‐term multilevel conductance states, and synaptic plasticity, making it highly suitable for neuromorphic visual recognition, image processing, and segmentation applications.

1. Introduction

The implementation of neuromorphic devices has recently garnered significant attention for their ability to revolutionize computational paradigms by emulating the data transfer mechanisms in biological neural systems. In contrast, due to the separation of memory and processing units, conventional von Neumann architecture encounters critical limitations when managing massive data processing tasks, such as natural language processing and image recognition, resulting in substantial execution delays and energy inefficiency.^{[ 1 ]} To date, the emulation of human brain's computing architecture has enabled the efficient processing of large‐scale information and facilitated the development of neurorobotics and artificial sensory systems, such as self‐driving systems and medical diagnostics applications.^{[ 2 ]}

Inspired by the remarkable capabilities of human vision, optical‐driven methods have attracted significant attention due to their wide bandwidth operation, low power consumption, and minimal cross‐talk effect.^{[ 3 ]} Moreover, visible‐light‐driven optical synaptic devices have emerged as promising candidates due to their unique advantages, including reduced scattering phenomena and biocompatibility, making them feasible for advanced imaging and sensing applications. These devices hold great promise for bridging the gap between biological sensory systems and next‐generation electronic technologies.^{[ 4 ]}

Numerous studies have investigated the impact of materials on the response of optically driven artificial synapses.^{[ 5 ]} Among these studies, organic synaptic transistors have attracted considerable interest due to their unique properties, such as flexibility and solution‐processability, making them promising candidates for emerging applications, especially in wearable devices.^{[ 6 ]} In addition, 2D materials have demonstrated outstanding potential for optoelectronic synaptic transistors owing to their distinctive chemical and physical characteristics. These properties facilitate the capability for emulating distinctive synaptic plasticity for neuromorphic computing.^{[ 7 ]} However, organic synaptic transistors are more sensitive to ambient temperature and humidity variations, compromising their long‐term stability.^{[ 8 ]} Meanwhile, the fabrication of 2D material fundamentally relies on the CVD processes or mechanical exfoliation, leading to high cost and complexity.^{[ 9 ]}

With these concerns, amorphous oxide semiconductors (AOS), especially InGaZnO (IGZO), have been regarded as an optimal choice for addressing the challenges of ambient stability and large‐scale fabrication, owing to their material characteristics and simpler manufacturing processes.^{[ 10 ]} AOS‐based synaptic transistors demonstrate significant potential in advanced optoelectronic applications with inherent advantages such as high carrier mobility, excellent photo‐sensitivity, and uniform deposition ability.^{[ 11 ]} Nevertheless, the wide bandgap of AOS materials hinders their optical performance in extended spectral wavelength regions, making them more suitable for ultraviolet‐driven synaptic devices. Therefore, bandgap narrowing strategies through heterojunction structure, colloidal quantum dots (QDs) implementation, and plasma treatment are vital to extending their functionality into the visible spectrum.^{[ 12 ]} However, the integration of QDs necessitates additional solution‐based deposition and precise patterning,^{[ 12e,f ]} complicating the fabrication processes and posing significant challenges to achieving high‐density integration. In contrast, a metal‐oxide heterojunction structure offers a simplified fabrication process and improved visible‐light modulation ability owing to enhanced charge separation phenomenon.^{[ 13 ]} As a result, IGZO‐based metal‐oxide heterojunction transistors stand out as an excellent candidate for constructing optical synapses due to their exceptional optoelectronic characteristics.^{[ 14 ]} However, gallium (Ga) is a rare element on Earth, leading to an urgent need for alternative materials to mitigate resource depletion. Fortunately, tungsten (W) is an effective carrier suppressor due to its strong bonding affinity with oxygen, which helps suppress the formation of excess oxygen vacancies and enhances the stability of the oxide film. Accordingly, tungsten‐doped InZnO (IWZO) is identified as an auspicious channel material for thin‐film transistor devices, and it has also demonstrated superior electrical performance in our previous work.^{[ 15 ]} On the other hand, WO₃ serves as a highly optically sensitive material, employed extensively in electrochromic applications,^{[ 16 ]} further emphasizing the significance of WO₃/IWZO integration for advanced light‐sensing devices. Additionally, only a few studies have explored WO₃/IWZO‐based optoelectronic synaptic devices, accentuating their potential for further investigation.

In this work, a novel heterojunction synaptic transistor based on crystalline WO₃ (c‐WO₃) and amorphous InWZnO (a‐IWZO) was proposed, demonstrating exceptional feasibility for visible‐light‐driven performance. The proposed devices successfully emulated key synaptic behaviors, including short‐term memory (STM), long‐term memory (LTM), paired‐pulse facilitation (PPF), and post‐tetanic potentiation (PTP). Furthermore, multilevel data storage states were achieved by applying negative gate voltages, which is rarely reported in AOS‐based synaptic transistors. The devices also exhibited high linearity in synaptic weight modulation when stimulated by visible light and electrical signals, enabling their application in advanced tasks such as RGB pattern recognition in challenging scenarios. Finally, an image segmentation system was simulated and implemented, paving the way for the integration of optoelectronic functionalities into advanced artificial visual perception systems.

2. Results and Discussion

2.1. Devices’ Structure and Electrical Performance

Figure 1a depicts the schematic of a human object perception process, which serves as the inspiration for designing heterostructure devices (Figure 1b) that emulate human visual functions. The devices incorporating a c‐WO₃ layer were denoted as WIST, while those without a c‐WO₃ layer were referred to as IST. Electrical measurements under drain voltage (V_D) of 0.1V, and 10 V were conducted and presented in Figure 1c to evaluate the fundamental electrical characteristics of the proposed devices. The results reveal that WIST devices incorporating an additional c‐WO₃ layer exhibit enhanced electrical performance compared to IST devices, which is attributed to a more efficient electron transfer mechanism within the WO₃ channel.^{[ 17 ]} In addition, the electrical parameters, including threshold voltage (V_th), subthreshold swing (SS), field effect mobility (µ_FE), and on‐off current ratio for WIST and IST are extracted and summarized in Table 1 , further substantiating the performance improvement in the proposed WIST devices. The transmittance spectra of c‐WO₃, a‐IWZO, and c‐WO₃/a‐IWZO films are shown in Figure 1d, while Figure 1e–g compares the optical bandgaps of a‐IWZO (3.23 eV), c‐WO₃ (2.8 eV), and the c‐WO₃/a‐IWZO heterostructure (2.87 eV), highlighting the tunable optical bandgap through heterostructure design.

Figure 1.

a) Schematic of visual perception of the human brain. b) Proposed WIST devices for emulating human visual perception functionality. c) Transfer characteristics of WIST and IST devices under V_D of 0.1 and 10 V with channel width/length of 500/50 µm. d) Optical transmittance of a‐IWZO, c‐WO₃, and c‐WO₃/a‐IWZO films. Optical absorption characteristics and bandgap extraction for e) a‐IWZO, f) c‐WO₃, and g) c‐WO₃/a‐IWZO films.

Table 1.

Extracted electrical parameters of WIST and IST devices.

Parameter	Vth (V)	SS (V dec−1)	µFE (cm2 V−1 s−1)	On‐off current ratio (VD = 0.1V)
Sample
WIST	−9.8	0.59	4.9	2.6 × 105
IST	−22.3	0.63	4.5	3.7 × 104

The surface morphology analysis in Figure 2a,b compares the roughness of IWZO films annealed at 300 °C for 1.5 h and WO₃ films annealed at 600 °C for 1 h. The annealed WO₃ films exhibit a high surface roughness (Rq: 5.08 nm, Ra: 4.13 nm), attributed to the crystallization. As shown in Figure 2c, additional X‐ray diffraction (XRD) measurements validate the occurrence of crystallization when the annealing temperature exceeds 500 °C. The intensity of the (002) major diffraction peak suggests a preferential crystallization growth along the [002] direction, and the grain size was estimated for different temperatures in Table S1 (Supporting Information) through the Scherrer equation.^{[ 18 ]} Furthermore, Figure 2d,e demonstrate the transmission electron microscopy (TEM) image and energy dispersive spectroscopy (EDS) mapping of the proposed WIST devices, visualizing the uniform heterointerface between the c‐WO₃ and a‐IWZO layers.

Figure 2.

a) Surface morphology of 300 °C IWZO. b) Surface morphology of 600 °C WO₃. c) XRD measurements of WO₃ films from 400 to 600 °C. d) Cross‐sectional TEM image of the c‐WO₃/a‐IWZO heterojunction. e) EDS elemental mapping of O, W, In, and Zn in the c‐WO₃/a‐IWZO heterojunction structure.

2.2. Devices’ Optoelectronic Performance

The superior optical responses of WIST devices under varying spectral wavelengths and intensities are comprehensively illustrated in Figure 3a–c. The incorporation of crystalline WO₃ film, which exhibits a lower density of gap states,^{[ 19 ]} enhances charge separation efficiency and renders it suitable as a light‐absorbing layer for improving photoelectric conversion efficiency in optoelectronic devices.^{[ 20 ]} Under 650, 525, and 460 nm light illumination with varied intensities (0.128, 0.254, and 0.301 mW cm⁻²) and fixed V_D (10 V), the WIST devices demonstrate a significant negative V_th shift with increasing illumination intensity. In contrast, IST devices also show minimal response under the same conditions, which can be attributed to the plasma process on the IWZO layer, as shown in Figure 3d–f. Notably, the WIST samples show a significant increase in the off‐state drain current region under the optical illumination. This enhancement in the off‐state region, identified as the light‐induced current, is mainly due to the increased concentration of photo‐generated carriers in the WO₃ region, which features a smaller bandgap. The unique combination of c‐WO₃ and a‐IWZO forms an efficient electron transport pathway, thereby enhancing carrier mobility and facilitating rapid charge separation. This synergistic structure enables WIST devices to achieve high photo‐response and ensures exceptional electron transport performance.^{[ 21 ]}

Figure 3.

Transfer characteristics of a) WIST d) IST devices under 650 nm light illumination with intensity of 0.128 to 0.301 mW cm⁻² at V_D = 10 V. Transfer characteristics of b) WIST e) IST devices under 525 nm light illumination with intensity of 0.128 to 0.301 mW cm⁻² at V_D = 10 V. Transfer characteristics of c) WIST f) IST devices under 460 nm light illumination with intensity of 0.128 to 0.301 mW cm⁻² at V_D = 10 V.

Additionally, key optical parameters, including optical responsivity (R), signal‐to‐noise ratio (SNR), and specific detectivity (D*) under 650, 525, and 460 nm wavelength illumination with V_D fixed to 10 V were systematically analyzed and summarized in Table 2 . Notably, the WIST devices demonstrated an impressive responsivity of 58.37 A W⁻¹ under 650 nm illumination. Furthermore, the proposed WIST devices exhibited superior optical responsivity to long‐wavelength light, achieving the highest performance among previously reported studies, as shown in Table 3 . These improvements highlight the potential of WIST devices to replicate human visual perception, owing to their remarkable sensitivity to visible light.

Table 2.

Optical parameters of proposed WIST devices under wavelengths of 650, 525, and 460 nm.

Light Source (0.301µW cm−2)	Optical responsivity (A W−1)	SNR	Detectivity (Jones)
λ  =  650 nm	58.37	188.74	4.52 × 1012
λ  =  525 nm	132.18	515.46	9.81 × 1012
λ  =  460 nm	676.89	3.4 × 103	4.67 × 1013

Table 3.

The proposed WIST devices in this work as compared to the reported work.

Material	Wavelength	Optical responsivity (A W−1)	SNR	Detectivity (Jones)
CsPbIxBr3‐x/IGZO[ 39a ]	λ  =  635 nm	26.48	8.71 × 106	8.42 × 1014
Perovskite/IGZO[ 39b ]	λ  =  620 nm	0.55	NA	5.40 × 1014
SnO/IGZO[ 29 ] Previous work	λ  =  640 nm	3.26	4.4 × 104	NA
c‐WO3/a‐IWZO (VD = 10V) This work	λ  =  650 nm	58.37	188.74	4.67 × 1013

2.3. Emulation of Visual Perception via RGB Light Stimulation

In biological neural networks, short‐term plasticity enables transient adjustments in synaptic strength over shorter time intervals. In contrast, long‐term plasticity supports sustained levels of synaptic strength, playing a crucial role in long‐term memory. Figure 4 illustrates the versatile synaptic plasticity performed by the proposed WIST devices under various spectral wavelengths of light stimuli (λ = 650, 525, and 460 nm). Simultaneously, the gate voltage (V_G) of −3 V and the V_D of 0.1 V remained constant throughout the measurements. With a longer pulse duration and fixed intensity (0.128 mW cm⁻²), the excitatory postsynaptic current (EPSC) demonstrates a progressive increment, followed by an exponential decay once the light is removed. Moreover, the employment of shorter spectral wavelengths (λ = 650 nm to λ = 460 nm) and longer spike duration (0.5 s to 2 s) led to an elevated enhancement of EPSC from 3.01 nA to 3.23 nA (λ = 650 nm), 3.19 nA to 3.65 nA (λ = 525 nm), and 7.93 nA to 19.67 nA (λ = 460 nm) as shown in Figure 4a. The enhancements are also accompanied by the significantly prolonged decay time constant from 6.34 to 21.01 s. These enhancements are attributed to the increased photo‐generated carriers induced by shorter‐wavelength light irradiation.

Figure 4.

a) EPSC stimulated by different pulse durations of the optical spike (0.5 s, 1 s, and 2 s) under 650, 525, and 460 nm light illumination with the fixed intensity of 0.128 mW cm⁻². b) EPSC stimulated by different numbers of optical spikes (5, 10, 15, and 20) under 650, 525, and 460 nm light illumination with fixed intensity of 0.128 mW cm⁻². c) EPSC stimulated by different intensities of 650, 525, and 460 nm light illumination (0.254 and 0.301 mW cm⁻²). PPF and PTP index of proposed WIST devices stimulated by successive optical illumination of d) 650 nm, e) 525 nm, and f) 460 nm light illumination with different time intervals. Optical potentiation performed by g) 650 nm, h) 525 nm, and i) 460 nm light illumination with intensity and frequency of (0.254 mW cm⁻², 1 Hz) and electrical depression conducted by electrical with amplitude and frequency of (6.5 V, 1 Hz), denoted as R, G, and B schemes.

Furthermore, spike‐number‐dependent plasticity is demonstrated in Figure 4b. As the number of input spikes increases from 5 to 20 with a fixed frequency of 0.5 Hz and intensity of 0.128 mW cm⁻², the EPSC exhibits a corresponding increase from 3.16 to 3.42 nA (λ = 650 nm), 3.35 nA to 4.03 nA (λ = 525 nm), and 11.79 to 26.11 nA (λ = 460 nm), reflecting the cumulative effect of successive spikes on the synaptic responses. This behavior highlights the WIST devices’ capability to emulate biological synapses, wherein the synaptic strength is dynamically modulated by the number of input spikes. Notably, in Figure 4c, the sustainability of the EPSC of WIST devices following consecutive optical pulses under different spectral wavelengths (λ = 650, 525, and 460 nm) and intensities (0.254 and 0.301 mW cm⁻²) is shown, which can be attributed to their unique structural properties. The c‐WO₃ layer enhances the transfer efficiency of photo‐excited carriers to the c‐WO₃/a‐IWZO interface under the negative biased gate voltage. Meanwhile, the heterointerface served as a defect‐rich region with a higher density of shallow and deep defects, facilitating charge storage due to the inherent properties of the amorphous IWZO film. The trapped carriers gradually recombine, resulting in an extended decay time.^{[ 22 ]} This slow decay characteristic is beneficial for synaptic devices, as it facilitates the emulation of long‐term synaptic behavior, which is crucial for mimicking memory retention in neuromorphic systems.^{[ 23 ]}

Meanwhile, the PPF and PTP dynamics were also examined to evaluate the short‐term synaptic plasticity of the WIST devices. These indices reflect that the peak intensity of the postsynaptic current triggered by the n^th optical spike (A_n ) exceeds that induced by the initial optical spike (A ₁). This phenomenon is ascribed to the residual photo‐generated carriers that persist after the first light stimulation. The corresponding PPF and PTP indices are defined by the following equations:^{[ 24 ]}

$$\mathrm{PPF}=\left({\mathrm{A}}_2/{\mathrm{A}}_1\right)\times 100\%$$ (1)

$$\mathrm{PTP}=\left({\mathrm{A}}_{10}/{\mathrm{A}}_1\right)\times 100\%$$ (2)

The PPF index quantifies the degree of synaptic behavior enhancement in response to two consecutive stimuli. A higher PPF index indicates effective facilitation, demonstrating the devices’ ability to mimic biological synapses under repetitive stimulation. In this study, the proposed WIST devices exhibit outstanding short‐term synaptic plasticity, with PPF indices ranging from 160% (λ = 650 nm) to 162% (λ = 525 nm) and up to 176% (λ = 460 nm). In addition, they demonstrate pronounced PTP indices of 414% (λ = 650 nm) to 580% (λ = 525 nm) and up to 890% (λ = 460 nm). These measurements were conducted with inter‐pulse intervals (Δt) varying from 1 s to 10 s, under an illumination intensity of 0.128 mW cm⁻² and a pulse duration of 0.5 s. As illustrated in Figure 4d–f, these results showcase the superior facilitation capability of the WIST devices.

As Δt increases, the PPF index decreases, exhibiting a decay pattern similar to the kinetics observed in biological synapses, which can be described by a double exponential decay, as shown in the following formula:^{[ 25 ]}

$$\mathrm{PPF}={\mathrm{C}}_0+{\mathrm{C}}_1\exp \left(-\frac{\mathrm{\Delta }\mathrm{t}}{{\tau }_1}\right)+{\mathrm{C}}_2\exp \left(-\frac{\mathrm{\Delta }\mathrm{t}}{{\tau }_2}\right)$$ (3)

C ₀, C ₁, and C ₂ represent the initial facilitation amplitudes, while τ₁ and τ₂ correspond to the characteristics of rapid and slow relaxation conditions, respectively. The extracted time constants (τ₁, τ₂) were determined to be (0.89 s, 10.21 s), (0.92 s, 10.81 s), and (0.95 s, 11.21 s) for λ = 650, 525, and 460 nm, respectively. These values indicate a rapid initial decay followed by a slower decline, which is crucial for temporal information processing in neuromorphic systems.

To emulate the human vision system, the linearity of conductance states in the RGB channels of WIST devices are pivotal parameters for their applicability in neuromorphic systems. As a result, RGB light illuminations were applied to assess the conductance modulation linearity through the following equation,^{[ 26 ]} while the V_G and V_D were fixed at −3 and 0.1 V, respectively.

$${\mathrm{G}}_{\mathrm{LTP}}=\mathrm{B}\left\{1-\exp \left(-\mathrm{P}/\mathrm{A}\right)\right\}+{\mathrm{G}}_{\min }$$ (4)

$${\mathrm{G}}_{\mathrm{LTD}}=-\mathrm{B}\left\{1-\exp \left(\frac{\mathrm{P}-{\mathrm{P}}_{\mathrm{MAX}}}{\mathrm{A}}\right)\right\}+{\mathrm{G}}_{\mathrm{MAX}}$$ (5)

$$B=\left(G_{MAX}-G_{min}\right)/\left\{1-\exp \left(\frac{-P_{MAX}}{A_{P,D}}\right)\right\}$$ (6)

G_LTP and G_LTD represent the conductance in the long‐term potentiation (LTP) and long‐term depression (LTD) regions, respectively. P and P_MAX denote the number of applied pulses and the total number of pulses. The maximum and minimum conductance values are presented as G_MAX and G_min , respectively. The A parameter reflects the nonlinearity (NL) in both the LTP and LTD regions. Meanwhile, the ratio of G_MAX to G_min is a crucial parameter in electronic memory and neuromorphic computing applications. A high G_MAX /G_min ratio enables clear identification of different conduction states, thereby enhancing data storage density and computational precision.^{[ 27 ]}

To evaluate the conductance modulation capability under the RGB schemes, the V_D was fixed at 0.1 V throughout the measurements. In the R scheme (Figure 4g), optical potentiation (650 nm, 0.254 mW cm⁻², 1 Hz) and electrical depression (6.5 V, 1 Hz), corresponding to the LTP and LTD regions, respectively. This approach yielded NL factors of 0.089 for the potentiation region (NL_LTP) and 0.106 for the depression region (NL_LTD) with a G_MAX /G_min ratio of 1.29. In the G scheme (Figure 4h), optical potentiation was performed under the conditions of (λ = 525 nm, 0.254 mW cm⁻², 1 Hz), while electrical depression was maintained at (6.5 V, 1 Hz). The corresponding nonlinear factors (NL_LTP, NL_LTD) were determined to be (0.132, −0.4), with a G_MAX /G_min ratio of 1.38. For the B scheme (Figure 4i), the NL parameters were evaluated as (0.102, 0.32), while the G_MAX /G_min ratio shifted to 1.84.

Notably, the endurance assessment under illumination at different spectral wavelengths (650, 525, and 460 nm) reveals consistent and stable synaptic responses over 1200 consecutive pulses, as illustrated in Figure S16 (Supporting Information). Additionally, the R scheme demonstrates remarkable temporal stability over a one‐month period, as evidenced by consistent electrical performance shown in Figure S12 (Supporting Information). Moreover, the learning and relearning capabilities of the WIST devices under these spectral wavelengths are clearly demonstrated in Figure S13 (Supporting Information). These results underscore the robust reliability of synaptic weight modulation and the spectral adaptability of the proposed WIST devices, highlighting their strong potential for next‐generation neuromorphic visual perception systems.

2.4. The Behavior of Multilevel Memory States

Furthermore, the WIST devices exhibit superior long‐term multilevel memory behavior under V_G and V_D of −5 and 0.1 V, as shown in Figure 5a. After being programmed with a series of write (525 nm, 0.301 mW cm⁻², 0.5 Hz) and erase (6.5 V, 0.5 Hz) signals, the devices exhibited negligible EPSC decay, indicating excellent memory retention capability. Furthermore, the multilevel data storage capability (4 states) of proposed WIST devices is exhibited in Figure 5b. This phenomenon was performed after 20 continuous optical programming pulses with spectral wavelengths of 650, 525, and 460 nm under identical conditions. Subsequently, the corresponding reliability assessment for each spectral scheme is exhibited in Figures S17 and S18 (Supporting Information). This stability evaluation validates the feasibility of long‐term multistate data storage within a single WIST transistor, contributing to improved memory density and operational efficiency for next‐generation artificial visual perception systems.^{[ 28 ]}

Figure 5.

a) The long‐term multilevel memory behavior was evaluated under optical writing conditions (525 nm, 0.301 mW cm⁻², 0.5 Hz) with 20 pulses, followed by electrical erasing (6.5 V, 0.5 Hz) with 10 pulses. b) Long‐term capability of 4 memory states after being programmed by consecutive 20 optical pulses of 650, 525, and 460 nm light illumination with intensity and frequency of (0.301 mW cm⁻², 0.5 Hz). c) Diagram of relative formation energy from oxygen vacancy state to oxygen lattice state. d) Probable underlying mechanism for synaptic (V_G = −3 V), memory (V_G = −5 V), and reset mode (V_G = 6.5 V).

The potential underlying mechanism of long‐term multilevel data storage states is proposed based on the probable energy band scenario demonstrated in Figure 5d. Due to the negative gate bias, the conduction and valance bands of the c‐WO₃ layer bend downward.^{[ 29 ]} This band alignment induces electron accumulation near the c‐WO₃/a‐IWZO heterointerface, while photo‐generated holes are gathered in the c‐WO₃/SiO₂ interface. When a moderate negative gate voltage (V_G = −3 V) is applied, photo‐generated electrons migrate toward the IWZO layer, while the holes accumulate near the WO₃/SiO₂ interface, which is denoted as synaptic mode in Figure 5d. In memory mode (V_G = −5 V), a decreased gate voltage enhances the band bending effect, thereby increasing the probability of photo‐generated holes accumulating and being trapped at the c‐WO₃/SiO₂ interface. Furthermore, the application of a stronger negative bias further restricts the release of trapped holes from the trap states, thereby suppressing the carrier recombination rate and inducing a pronounced photogating effect.^{[ 30 ]} This phenomenon enhances photocurrent retention and facilitates multistate data storage functionality, contributing to the unique optoelectronic properties of the WIST heterostructure devices.^{[ 31 ]} In reset mode (V_G = 6.5 V), applying a positive gate pulse triggers the release of trapped holes, thereby promoting electron‐hole recombination and resetting the device to its initial state.

To further elucidate the mechanism underlying the multistate data storage behavior of WIST devices, the temperature‐dependent decay of photocurrent was systematically investigated. The photocurrent decay is intrinsically linked to the recombination mechanism of photo‐generated carriers, which are directly associated with the neutralization of ionized oxygen vacancies. This recombination process is thermally activated, a specific activation energy is essential for its occurrence. Figure 5c systematically illustrates the energy barrier associated with the neutralization process from oxygen vacancies to oxygen lattices. Within the c‐WO₃/a‐IWZO heterojunction structure, the larger activation energy is required to neutralize ionized oxygen vacancies, enhancing memory behavior effectively.^{[ 32 ]} Therefore, the temperature dependence of the decay time constant exhibited by WIST devices should be carefully examined.

The temperature varied from 298 to 313 K, and the drain current stimulated by consecutive 650 nm light illumination follows the dynamics of Kohlrausch stretched exponential function as the following equation:^{[ 33 ]}

$${\mathrm{I}}_{\mathrm{D}}\left(\mathrm{t}\right)={\mathrm{I}}_{\mathrm{D}}\left(0\right)\exp \left(-{\left(\mathrm{t}/\tau \right)}^{\beta }\right)$$ (7)

I_D is the drain current, t is the time, τ is the relaxation time constant, and β is the stretching exponent. The decay of photocurrent was measured and shown in Figure S10a (Supporting Information). Furthermore, the logarithmic relaxation time constants at various temperatures were also fitted in Figure S10b (Supporting Information). The neutralization activation energy was calculated using the Arrhenius equation given below:^{[ 34 ]}

$$\tau =\mathrm{Aexp}\left(\frac{-{\mathrm{E}}_{\mathrm{a}}}{\mathrm{kT}}\right)$$ (8)

A represents the pre‐exponential factor, E_a denotes the activation energy, k is Boltzmann's constant, and T is the absolute temperature. According to this relationship, the activation energy is determined by analyzing the slope of ln(τ) as a function of the inverse absolute temperature. The calculated activation energy required for the neutralization of ionized oxygen vacancies in the WIST devices is ≈630 meV, which is higher than the previously reported value,^{[ 35 ]} promoting superior long‐term memory performance.

2.5. Learning and Forgetting Behavior of a 2 × 2 Array of WIST Devices

Most of the external information can be perceived by the visual perception system. The human retina can detect the incident light of different spectral wavelengths, each carrying distinct information, which is converted into neural impulses and transmitted to the visual cortex, where it is further processed and stored within synapses.^{[ 36 ]} The proposed WIST devices enable real‐time optical signal sensing and processing, enhancing power efficiency and minimizing signal interference during transmission. Furthermore, the WIST devices exhibit significant photoresponsivity and long‐term memory characteristics under visible light illumination, thereby facilitating the emulation of human perceptual functions for colored image recognition. In this regard, a 2 × 2 array of WIST devices was proposed as an imaging chip to emulate the human visual perception system, demonstrated in Figure 6a. After stimulating by 3 cycles of 5 optical spikes with intensity and frequency of (0.254 mW cm⁻², 0.5 Hz), EPSC exhibits varied values by different spectral wavelengths as demonstrated in Figure 6b. The EPSC stimulated by 460 nm light shows the highest value of 23.4 nA, while that excited by 650 nm light exhibits a value of 3.9 nA.

Figure 6.

a) Schematic diagram of a 2 × 2 array based on WIST devices for mimicking visual perception functionality in colored image processing. b) EPSC responses under different wavelengths and learning spikes. c) EPSC variations of 2 × 2 array during learning (upper panel) and forgetting (lower panel) processes after three learning and three forgetting states under 650, 525, and 460 nm light illumination with fixed condition of (0.254 mW cm⁻², 0.5 Hz).

As illustrated in Figure 6c, darker regions correspond to higher photocurrent levels, whereas lighter regions indicate weaker photocurrent responses. When stimulated by different learning spikes and spectral wavelengths, the 2 × 2 WIST array undergoes multiple learning processes. As the number of optical spikes increases, the color gradually darkens, mimicking the dynamic learning behavior of the human brain. In contrast, the color of the pixels remains nearly unchanged after removing the light pulses for 40, 80, and 160 s, indicating the strong memory characteristics of the WIST devices. Furthermore, owing to the inherent scalability of the WIST devices, larger arrays (e.g., 16 × 16 or 32 × 32) can be readily realized using the same device configuration and optical stimulation strategy. Collectively, these features highlight the great potential of WIST arrays for mimicking retinal perception and image memory, representing a significant step toward developing advanced neuromorphic vision systems.

2.6. Multilayer Perceptron (MLP) Structure for Blurred Handwritten Recognition

The artificial neuron (perceptron) is a fundamental building block of artificial neural networks. It was introduced by Frank Rosenblatt in 1958 as a simple linear classifier. The perceptron processes weighted inputs and transmits the computed result to the output layer. If the sum exceeds a predefined threshold, the perceptron fires and outputs 1; otherwise, it outputs 0.^{[ 37 ]} Figure S14 (Supporting Information) illustrates the fundamental perceptron unit for realizing AND and OR operations by applying sequential input digits and bias (V_G). In this framework, WIST devices serve as core computational elements, dynamically adjusting the summation based on input digits and applied bias to deliver the corresponding output. This result highlights the versatility of the perceptron model in executing fundamental logic operations within a neuromorphic computing paradigm.

To further validate the applicability of the proposed WIST devices in more complex tasks, handwritten digit recognition tests were conducted, as shown in Figure 7a. A four‐layer multilayer perceptron (MLP) model was implemented, utilizing the MNIST dataset to assess the feasibility of WIST synaptic devices for handwritten recognition. The MLP model comprises an input layer of 784 neurons, followed by two hidden layers. The first hidden layer comprises 256 neurons, while the second contains 128 neurons. Finally, the output layer includes 10 neurons, corresponding to the classification results. In order to simulate real‐world conditions and evaluate devices’ performance under image distortions, the MNIST dataset was modified by introducing Gaussian, Pepper and salt, and Stripe noise. These perturbations were designed to test the accuracy and robustness of the WIST synaptic devices in processing blurred and noisy images, further demonstrating its potential for practical implementation in artificial neural networks. By leveraging the high linearity achieved through RGB schemes, the results indicate that under Gaussian noise conditions, the recognition accuracy reached optimal levels, achieving 94.2%, 91.2%, and 91.4% for the R, G, and B schemes, respectively, as illustrated in Figure 7b. These findings highlight the robust performance and high accuracy of the proposed WIST synaptic devices in real‐world color perception scenarios, demonstrating its feasibility for emulating the human visual system.

Figure 7.

a) Handwritten recognition test through RGB channels by introducing different types of noise (Gaussian, Pepper and salt, Stripe) performed by four layers of MLP structure. Handwritten recognition accuracy and corresponding confusion matrixes of b) Gaussian, c) Pepper and salt, d) Stripe noise through RGB channels.

2.7. Image Segmentation Simulation

Furthermore, the demand for accurate and reliable image segmentation across RGB channels is crucial for enhancing the efficiency and safety of vision systems. In this work, the U‐Net architecture was employed to validate the application of image segmentation using the proposed WIST devices under RGB‐scheme stimulation. As illustrated in Figure 8a, U‐Net consists of an encoder and a decoder, which are connected by skip connections. The encoder is responsible for extracting features of the input image through repeated processing of convolutional kernels, followed by pooling layers to minimize the spatial dimensions. The decoder progressively restores the spatial features using transposed convolution kernels, which enables precise localization of images. Additionally, skip connections between corresponding encoder and decoder layers allow the direct transfer of high‐resolution features, improving segmentation accuracy and preserving detail features.^{[ 38 ]} The complete explanation of the U‐Net structure utilized in this work is elucidated in Figure S19 (Supporting Information). The input images, containing RGB channels, are first flattened and processed by the encoder to extract key image features. The decoder then reconstructs and outputs the segmented regions, as shown in Figure 8b. Despite the presence of blurring, as depicted in Figure 8b, the segmentation result in Figure 8c demonstrates that the segmentation process remains effective, achieving high accuracies of 74.3%, 71.4%, and 74.5% for the R, G, and B schemes, respectively, after 10 epochs. These values closely approximate the ideal U‐Net accuracy of 74.6%, underscoring the superior performance of the proposed WIST devices. Figure 8d exhibits a relevant hybrid optical‐electronic computing architecture for realizing image segmentation tasks in artificial neural networks. The optical and electrical inputs lead to an increased or decreased current, and the resulting change in the devices’ conductance sensed by the readout ADC is fed back to the weight parameter update module, modifying the following input accordingly. This synergy enables applications in image segmentation by dynamically adjusting weights to enhance feature extraction and boundary detection, which plays a fundamental role in providing precise pixel‐level boundaries and classifications essential for advanced object‐tracking applications.^{[ 2 ]} These achievements collectively contribute to developing a vision system that closely emulates the human visual perception system. Compared to the previous studies, as summarized in Table 4 , this work exhibits more diverse synaptic behaviors and a wider range of applications, especially in achieving long‐term multilevel conductance states and enabling the feasibility of advanced visual systems. These attributes pave the way for next‐generation artificial intelligence and advanced vision‐based technologies.

Figure 8.

a) Structure of proposed U‐Net architecture and input image with RGB channels. b) Segmentation masks (top) and corresponding predicted results (bottom) c) Accuracy of image segmentation by proposed WIST devices via RGB channels and ideal conditions. d) Schematic of image segmentation realized by WIST devices via RGB optical and electrical input.

Table 4.

Performance comparison between this work (WIST) and previously reported studies.

Material	Drain voltage (Read voltage)	PPF / PTP index	Nonlinearity (NLLTP/NLLTD)	Wavelength of incident light	Long‐term multilevel conductance state	Application
IGZO/Ag2O[ 40 ]	0.5 V	～180%/—	—	830–405 nm	—	Learning and relearning
IGZO (absorption layer)[ 3a ]	NA	～200%/—	—	635 nm	—	Pavlov's classical conditioning
PP/SnOx/IGZO[ 41 ]	0.1 V	～160%/—	—	650–400 nm	—	—
In2O3/ZnO[ 42 ]	1 V	～113%/—	3.18 / 3.09	405 nm	—	Handwritten recognition
QDs/SnO2 [ 43 ]	0.1 V	～180%/—	—	520 nm	—	Handwritten recognition
ZnO[ 44 ]	3 V	～114%/—	—	355 nm	—	Learning and relearning
ITO[ 45 ]	5 V	146% /～550%	—	395 nm	—	Handwritten recognition
This work (WIST)	0.1 V	～180%/890% (B scheme)	0.102 / 0.32 (B scheme)	650–460 nm	Y	Handwritten recognition, image segmentation

3. Conclusion 

We have successfully developed a novel metal oxide‐based c‐WO₃/a‐IWZO heterojunction optoelectronic synaptic transistors with enhanced spectral response range and photoresponsivity within the visible‐light region. By utilizing the 650, 525, and 460 nm light‐driven conditions, we demonstrated various forms of synaptic plasticity capabilities, including a remarkable PPF index of 160%, 162%, and 176% under different spectral wavelengths. The long‐term multilevel memory functionality was successfully demonstrated through the photogating effect under 650, 525, and 460 nm light illumination. The proposed WIST devices represent a significant advancement in neuromorphic computing, progressing from single‐perceptron operations to complex multilayer perceptron architectures while maintaining exceptional resilience under extreme environmental conditions. The devices achieved a high recognition accuracy of 87.4% under the most challenging testing conditions in handwritten digit recognition simulations. Additionally, the WIST devices were successfully applied to an image segmentation system, demonstrating an optimal accuracy of 74.5%, approaching the ideal value. These outstanding results underscore the potential of WIST devices for advanced neuromorphic computing applications and highlight their integration potential in vision‐based systems, paving the way for next‐generation adaptive vision technologies.

4. Experimental section

Devices Fabrication

Thin film transistor (TFT) devices were fabricated on a highly doped Si wafer covered with a 100 nm thermally grown SiO₂ layer as the buffer oxide. First, a 6 nm WO₃ thin film was deposited using RF sputtering with a power of 50 W. The deposition utilized a pure WO₃ target under an Ar/O₂ gas mixture of 30 sccm/0 sccm. Following the deposition, the WO₃ film was annealed in a horizontal furnace at 600 °C for 1 h in an oxygen atmosphere. Subsequently, an IWZO layer was deposited using a target of IWZO (WO₃ 4wt.%) on the annealed WO₃ film for 6 nm by RF sputtering under the power of 50 W and Ar/O₂ gas mixture of 28 sccm/2 sccm, forming a heterojunction structure. The entire devices were then annealed in a furnace for 1.5 h in an oxygen atmosphere. After the annealing process, the source and drain electrodes were formed by a DC‐sputtered 26 nm‐thick Mo metal layer at a power of 30 W, ensuring a low work function and highly conductive metal contact.^{[ 29 ]} Finally, the 20 s plasma treatment was conducted on the devices to enhance the optical response. All devices were patterned using the shadow mask. Devices incorporated with the WO₃ layer were denoted as WIST, while those without the WO₃ layer were labeled as IST for comparison.

Characterizations and Electrical Measurement

The optical transmittance of the IWZO films was characterized using an ultraviolet‐visible (UV‐vis) spectrophotometer (JASCO 750). X‐ray photoelectron spectroscopy (XPS, Thermo Fisher Scientific Theta Probe) was utilized to analyze the chemical composition of the deposited films, with the obtained binding energies calibrated using the C 1s peak at 284.8 eV. Additional characterizations, including TEM (Jeol JEM‐F200), EDS, XRD (PANalytical X'Pert Pro (MRD)), and AFM (Bruker Dimension Icon Scanning Probe Microscope) were employed to further examine the material properties and structure of the deposited films.

Electrical measurements were performed at room temperature under ambient air conditions, with device dimensions of channel width (W) and channel length (L) fixed to 500 and 50 µm, respectively. The field effect mobility (µ_FE) was extracted from the transconductance at a drain current in the linear region (V_D = 0.1 V), and the threshold voltage (V_th) was defined by the gate voltage at a drain current of 10 nA. Optical illuminations at spectral wavelengths of 650, 525, and 460 nm were provided by LED sources with different intensities of 0.128, 0.254, and 0.301 mW cm⁻², while electrical pulses were applied through a Keithley 4225 PMU unit, with measurements conducted via a Keithley 4200 SCS parameter analyzer. Finally, the optical responsivity (R), signal‐to‐noise ratio (SNR), and specific detectivity (D*) under different spectral wavelength illumination and intensity were examined at subthreshold region (V_G = −10 V for WIST, V_G = −20 V for IST). These measurements were performed at the V_D of 10 V through the following equation:^{[ 39 ]}

$$\mathrm{R}=\frac{{\mathrm{J}}_{\mathrm{ph}}}{{\mathrm{P}}_{\mathrm{in}}}=\frac{{\mathrm{J}}_{\mathrm{photo}}-{\mathrm{J}}_{\mathrm{dark}}}{{\mathrm{P}}_{\mathrm{in}}/{\mathrm{A}}_{\mathrm{LED}}}$$ (9)

$$\mathrm{SNR}=\frac{{\mathrm{I}}_{\mathrm{photo}}-{\mathrm{I}}_{\mathrm{dark}}}{{\mathrm{I}}_{\mathrm{dark}}}$$ (10)

$${\mathrm{D}}^{\ast }=\frac{\mathrm{R}}{\sqrt{2\mathrm{q}{\mathrm{J}}_{\mathrm{dark}}}}$$ (11)

These three equations describe key performance for optoelectronic synaptic devices. Responsivity quantifies the signal conversion from incident optical illumination into electrical photocurrent, indicated by the net photocurrent density (J_photo − J_dark ) over the incident power density (P_in /A_LED ). J_photo is the photocurrent density under optical illumination, J_dark represents the dark current density, P_in indicates the power of optical input, and A_LED is the illuminated area of the optoelectronic synaptic devices. Subsequently, signal‐to‐noise ratio implies the strength of the photocurrent related to the dark current, indicating the devices’ ability to distinguish light‐induced signals (I_photo − I_dark ) from dark current (I_dark ). I_photo represents the photocurrent, while I_dark indicates the dark current. Finally, specific detectivity evaluates the optoelectronic synaptic devices’ sensitivity by normalizing responsivity with respect to noise, where q is the elementary charge.

The synaptic devices were irradiated with optical illumination at spectral wavelengths of 650, 525, and 460 nm and stimulated by electrical pulses with different amplitudes. Various stimuli were applied to demonstrate different synaptic characteristics of proposed WIST devices.

Handwritten Recognition and Image Segmentation Simulation

The dataset of handwritten recognition simulation was utilized from the Modified National Institute of Standards and Technology (MNIST). In the handwritten digit recognition test, the dataset size, distribution, and preprocessing steps applied to the training and validation sets were specified. During the training process, a customized optimizer was employed to adaptively adjust weight updates according to the conditions of the R, G, and B schemes.

To evaluate the model's generalization ability under different noise conditions, various types of noise, such as Gaussian, salt‐and‐pepper, and stripe noise were introduced, simulating real‐world visual perception systems. These noise perturbations were incorporated to ensure the robustness and adaptability of the proposed MLP model. Furthermore, dropout and regularization techniques were applied to mitigate the risk of overfitting and enhance the model's generalization capability. Specifically, dropout was employed to randomly deactivate a fraction of neurons during training, preventing the model from relying on specific features. Additionally, regularization was implemented to constrain the magnitude of model parameters, thereby reducing complexity and improving stability. These strategies strengthen the model's resilience against external noise while preserving its predictive accuracy in practical applications.

In addition, the image dataset originated from the COCO consortium and was used as the basis for image segmentation through the simulation method. The same customized optimizer and overfitting mitigation strategies were employed to train the U‐Net architecture (Figure S19, Supporting Information), ensuring that the WIST device‐based model effectively mimics real‐world human visual perception systems.

This combined approach leveraged the simulation method to demonstrate the high potential of proposed WIST devices in advanced neuromorphic applications.

Conflict of Interest

The authors declare no conflict of interest.

Supporting information

Supporting Information

Chen J.‐L., Chiang T.‐C., Liu P.‐T., All‐Metal‐Oxide Heterojunction Optoelectronic Synapses with Multilevel Memory for Artificial Visual Perception Applications. Small 2025, 21, 2502271. 10.1002/smll.202502271

Data Availability Statement

The data that support the findings of this study are available in the supplementary material of this article.