Defect-induced subgap state engineering in neuromorphic metal-oxide phototransistors for in-sensor color processing

Abstract

Recently, with the rapid development of autonomous vehicles, intelligent robots, and mobile electronics, retina-inspired neuromorphic photosensors have attracted growing interest as color image processors for machine vision systems. These devices typically mimic essential functions of the human retina, such as multicolor detection and the processing of raw visual information. However, most neuromorphic photosensors have been implemented with heterojunction channel structures or complex circuit architectures, resulting in system complexity, low resolution, and inefficient energy consumption. Here, we propose a neuromorphic phototransistor with a homogeneous channel structure using subgap-engineered metal-oxide (MO) semiconductors. Despite the absence of conventional heterojunction channel architectures, subgap-engineered MO phototransistors, achieved through intentional doping with alkali metal ions, demonstrate broad spectral responsivity and analog conductance modulation in a compact device structure. Particularly, when doped with a small amount of alkali metal ions (Li 5 at%-MO), the device exhibits in-sensor color image processing capabilities, including full-color detection, enhanced analog conductivity, and distinct sensing performance based on the input color. By applying a 7 × 7 neuromorphic phototransistor array using the Li 5 at%-MO semiconductor as the frontend device, innovative refinement tasks of raw color images such as color character sharpness, noise reduction, and contrast enhancement were successfully achieved, significantly contributing to the overall performance enhancement of the machine vision system.

---

New concepts

We present the first demonstration of a homostructure neuromorphic phototransistor based on subgap-modulated oxide semiconductors, enabling full-color in-sensor image processing without the need for complex heterojunction architectures. While conventional neuromorphic photosensors rely on multilayer heterostructures for color discrimination, our approach utilizes interstitial lithium doping to precisely engineer shallow subgap states within a singular oxide semiconductor channel, thereby achieving color-selective analog memory functionality. This dual-state modulation mechanism—combining Li-induced shallow traps and oxygen vacancy-induced deep traps—enables non-volatile analog conductance tuning across the entire visible spectrum. Our work replaces structural complexity with electronic subgap control, simultaneously enhancing spectral sensitivity and simplifying device fabrication, and thus introduces a paradigm shift in the design of neuromorphic optoelectronic devices. Furthermore, by integrating these devices into arrays, we demonstrate real-time, energy-efficient, and scalable in-sensor image refinement, offering a compact and low-power alternative for machine vision frontends. Ultimately, by leveraging this frontend device capable of in-sensor color image processing, we achieved a significant improvement in both recognition performance and processing speed in machine vision systems.

Introduction

A machine vision system, inspired by biological vision systems, consists of a retina-inspired frontend (responsible for image detection and preprocessing) and a brain-inspired backend (responsible for image learning and recognition).¹ To date, the retina-inspired frontend part for machine vision systems has been realized using complex circuits and many components, including conventional photosensors, digital memories, and preprocessing chips, leading to issues such as redundant data and high power consumption.^2,3 To address these issues, as a state-of-the-art retina-inspired frontend device, recent developments have focused on neuromorphic photosensors with in-sensor processing capabilities, which enable image detection and complex image preprocessing within a single device.^4–6 These neuromorphic photosensors, equipped with both photodetection capabilities and analog memory functions, can deliver well-refined image data with improved contrast through in-sensor processing of raw images that contain high noise and blurred main features. After in-sensor processing, these well-refined data are transmitted to the backend part of the machine vision system. Meanwhile, the full-color in-sensor processing capability and color-discriminative contrast enhancement performance are crucial for achieving precise color information processing, ultimately contributing to enhancing image learning and recognition performance in machine vision systems.^7–9

In general, implementing retina-inspired neuromorphic phototransistors has often relied on the heterostructure channel architecture in phototransistors, which typically include a photoabsorber for color light reception, a semiconducting counter layer for photocharge transport, and a hetero-interface region for photocharge storage.^10–13 For the visible light absorber and semiconducting counter layer, high-efficiency solar light absorbers with nanostructures and crystalline semiconductors with high electrical mobility have been adopted, resulting in excellent color response and low-power operation. Although neuromorphic color vision sensors with a heterojunction channel structure have been successfully demonstrated, they have suffered from several challenges associated with multilayer stacking and nanoscale material usage, which include complex fabrication processes, device instability, and non-uniformity, hindering the practical realization of neuromorphic image sensor chips. In contrast, metal-oxide (MO) semiconductors such as InGaZnO, InZnO, and ZnSnO enable neuromorphic photosensing in phototransistors without the need for a heterostructure channel, allowing the fabrication of neuromorphic homostructure phototransistors with a singular channel layer.^14–16 These capabilities are possibly due to the persistent photoconductivity (PPC) effect, induced by the photoionization of point defects in MO semiconductors.^15,17 However, the localized energy states associated with naturally formed oxygen vacancies in MO semiconductors are mainly located in the deep subgap energy levels within their wide bandgap, limiting the photoionization to the high-energy visible light spectrum.^15,16 Consequently, MO-based homostructure phototransistors typically exhibit neuromorphic photosensing functions only within a limited spectrum range such as deep blue or ultraviolet (UV) light and often suffer from low sensitivity in most of the visible spectrum range. Due to the absence of shallow subgap states, MO-based phototransistors struggle to achieve key characteristics essential for neuromorphic image processing, such as high sensitivity, analog operation, and color-discriminative conductance changes for refined color image processing. Fortunately, interstitial elements such as alkali metals or protons are well-known to introduce shallow subgap states in MO semiconductors, enhancing their photoresponse in the red and green regions of the visible spectrum.^18–22 Therefore, the strategy of doping interstitial elements into conventional MO semiconductors with oxygen defects has great potential to realize full-color neuromorphic phototransistors for in-sensor image processing without heterojunction channel architectures. While extended photoresponse properties in the visible spectrum have been reported in classical photosensor applications, their neuromorphic spectral characteristics and applications for in-sensor image processing have rarely been demonstrated.¹⁸

Here, we report the development of MO semiconductor-based neuromorphic phototransistors with full-color sensing and in-sensor image processing capabilities, replicating essential retinal functions. These devices work as frontend processors in machine vision systems, enabling efficient preprocessing of raw color images. To address the inherent spectral limitations of conventional MO semiconductor channels, which arise from the absence of shallow subgap states, we introduce a subgap-modulated homostructure MO semiconductor via alkali metal doping. The incorporation of interstitial alkali metal dopants induces shallow subgap states, thereby extending the photoresponse to the full visible spectrum. Notably, lithium (Li)-doped ZnSnO (ZTO) channels exhibit both shallow subgap states, generated by the interstitial Li atoms, and deep subgap states, associated with oxygen vacancies (V_O), enabling enhanced spectral selectivity and neuromorphic functionality. The Li-doped ZnSnO (Li-ZTO) exhibits photo-induced non-volatile conductance switching and analog conductance updating, attributed to the photo-ionization of point defects. Leveraging these properties, we demonstrate a retina-inspired neuromorphic phototransistor including full-color detection with analog memory-like conductance updating functions for simultaneous image sensing and processing. To demonstrate the viability of the developed device for neuromorphic color image processing, a 7 × 7 phototransistor array based on subgap-modulated Li-ZTO was fabricated, enabling direct acquisition and in situ refinement of raw color images. Applying these refined images to pattern recognition tasks markedly enhanced recognition accuracy and computational efficiency in machine vision systems, highlighting the potential of the simplified MO-based neuromorphic phototransistors for advanced in-sensor processing.

Results and discussion

Fig. 1 shows a schematic representation of both the human visual system and the artificial neuromorphic image recognition system for high-efficiency color image recognition. Comprising the retina, optic nerve, and visual cortex, the human visual system performs complex and highly sophisticated color perception processes. Upon exposure to external visible light, the retina, as a biological sensor and preprocessing element, absorbs color signals and carries out initial visual information processing, thereby collecting contrast-enhanced image data. Color discrimination is mediated by variable cone photoreceptors, each selectively sensitive to specific wavelength ranges, demonstrating full-color perception through their combined activation.^3,4 Following biological in-sensor processing in the retina, the refined image is transmitted to the visual cortex of the brain. Color image pattern recognition is achieved through a computational process that utilizes the biological neural network structure of the visual cortex. In parallel, the artificial neuromorphic image sensor emulates the biological in-sensor processing of the retina, simultaneously capturing raw color information and performing preprocessing within a single device. Specifically, phototransistors function analogously to retinal photoreceptors. This well-refined image information is subsequently processed via artificial neural networks along with machine-learning algorithms designed to mimic neuromorphic functionalities. By leveraging the biomimetic approaches, the artificial neuromorphic system facilitates efficient color image recognition, offering promising advancements for next-generation machine vision applications. To realize a retina-inspired neuromorphic full-color image sensor, the phototransistors should exhibit broadband visible light detection characteristics and neuromorphic processing capabilities, including non-volatile conductance switching and analog updating in response to color signals.

Fig. 1 Schematics of biological and artificial vision systems for high-efficiency color image recognition. (a) Human visual system and (b) neuromorphic image recognition system.

However, as shown in Fig. 2(a), typical MO-based phototransistors using InGaZnO, InZnO, and ZnSnO indicate limited color sensing, predominantly in the blue spectrum, despite their non-volatile storage characteristics. The narrow spectral response may be attributed to their subgap density-of-states (DOS), primarily determined by oxygen vacancy (V_O)-related states. These deep subgap states, centered around 2.3 eV below the conduction band minimum (CBM) with a Gaussian distribution, can limitedly contribute to photo-electron generation under high-energy visible light illumination via the photo-ionization process of V_O, following the reaction V_O → V⁺⁺_O + 2e⁻ (V_O = oxygen vacancy; V⁺⁺_O = doubly ionized oxygen vacancy; e⁻ = photo-electron), as illustrated in Fig. 2(b).^23,24 Therefore, conventional optoelectronics utilizing homostructure MO channels with deep-state dominant subgap DOS show partial detection capabilities, mostly for high-energy blue light.^15,17 In contrast, as shown in Fig. 2(c), the interstitial-metal-doped MO semiconductors achieve broadband light responsivity due to a modified subgap DOS structure that consists of both shallow subgap states from interstitial metal doping and V_o-related deep subgap states, enabling full-color sensing and photo-induced conductance switching. Specifically, interstitial elements with small atomic radii, such as alkali metals and protons, can introduce additional shallow subgap states near the CBM.^18–22 These states facilitate photo-electron generation even under low-energy visible light by compulsory photo-ionization in response to light stimuli, following the reaction M_i → M_iⁿ⁺ + ne⁻ (M_i = interstitial metal; M_iⁿ⁺ = metal ion in the interstitial site, e⁻ = photo-electron), as illustrated in Fig. 2(d).^18,21 Next, after the termination of light illumination, the photo-charges generated during light exposure may undergo a recombination process. In general, the band-to-band recombination process between delocalized photo-charges (i.e., photo-electron and photo-hole) shows spontaneous characteristics (Fig. S1a), characterized by a high recombination probability and low activation energy.^25,26 As a result, the created photo-charges are rapidly annihilated upon the termination of light illumination, preventing memory-related behaviors. In contrast, the recombination between delocalized photo-electrons and localized ionized point defects in our device (i.e., V⁺⁺_O + 2e⁻ → V_O and M_iⁿ⁺ + ne⁻ → M_i) follows a non-spontaneous process, requiring relatively high activation energy and exhibiting a prolonged reaction time with low recombination probability (Fig. S1b).²⁷ Consequently, the photo-induced (updated) conductivity states generated by the compulsory photoionization of point defects persist even after the termination of light exposure, demonstrating the memory-like PPC effects. These conductance switching characteristics, driven by the interstitial metal- and oxygen vacancy-related PPC phenomenon, enable non-volatile memory functionalities over a broad spectrum range, offering new opportunities for neuromorphic vision sensing applications.

Fig. 2 Subgap-engineered MO semiconductor-based neuromorphic phototransistors for full-color image processing. (a) and (b) Conventional MO-based neuromorphic phototransistor: (a) narrow spectrum (limited to blue) sensing dynamics and (b) driving mechanism. (c) and (d) Bandgap-engineered MO (interstitial Li doped MO)-based neuromorphic phototransistor: (c) full-color sensing dynamics and (d) driving mechanism.

Fig. 3(a) illustrates the device structure of a phototransistor employing a Li-doped ZTO (Li-ZTO) channel, fabricated to demonstrate a neuromorphic full-color image processor with a homostructure channel layer. Among various interstitial cation candidates, including alkali metals and alkaline earth metals, Li, owing to its smallest radius, is readily incorporated into ZTO films without significantly altering the microstructure. Furthermore, Li incorporation offers superior compositional control compared to the proton, facilitating precise modulation of the electronic properties of the films. Similar to undoped ZTO, the 10-nm-thick Li (5 at%)-ZTO film exhibits an amorphous phase, as evidenced by the halo ring-shaped selective-area diffraction pattern image, the smooth film surface, and the well-defined channel/gate-dielectric interface (Fig. 3(b) and Fig. S2). Meanwhile, due to theoretical limitations in energy dispersive X-ray spectroscopy using a transmission electron microscope (TEM), direct composition mapping of the Li element is not feasible. Instead, the presence of the Li dopant within the amorphous ZTO thin film was verified through the Li-related peaks in the X-ray photoelectron spectroscopy (XPS) spectrum (Fig. 3(c)). Additionally, photo-excited charge-collection spectroscopy (PECCS) analysis was performed to compare the subgap DOS profiles between undoped ZTO and Li (5 at%)-ZTO films. For the PECCS analysis, monochromatic photons with varying photonic energies, but with a constant light power intensity of 0.1 mW cm⁻², were irradiated onto samples (Fig. S3).^28,29 As shown in Fig. 3(d), undoped ZTO exhibited a negligible density of deep subgap states near the CBM, while a significant density of subgap states were observed in the energy range of 2.0 eV to 2.6 eV below the CBM. These deep subgap states in amorphous ZTO are attributed to V_O defects. In contrast, the Li (5 at%)-ZTO film showed both deep subgap states and a clear emergence of shallow subgap states near the CBM, spanning from 1.2 eV to 1.8 eV below the CBM, demonstrating a successful modification of the electronic structure upon Li incorporation.

Fig. 3 Device configuration and fundamental optoelectronic characteristics of MO phototransistors using a homostructure channel. (a) Device schematics, (b) conventional bright-field TEM image, (c) XPS Li 1s peaks, (d) subgap DOS profile in bandgap (directly extracted using the PECCS method), (e) transfer curves, and (f) the summary of figure-of-merits. (g)–(i) Band structures of various ZTO films: (g) undoped ZTO, (h) Li (5 at%)-ZTO, and (i) [Li (10 at%)-ZTO]. Color sensing performance of phototransistors using various ZTO films [undoped ZTO, Li (5 at%)-ZTO, and Li (10 at%)-ZTO]. (j) The variation of transfer curves under illumination with three colors (R, G and B) and (k) their photocurrent values. The light emitting diodes (LEDs) with peak wavelengths of 454 nm, 519 nm, and 635 nm were used as B, G, and R light sources, respectively, with a light power density (P) of 7 mW cm⁻².

Subsequently, the optoelectronic properties of various MO phototransistors employing the homostructure channel layer were investigated to elucidate the impact of Li incorporation on the subgap states of ZTO films. Due to its small atomic radius, Li element can act as either an interstitial or a substitutional dopant in MO semiconductors, depending on the doping concentration.^30,31 At low concentrations, as in the case of Li (5 at%)-ZTO, the Li atoms preferentially occupy interstitial sites rather than substitute metal cations (Zn²⁺ and Sn³⁺). These interstitial Li dopants (Li_i) introduce shallow subgap states near the CBM and generate free electrons via the interstitial doping process, following the reaction Li_i → Li_i⁺ + e⁻ (interstitial doping). In fact, as shown in Fig. 3(e) and (f) along with Fig. S4 and S5, the Li (5 at%)-ZTO transistor exhibited negatively shifted threshold voltage (V_Th) and increased subthreshold swing (S.S.) values compared to the undoped ZTO transistor, indicating the increase of electron concentration and the generation of shallow subgap states, respectively. Energy band structure analysis using ultraviolet photoelectron spectroscopy and UV-vis spectroscopy further support these observations (Fig. S6). The Fermi level (E_F) of the Li (5 at%)-ZTO film was measured at 4.10 eV, which is closer to the CBM than that of undoped ZTO (E_F = 4.3. eV). As a result, the Li (5 at%)-ZTO films are expected to show broadband photo-sensing capability, attributed to the presence of a widely distributed subgap DOS structure, incorporating both interstitial Li-induced shallow states and Vo-induced deep states. As shown in Fig. 3(g) and (j), undoped ZTO exhibited a photo-response predominantly limited to the blue (B) light spectrum. This behavior is attributed to the deep state-dominant subgap DOS structure, resulting in photocurrent generation and a negative shift of V_Th (ΔV_Th = 6 V). Meanwhile, under red (R) and green (G) light illumination, the device indicates negligible photoresponse due to the absence of shallow subgap states. In comparison, the Li (5 at%)-ZTO phototransistor demonstrates broadband RGB detection, facilitated by the widely distributed subgap DOS structure (Fig. 3(h)), enabling high-performance and color-discriminative photo-sensing characteristics such as distinct V_Th shifts (ΔV_Th = 5.1 V, 7.5 V, and 14.3 V under R, G, and B illumination, respectively) and varying photocurrent responses (I_Photo = 5.68 × 10⁻¹¹ A, 3.1 × 10⁻¹⁰ A, and 6.66 × 10⁻⁸ A under R, G, and B illumination, respectively) (Fig. 3(j) and (k)).

In contrast, excessive incorporation of Li elements [doping concentration > 5 at%] leads to a reduction of color-responsive subgap states. When the Li content exceeds the limit for interstitial doping, additional Li atoms act as substitutional dopants, forming hole carriers and lithium–oxygen (Li–O) bonds via the following reaction: Li → Li_Zn–O + h⁺ or Li_Sn–O + 2h⁺ (substitutional doping).³¹ As shown in Fig. 3(i), the E_F of the Li(10%)-ZTO films shift further from the CBM (E_F = 4.5 eV), indicating a significant reduction in electron carrier concentration. Furthermore, the formation of Li–O bonds suppresses the density of V_O-induced deep states owing to the high oxidation potential of Li (Li: 3.04 V) compared to other metal cations such as Zn (0.76 V) and Sn (0.14 V).³² XPS analysis (Fig. S7) further confirms that Li (10 at%)-ZTO exhibits an increased density of metal–oxygen (M–O) bonds and a reduced V_O concentration, consistent with the substitutional doping behaviors of excessive Li. Consequently, compared to Li (5 at%)-ZTO, Li (10 at%)-ZTO demonstrates a more positive V_Th shift and suppression of S.S. (Fig. 3(f)), validating a reduction in color-responsive subgap states and electron concentration due to substitutional Li doping. These results suggest that, unlike Li (5 at%)-ZTO, the Li (10 at%)-ZTO film is less favorable for achieving broadband or full-color neuromorphic color sensors.

The machine vision system consists of a retina-like frontend processor for image acquisition and preprocessing, and a brain-like backend processor for image recognition. Phototransistors, which enable both color image detection and analog conductance updating, can be used as retina-inspired neuromorphic image processors, especially for key frontend components in machine vision architectures. To assess the potential for neuromorphic color image sensing, the photo-induced conductance state variations of the ZTO-based phototransistors were systematically examined by monitoring drain current variations under dynamic R/G/B-light pulse conditions. The conductance of phototransistors was measured in the subthreshold region (V_GS − V_Th = −1.0 V), where the initial current in the dark state remains low, while photosensitivity (i.e., the ratio of photocurrent to initial current) is maximized compared to other operating regions (Fig. S8).³³ The conductance switching behaviors under a single pulse of R/G/B-light was initially evaluated. As shown in Fig. 4(a), the undoped ZTO phototransistor shows a rapid increase in conductance under B-light illumination due to the active generation of photocarriers from the deep subgap states associated with V_O. Moreover, the elevated conductance can be maintained for an extended period following the termination of B-light illumination, indicating that the conductance state can be switched to a higher conducting state under the B-light pulse. The conductance retention behaviors can be attributed to the incomplete recombination of photo-electrons and photo-ionized deep subgap states, which is a kind of non-spontaneous process requiring high activation energy. Unlike the spontaneous and rapid recombination process between delocalized photo-electrons and delocalized photo-holes, which can freely move within the material, the relaxation between delocalized photo-electrons and localized photo-ionized deep subgap states is not easily activated and occurs at a significantly slower rate.^15,26 In contrast, under R- and G-light pulse illumination, the device fails to exhibit clear conductance changes and non-volatile behavior owing to the absence of localized shallow subgap states induced by point defects. Instead, slight conductance transitions are observed under R and G light pulse illumination, which are believed to be caused by photo-electron generation from band-tail states near the conduction band, formed due to disordered MO bonding networks (i.e., disordered bonding angles and lengths) within the amorphous metal oxide. Such conductance changes induced by photoexcitation from band-tail states can exhibit volatile behavior due to rapid recombination phenomena between delocalized photo-electrons and the widely distributed band-tail states within the MO semiconductor (Fig. S9). Comparatively, the Li-ZTO phototransistor exhibited non-volatile switching behaviors for all R/G/B-light pulses (Fig. 4(b) and (c)), which is attributed to the presence of widely distributed subgap states in the bandgap. These non-volatile switching characteristics under R/G/B light illumination were also similarly observed in devices doped with other alkali metals, such as Na and K (Fig. S10). Meanwhile, the Li (5 at%)-ZTO phototransistor exhibits distinct conductance switching characteristics, featuring a high conductance and robust non-volatile retention performance compared to the Li (10 at%)-ZTO phototransistor. Particularly, the conductance increased in response to the photon energy of the incident light, yielding conductance values of 2.04 × 10⁻¹¹ S, 4.35 × 10⁻¹¹ S, and 1.29 × 10⁻⁹ S under R-, G-, and B-light pulses, respectively, at a light intensity of 7 mW cm⁻². These observations are likely due to the increased involvement of deeper subgap states in photo-electron generation with higher incident photon energy.

Fig. 4 Color detection and analog conductance switching characteristics of MO phototransistors using a homostructure channel. (a)–(c) The variation of conductance under a single pulse of R/G/B-light: (a) undoped ZTO, (b) Li (5 at%)-ZTO, and (c) Li (10 at%)-ZTO. (d)–(f) The variation of conductance under multiple pulses of R/G/B-light (pulse number = 30): (d) undoped ZTO, (e) Li (5 at%)-ZTO, and (f) Li (10 at%)-ZTO. (g) and (h) Dynamic photoresponse of ZTO-based phototransistors to 30 pulses of R/G/B-lights: (g) undoped ZTO and (h) Li (5 at%)-ZTO. (i) Generation mechanism of the analog conductance update function in Li-ZTO phototransistors.

Furthermore, the analog conductance updating capability of the phototransistors was evaluated under repeated R/G/B-light pulses (30 pulses). As shown in Fig. 4(d), the undoped ZTO transistor exhibits conductance updating behavior exclusively under B-light pulses. The conductance progressively increases corresponding to the number of B-light pulses (Fig. 4(g)), resulting in a cumulative conductance update. Moreover, even after the termination of the B-light pulses, the elevated conductance persists with slow decay characteristics, indicating long-term retention of the conductance state. On the other hand, under R/G-light pulses, the conductance (C) values show transient detection behavior with rapid decay characteristics. The increased conductance after 30 pulses (C_30th) of R- and G-light rapidly decreased to 37% and 34% within 50 s after termination of R- and G-light pulses, respectively. In contrast, the Li-ZTO phototransistors exhibit analog conductance update behaviors under all R/G/B-light pulses (Fig. 4(e) and (f)), which provides a basis for obtaining preprocessed color image data with enhanced contrast and reduced noise. Notably, regardless of the light color, the conductance reveals incremental updates with each successive light pulse. This behavior is possibly due to the excessive generation and incomplete recombination of photocarriers within the Li-ZTO channel, leading to a gradual accumulation of photocarriers (Fig. 4(i)). Specifically, in the case of the Li (5 at%)-ZTO transistor, the conductance state could be clearly enhanced depending on the number of repeated light pulses, resulting in a dramatic increase in conductance (Fig. 4(h)). Compared to the conductance value obtained after a single pulse (C_1st) of R/G/B-light, the conductance values after 30 pulses of R/G/B-light (C_30th) increase by factors of 15.5, 22.1, and 70.6, respectively. Additionally, a linear change in the conductivity state is observed during the light pulse, with low non-linearity (NL) values close to zero (Fig. S11). An NL value of zero indicates an ideal straight-line response.³⁴ Furthermore, the updated conductance state can be maintained, showing robust retention behavior with slow decay characteristics. The C/C_30th values remained at 77%, 63%, and 59% after 50 s following the termination of R/G/B-light pulses, respectively, demonstrating the stability of the conductance state.

Fig. 5(a) shows the neuromorphic image processor developed as a frontend device, which can further enhance machine vision systems by simultaneously performing color image detection and data preprocessing on the same device based on contrast enhancement and noise reduction capabilities. To verify the viability of the developed device as a neuromorphic color image processor, a simple prototype of the neuromorphic color image processor was implemented using a 7 × 7 device array of Li (5 at%)-ZTO phototransistors (Fig. 5(b)). An individual phototransistor can be considered as one pixel of the refined output image. The device array demonstrates exceptional device-to-device uniformity, as evaluated through the electrical characteristics of 49 devices uniformly distributed across a 9.9 cm² Si substrate (Fig. S12). To demonstrate simple image preprocessing tasks, various raw images of letter ‘A’ were prepared, maintaining a consistent main feature while introducing randomly distributed background noise. In all raw datasets, letter ‘A’ occupies the identical 17 pixels within the 7 × 7 pixel array, while the remaining 32 pixels contain stochastic noise. As illustrated in Fig. 5(c), image processing was performed using the neuromorphic color image processor by mapping the light intensity of each pixel in the raw image. Strong (P = 7 mW cm⁻²) and weak (P = 3 mW cm⁻²) color light pulses were employed as input signals for the main feature and noise pixels, respectively. During sequential input of the noisy character image, strong light pulses (7 mW cm⁻² and a pulse duration of 10 s) were constantly irradiated to the pixel locations corresponding to letter ‘A’, whereas weak pulses were randomly distributed across the background pixel. As a representative case, the detailed processing results for the red-colored letter ‘A’ are provided in Fig. S13.

Fig. 5 Neuromorphic color image processor. (a) The schematics of image preprocessing through a neuromorphic image processor, (b) photograph and optical microscopy images of the 7 × 7 array of ZTO-based phototransistors and an individual pixel. The demonstration of the image processing task. (c) The mapping results of output currents measured on each neuromorphic sensor after preprocessing of colored “A” letters with noise. (d)–(f) Update behavior and distribution of drain current values measured from phototransistors present at each pixel of the image processor during iterative image preprocessing of colored “A” letters with noise. Comparison of C/C₀ values at the 19 pixels corresponding to the colored ‘A’ letter and C/C₀ at 30 background pixels: (d) R-light, (e) G-light, and (f) B-light.

As shown in Fig. 5(d)–(f), increasing the number of preprocessing cycles leads to a gradual enhancement in the conduction state of the phototransistors corresponding to letter ‘A’, resulting in an increased C/C₀. Here, C₀ and C denote the initial conductance and the light-stimuli updated conductance, respectively. These preprocessing operations can be achieved for letter ‘A’ with RGB colors owing to the full color response characteristics of the Li (5 at%)-ZTO phototransistor. Meanwhile, phototransistors corresponding to the background region of the raw image exhibited minimal changes in conductance, as the weak light pulses representing noise were intermittently incident and transient. Consequently, repeated input of the raw image resulted in progressively higher output signals (C/C₀) in the phototransistors, leading to a more refined output image. This can be clearly observed in the output images shown in Fig. 5(c), where the main feature, letter ‘A’, becomes increasingly distinct compared to the initial output image, while the background noise is gradually suppressed. Additionally, despite an identical number of preprocessing operations, the neuromorphic image process can provide distinct output signals depending on the color of the input image. When the blue letter ‘A’ with high photon energy was utilized, the conductance state of the phototransistor indicated a more pronounced increase compared to the red letter ‘A’, owing to a greater generation of photocarriers. This resulted in a significant enhancement in the conductance compared to the initial state with C/C₀ = 15 145, 779, and 184 for letter ‘A’ positions after 20 light pulses of B-, G- and R-light, respectively. The results presented here suggest that a neuromorphic image sensor consisting of the Li (5 at%)-ZTO phototransistor array can be viable as the frontend processor of a machine vision system, capable of transformation of raw color images into refined image data with enhanced contrast. Meanwhile, to process new image data with different colors or shapes, it is essential to reset the updated conductance states of the phototransistors, restoring them to their initial state. As shown in Fig. S14 and S15, the conductance states, updated by light stimulation, can be fully reset by applying a gate bias pulse (V_GS for the reset operation = +2.5 V for 2 s).

As shown in Fig. 6(a), a machine vision system for machine recognition of color images was established using a neuromorphic phototransistor array (frontend part for image preprocessing) and AI algorithms (backend part for image recognition). For the raw image data, two types of 7 × 7 pixel character images (‘A’ and ‘B’) with six RGB color combinations for the main features and background areas were used (Fig. 6(b)). For each color configuration, multiple variations of letters ‘A’ and ‘B’ were extracted from a dataset of 100 images (Fig. S16). Therefore, a total of 1200 color character images [2 characters × 6 RGB color combinations × 100 shape variations] were prepared for training and testing. The raw images were repeatedly processed using the image preprocessor to enhance contrast through iterative image refinement. In this process, visible light signals corresponding to each pixel color in the raw image were applied to individual phototransistors within the image preprocessor at a constant light intensity of 7 mW cm⁻². Following repeated image iteration, the I_Photo/I₀ values of each phototransistor were extracted and converted into grayscale values from 0 to 255. By mapping these grayscale values, a refined image was reconstructed through image preprocessing. The image preprocessing was conducted for all 1200 raw images. Depending on the number of preprocessing cycles and the specific neuromorphic image processor used, refined images with varying grayscale quality were obtained. As a representative example, Fig. 6(c)–(g) present the grayscale images obtained by preprocessing the raw color images of letters ‘A’ and ‘B’ in Fig. 6(b). As demonstrated in Fig. 6(c)–(f), iterative preprocessing progressively enhances image contrast, producing grayscale character images with more distinct differences between the key features and the background. It is worth noting that the enhancement of image contrast through image preprocessing is a unique capability of neuromorphic image sensors, which is generally unattainable with conventional sensors. During the convolutional neural network (CNN) model training and pattern recognition using refined data, it is observed that multiple preprocessing iterations improved learning efficiency validated by a reduction in the number of training epochs and enhanced recognition rates compared to the single preprocessing case (Fig. 6(h)). However, increasing the number of preprocessing steps also results in higher power and time consumption. Notably, beyond ten times of image preprocessing, no significant improvement in learning efficiency or recognition rates is observed. These findings highlight the importance of optimizing the number of preprocessing cycles to balance performance gains with recognition efficiency.

Fig. 6 Demonstration of pattern recognition in color images using an artificial vision system. (a) A machine vision system consisting of a neuromorphic phototransistor array and AI algorithms on a commercial computer. (b) Representative examples of raw color character (‘A’ and ‘B’) images. (c)–(g) Representative examples of refined images after the image preprocessing task though a neuromorphic image processor. The cases using the Li (5 at%)-ZTO based image processor: preprocessing numbers of (c) 1, (d) 5, (e) 10, and (f) 20. The case using the ZTO based image processor: a preprocessing number of (g) 10. (h) and (i) Comparison of the pattern recognition performance of color images using an artificial vision system according to (h) the preprocessing number [the case using the Li (5 at%)-ZTO based image processor] and (i) the type of image preprocessor.

Additionally, the performance of the Li (5 at%)-ZTO-based neuromorphic processor was compared with that of the undoped ZTO-based neuromorphic processor. As shown in Fig. 6(e) and (g), with an identical preprocessing number of 10, the Li (5 at%)-ZTO-based neuromorphic processor produces high-quality refined images with distinct contrast, outperforming the ZTO-based processor. In the ZTO phototransistors, extracting clear character patterns from raw images with R/G color combinations is challenging owing to the low photoresponse to R- and G-light. As a result, during CNN-based learning and pattern recognition, the Li (5 at%)-ZTO-based image processor exhibits a high recognition accuracy of 92% and enhanced learning efficiency (saturation of recognition accuracy at 22 epochs) compared to the ZTO-based image processor, which achieves a recognition accuracy of 65% with a saturation point at 30 epochs (Fig. 6(i)). For a more comprehensive discussion, the color pattern recognition performance of the image processor was compared between the homostructure Li (5 at%)-ZTO and heterojunction channel structures, such as our previously developed heterostructure MO phototransistor with a mixed quantum dot light-absorbing layer.⁹ The Li (5 at%)-ZTO processor exhibits a comparable recognition accuracy of more than 90%, similar to that of the heterojunction channel structure (Fig. S17). These results show that, even in the absence of a heterostructure, the MO semiconductor-based optoelectronic synapse array with a subgap-engineered homostructure channel layer can possess excellent color image preprocessing capabilities for both monochromatic RGB and mixed-color images (Table S1), thereby providing critical insights into the development of compact and low-power neuromorphic machine vision systems (Table S2).

Experimental

Preparation of precursor solutions

Individual precursor solutions for Zn, Sn, and Li elements at 0.1 M were prepared by dissolving zinc nitrate hydrate (Zn(NO₃)₂·xH₂O; Sigma-Aldrich), tin chloride hydrate (SnCl₂·xH₂O; Sigma-Aldrich), and lithium nitrate (LiNO₃; Sigma-Aldrich) in 2-methoxyethanol (CH₃OCH₂CH₂OH; Sigma-Aldrich), respectively. The ZnSnO (ZTO) precursor was synthesized with the same composition of Zn and Sn (Zn : Sn = 5 : 5). Next, Li-ZTO precursor solutions for Li-ZTO films (Li = 5 at% and 10 at%) were prepared by adding the appropriate amount of Li precursor solution to the ZTO precursor solution according to their composition ratios [Li : (Zn + Sn) = 5 : 100 and 10 : 100]. Then homogeneous solutions were prepared via magnetic stirring at 400 rpm for 12 h at room temperature and filtered using a 0.2-μm syringe before sol–gel spin-coating.

Fabrication of AOS phototransistors and neuromorphic image processors

To fabricate MO phototransistors for neuromorphic color image processing, the precursor solutions for undoped ZTO and Li-ZTO films were spin-coated at 3000 rpm for 30 s onto p⁺⁺-Si wafers with a 200-nm-thick SiO₂ layer. The spin-coated precursor solution was chemically transformed into undoped ZTO and Li-ZTO films with 10 nm thickness via sequential soft-bake (250 °C for 10 min) and hard-bake processes (400 °C for 1 h). The channel area for MO phototransistors was defined via conventional photolithography and wet etching processes. The Mo source and drain electrodes with 100 nm thickness were deposited and defined via DC sputtering and conventional lift-off processes, respectively. To function as a neuromorphic image processor, AOS phototransistors were produced in a 7 × 7 device array configuration.

Characterization of the films and devices

The local microstructure and compositional characteristics of MO films (undoped ZTO and Li-ZTO) were characterized using conventional bright-field images and energy dispersive X-ray spectroscopy (EDX) compositional mapping using Cs-corrected TEM (JEM-ARM200F, JEOL). The transmission electron microscopy (TEM) samples were prepared using a focused ion beam (JIB-4601F, JEOL). The optical absorbance spectra and energy bandgap of MO films were measured using an ultraviolet–visible spectrometer (UV-Vis; Lambda 35). The chemical bonding states of MO films were evaluated via X-ray photoelectron spectroscopy (XPS; Thermo VG Scientific) using a monochromatic Al Kα (1486.6 eV) X-ray source. The work functions and valence band maximum energy (E_VBM) of the MO films were determined using an ultraviolet photoelectron spectroscopy (UPS) system (Nexsa, Thermo Fisher Scientific) with a He–I (21.2 eV) radiation source. The work function was calculated by subtracting the secondary electron cut-off energy from the incident photon energy (21.2 eV), as observed in the UPS spectrum. Additionally, the valence band edge (E_VBM) was derived from the onset of the valence band in the UPS spectrum. The photo-excited charge collection spectroscopy (PECCS) measurement setup to monitor the subgap states of MO films, utilizing a Hg (Xe) arc lamp (2000–200 nm), optical filters (NIR-VIS-UV), a grating monochromator, and a 200 μm diameter optical fiber, maintained an average optical power intensity of approximately 0.1 mW cm⁻² for all spectral photons.

Next, all optoelectronic characteristics of MO phototransistors were evaluated using a semiconductor parameter analyzer (Agilent 4156C, Keysight) at room temperature in ambient air. The drain current–gate voltage (I_DS–V_GS) transfer curve with V_GS sweep from −20 to +20 V was measured at a specific drain voltage (V_DS) of 10 V. The V_Th was determined as the specific V_GS at I_DS = L/W × 10 nA. The subthreshold swing (S.S.) value was extracted from the transfer curve (V_DS = 10 V) using the following equation: S.S. = dV_GS/dlog(I_DS). The density of charge trap states (N_it) can be calculated from the following equation: N_it = [(S.S. × log e)/(kT/q) − 1] × (C_i/q), where k and T are Boltzmann's constant and absolute temperature, respectively. The static and dynamic photoresponses of phototransistors under primary visible-light illumination were evaluated using light-emitting diodes (LEDs) with peak wavelengths at 454 (B), 519 (G), and 635 nm (R) being used as color light sources. The optical input-spike signals were generated by controlling visible LEDs using arbitrary function generators (Tektronix AFG 3252C).

Conclusions

In summary, a full-color neuromorphic image processor based on MO phototransistors with a homostructure channel layer was developed. By introducing an interstitial alkali metal dopant (e.g., lithium), we engineered the bandgap states of the MO semiconductor (e.g., ZTO) to achieve full-spectrum visible light responsiveness. This bandgap-engineered MO (Li-ZTO) enabled the implementation of a more compact and efficient device architecture, demonstrating RGB color detection and analog conductance updating capabilities. During the preprocessing of raw color images using Li-ZTO-based optoelectronic synapses, well-refined output images with enhanced contrast both between the background and key features, as well as among various colors, were directly obtained through on-chip processing, without requiring additional software. Finally, the refined image data were fed into a subsequent image recognition algorithm for color pattern machine recognition, achieving an improved recognition accuracy of up to 92% and enhanced learning efficiency compared to the conventional undoped ZTO phototransistor, which exhibited a limited spectral response and a recognition accuracy of 65%.

Author contributions

E. C. J.: writing – original draft, data curation, formal analysis, investigation, methodology, and visualization. D. H. B.: formal analysis, methodology, software, and visualization. J. M. L.: formal analysis, investigation, validation, and visualization. Y.-J. C.: methodology and validation. H. G. Shin: data curation and formal analysis. S. I.: data curation and formal analysis. J.-W. J.: validation. Y.-H. Kim: writing – review & editing and validation. S. K. P.: writing – review & editing, funding acquisition, project administration, resources, supervision, and validation. S. W. C.: writing – review & editing, conceptualization, data curation, investigation, and visualization.

Conflicts of interest

There are no conflicts to declare.

Data availability

The data supporting this article have been included as part of the supplementary information (SI). Supplementary information is available: See DOI: https://doi.org/10.1039/d5mh01018g.

Acknowledgements

This work was supported by the Nano & Material Technology Development Program through the National Research Foundation of Korea (NRF) funded by Ministry of Science and ICT(RS-2023-00281346). This work was supported by the Industry & Energy (MOTIE, Korea), by the Technology Innovation Program (or Industrial Strategic Technology Development Program-Development of Nano Convergence Innovative Product Technology) (20025062, Development of 100 um Pixel-Pitch High Resolution Flexible Xray Detectors based on Nanocomposites). This research was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea Government (MSIP) (No. RS-2024-00347845).

Notes and references

1. H. Kolb, Am. Sci., 2003, 91, 28–35.
2. C. Choi, M. K. Choi, S. Liu, M. S. Kim, O. K. Park, C. Im, J. Kim, X. Qin, G. J. Lee, K. W. Cho, M. Kim, E. Joh, J. Lee, D. Son, S. H. Kwon, N. L. Jeon, Y. M. Song, N. Lu and D. H. Kim, Nat. Commun., 2017, 8, 1664.
3. F. Zhou and Y. Chai, Nat Electron., 2020, 3, 664–671.
4. W. Chen, Z. Zhang and G. Liu, iScience, 2022, 25(2), 103719.
5. F. Zhou, Z. Zhou, J. Chen, T. H. Choy, J. Wang, N. Zhang, Z. Lin, S. Yu, J. Kang, H. S. P. Wong and Y. Chai, Nat. Nanotechnol., 2019, 14, 776–782.
6. F. Wang, F. Hu, M. Dai, S. Zhu, F. Sun, R. Duan, C. Wang, J. Han, W. Deng, W. Chen, M. Ye, S. Han, B. Qiang, Y. Jin, Y. Chua, N. Chi, S. Yu, D. Nam, S. H. Chae, Z. Liu and Q. J. Wang, Nat. Commun., 2023, 14, 1938.
7. R. Dutta, M. Naqi, Y. Cho, J. Oh, T. Kim, U. Jeong, Y. Yu, Y. Lee and S. Kim, Adv. Funct. Mater., 2024, 34, 202315058.
8. Y. Cheng, Z. Li, Y. Lin, Z. Wang, X. Shan, Y. Tao, X. Zhao, H. Xu and Y. Liu, Adv. Funct. Mater., 2024, 34, 2414404.
9. C. Jo, J. Kim, J. Y. Kwak, S. M. Kwon, J. B. Park, J. Kim, G. S. Park, M. G. Kim, Y. H. Kim and S. K. Park, Adv. Mater., 2022, 34, 2108979.
10. Z. Li, G. Zou, Y. Xiao, B. Feng, J. Huo, J. Peng, T. Sun and L. Liu, Nano Energy, 2024, 127, 109733.
11. C. Choi, J. Leem, M. S. Kim, A. Taqieddin, C. Cho, K. W. Cho, G. J. Lee, H. Seung, H. J. Bae, Y. M. Song, T. Hyeon, N. R. Aluru, S. W. Nam and D. H. Kim, Nat. Commun., 2020, 11, 5934.
12. Q. B. Zhu, B. Li, D. D. Yang, C. Liu, S. Feng, M. L. Chen, Y. Sun, Y. N. Tian, X. Su, X. M. Wang, S. Qiu, Q. W. Li, X. M. Li, H. B. Zeng, H. M. Cheng and D. M. Sun, Nat. Commun., 2021, 12, 1798.
13. S. S. Cho, J. Kim, S. Jeong, S. M. Kwon, C. Jo, J. Y. Kwak, D. H. Kim, S. W. Cho, Y. H. Kim and S. K. Park, Nano Energy, 2022, 104, 107991.
14. J. Shi, J. Zhang, L. Yang, M. Qu, D. C. Qi and K. H. L. Zhang, Adv. Mater., 2021, 33, 2006230.
15. M. Lee, W. Lee, S. Choi, J. W. Jo, J. Kim, S. K. Park and Y. H. Kim, Adv. Mater., 2017, 29, 1700951.
16. Q. Wu, J. Wang, J. Cao, C. Lu, G. Yang, X. Shi, X. Chuai, Y. Gong, Y. Su, Y. Zhao, N. Lu, D. Geng, H. Wang, L. Li and M. Liu, Adv. Electron. Mater., 2018, 4, 1800556.
17. S. Jeon, S. E. Ahn, I. Song, C. J. Kim, U. I. Chung, E. Lee, I. Yoo, A. Nathan, S. Lee, J. Robertson and K. Kim, Nat. Mater., 2012, 11, 301–305.
18. B. H. Kang, W. G. Kim, J. Chung, J. H. Lee and H. J. Kim, ACS Appl. Mater. Interfaces, 2018, 10, 7223–7230.
19. K. S. Kim, M. S. Kim, J. Chung, D. Kim, I. Sak Lee and H. J. Kim, ACS Appl. Mater. Interfaces, 2022, 14, 21150–21158.
20. S. Y. Park, B. J. Kim, K. Kim, M. S. Kang, K. H. Lim, T. Il Lee, J. M. Myoung, H. K. Baik, J. H. Cho and Y. S. Kim, Adv. Mater., 2012, 24, 834–838.
21. S. B. Orlinskii, J. Schmidt, P. G. Baranov, D. M. Hofmann, C. de Mello Donegá and A. Meijerink, Phys. Rev. Lett., 2004, 92, 047303.
22. S. Kim, J. Maassen, J. Lee, S. M. Kim, G. Han, J. Kwon, S. Hong, J. Park, N. Liu, Y. C. Park, I. Omkaram, J. S. Rhyee, Y. K. Hong and Y. Yoon, Adv. Mater., 2018, 30, 1705542.
23. K. Nomura, T. Kamiya, E. Ikenaga, H. Yanagi, K. Kobayashi and H. Hosono, J. Appl. Phys., 2011, 109, 073726.
24. T. Kamiya and H. Hosono, NPG Asia Mater., 2010, 2, 15–22.
25. J. Miao, W. Hu, N. Guo, Z. Lu, X. Zou, L. Liao, S. Shi, P. Chen, Z. Fan, J. C. Ho, T. X. Li, X. S. Chen and W. Lu, ACS Nano, 2014, 8, 3628–3635.
26. S. M. Sze and K. N. Kwok, Physics of Semiconductor Devices Physics of Semiconductor Devices, John Wiley & Sons, Hoboken, New Jersey, 3rd edn, 2007.
27. M. Kumar, R. K. Kher, B. C. Bhatt and C. M. Sunta, J. Phys. D: Appl. Phys., 2006, 39, 2670–2679.
28. K. Lee, M. S. Oh, S. J. Mun, K. H. Lee, T. W. Ha, J. H. Kim, S. H. K. Park, C. S. Hwang, B. H. Lee, M. M. Sung and S. Im, Adv. Mater., 2010, 22, 3260–3265.
29. J. H. Park, H. S. Lee, S. Park, S. W. Min, Y. Yi, C. G. Cho, J. Han, T. W. Kim and S. Im, Adv. Funct. Mater., 2014, 24, 1109–1116.
30. D. Wan, X. Liu, L. Xu, C. Liu, X. Xiao, S. Guo and L. Liao, IEEE Electron Device Lett., 2016, 37, 50–52.
31. J. Lee, S. Cha, J. Kim, H. Nam, S. Lee, W. Ko, K. L. Wang, J. Park and J. Hong, Adv. Mater., 2011, 23, 4183–4187.
32. P. Vany, Electrochemical series, CRC Handbook Chem. Phys, 2000.
33. B. H. Jeong, J. Lee, M. Ku, J. Lee, D. Kim, S. Ham, K. T. Lee, Y. B. Kim and H. J. Park, Nano-Micro Lett., 2025, 17, 78.
34. J. Fu, Z. Liao and J. Wang, IEEE Trans. Very Large Scale Integr. (VLSI) Syst., 2019, 27, 2745–2754.

---