Flexible optoelectronic synaptic transistors for neuromorphic visual systems

The human visual system senses external light signals and parallelly preprocesses signals through the retina, which is highly efficient and low-energy consuming.^1,2 To develop artificial visual systems by imitating the human visual system is of great significance to address the needs of driverless cars, drones, robot visual perception, and environmental security monitoring.^3–5 As a novel artificial visual system, a neuromorphic visual system based on non-volatile optoelectronic synaptic devices can mimic the human retina and integrate the functionalities of sensing, memory, and processing, overcoming the shortcomings of a conventional artificial visual system, such as data redundancy, data access delay, and high-energy consumption.^6–9 

Soft robot surface and biological bodies consist of many curved surfaces with arbitrary shapes. Conventional synaptic devices based on rigid substrates are hard to attach with curved surface and bear the physiological strain of biological organisms, which seriously hampers their applications in skin-attachable, wearable, and implantable fields for healthcare.^10,11 Optoelectronic synaptic devices with flexibility and biocompatibility are highly desired in neuromorphic visual systems for robot visual perception, visual prosthetics, and artificial intelligence.^12–14 An integrated array of the flexible optoelectronic synaptic devices may be surgically implanted into a human eye to help blind people regain their optesthesia.¹⁵ 

Various flexible optoelectronic synaptic devices have been developed based on two-terminal (2T) memristors¹⁶ and three-terminal (3T) transistors.¹⁷ The 3T structure is relatively more complicated than the 2T structure, but they can implement various synaptic functionalities by exploiting functional expandability, such as multi-gated structures.¹⁰ In optoelectronic synaptic transistors, light is additionally introduced to modulate the channel conductance and the synaptic weight besides the gate,^18–23 providing the devices with wide bandwidth, fewer crosstalk, lower energy consumption, and faster signal transmission.^24,25 When the basic synaptic behaviors of biological synapses are emulated, electrical and optical pulses are both treated as presynaptic action potentials. The channel current is treated as a postsynaptic response. The photoelectric synergistic modulation of the synaptic weight endows a neuromorphic visual system with versatile functionalities, such as cross-modal learning,²⁶ pattern memory under different mood states,²⁷ and hetero-synaptic plasticity.²⁸ 

In this review, we mainly discuss the research progress of 3T flexible optoelectronic synaptic transistors (FOSTs). First, the main working mechanisms and device structures of FOSTs are introduced. The active materials that are commonly used in FOSTs are discussed in detail, including organic semiconductors, metal oxides, perovskites, and two-dimensional (2D) materials. The applications of FOSTs in neuromorphic visual systems are presented. Perspectives on the development of FOSTs are finally outlined. Figure 1 shows the overview of this review.

Most 3T FOSTs have similar device structures as traditional transistors, including dielectric layer, active layer, and conductive 3T electrodes. However, each FOST has special working mechanism, device design, and material to perform conductance modulation and synaptic weight update.²⁹ To date, 3T FOSTs can be roughly divided into the following four types according to their working mechanisms (Fig. 2): (a) heterogeneous channel FOSTs, (b) electrolyte-gate FOSTs, (c) ferroelectric FOSTs, and (d) floating-gate FOSTs.³⁰ 

Generally, as shown in Fig. 2(a), a heterogeneous channel consists of a photoabsorber for light reception, high-mobility semiconductors for photogenerated carrier transportation, and a hetero-interface for carrier separation and storage. Since photogenerated holes and electrons are trapped in different materials, they can be stored for a long time without recombination, which results in slow decay of the photocurrent and synaptic behaviors. However, the channel conductance can only be reversibly reduced using gate bias. For instance, Zhu et al. prepared heterogeneous channel FOSTs by using CsPbBr₃ quantum dots (QDs) to absorb the light and carbon nanotube (CNT) to transport the photogenerated holes [Fig. 3(a)].³¹ An effective separation of the photogenerated carriers occurred due to the built-in electric field formed at the CsPbBr₃ QDs/CNT interface [Fig. 3(b)]. The channel conductance can be modulated by the light intensity and optical pulse number. The spike-intensity-dependent plasticity (SIDP) and spike-number-dependent plasticity (SNDP) of the biological synapse were mimicked by the transistor [Fig. 3(c)]. A reset gate pulse voltage was applied to discharge the trapped photogenerated carriers [Fig. 3(d)]. Similarly, the heterogeneous channel of the CsPbBr₃ QDs/organic semiconductor [Fig. 3(e)] and carbon dots/pentacene [Fig. 3(f)] was also used to realize the FOSTs by Zhang et al.¹⁷ and Lv et al.,³² respectively.

Electrolyte-gate transistors employ a high capacitance electrolyte as the gate insulator [Fig. 2(b)]; the high capacitance increases drive current and lowers operating voltages and power consumption benefited from the electric double layer (EDL) effect.^33,34 To date, electrolytes used in FOSTs can be classified into inorganic solid electrolytes, ionic liquid or ionic gel, and polymer electrolytes.

Inorganic solid electrolytes, such as Al₂O₃,³⁵ nanogranular SiO₂,^36–38 and graphene oxide proton conductor,³⁹ show great application potential in FOSTs. For instance, Wang et al. described a FOST based on a polyethylene terephthalate (PET) substrate, Al₂O₃ dielectric layer, and amorphous In–Ga–Zn–O (a-IGZO) channel layer [Fig. 4(a)].³⁵ The interfacial hydrogen doping from hydrogen-rich Al₂O₃ into a-IGZO generated subgap states and improved the photosensing performance of the transistor. The excitatory postsynaptic current (EPSC), paired-pulse facilitation [PPF, Fig. 4(b)], and SNDP [Fig. 4(c)] were emulated by the transistor under the 300 nm light stimulation. Wan et al. prepared a flexible neuromorphic transistor with PET as the substrate, graphene oxide as the dielectric layer, and In–Zn–O as the channel layer [Fig. 4(d)].³⁹ The transistor successfully mimicked the spike-timing-dependent plasticity (STDP) [Figs. 4(e) and 4(f)].

Li et al. presented a FOST based on an inorganic vanadium dioxide (VO₂) film grown on a mica substrate [Fig. 5(a)].⁴⁰ The transistor exhibited non-volatile modulation of channel conductance under 375 nm ultraviolet (UV) stimulation, which can be reversibly modulated using ionic liquid [N,N-diethyl-N-(2-methoxyethyl)-N-methylammoniumbis-(trifluoromethylsul phonyl)-imide (DEME-TFSI)] gating. During the reset process, ionic liquid gating with negative gate voltage could insert O²⁻ into a VO₂ film through hydrolysis reaction to make the device return to its initial state. The SIDP, spike-duration-dependent plasticity (SDDP), SNDP, spike-rate-dependent plasticity (SRDP), and spike-wavelength-dependent plasticity (SWDP) [Figs. 5(b)–5(f)] were emulated by using the transistor under the UV light stimulation. In addition, the transistor can still exhibit stable EPSC after 10 000 bending cycles at a bending radius of 6 mm.

Various polymer electrolytes are expanding the potential applications of electrolyte-gate transistors in FOSTs,³³ including poly(4-vinylphenol) (PVP),^15,41–43 sodium alginate (SA),^44,45 polyethylene oxide/lithium perchlorate (PEO/LiClO₄),⁴⁶ cellulose,^17,47 and chitosan.⁴⁸ For instance, Sun et al. constructed a FOST with polyimide (PI) as the substrate, a SA film as the dielectric layer, and zinc oxide nanowires (ZnO NWs) as the channel layer [Fig. 6(a)].⁴⁴ The SNDP was mimicked under the stimulation of gate voltage pulses [Fig. 6(b)] due to the EDL effect. In addition, the SIDP was mimicked under the stimulation of 365 nm light pulses [Fig. 6(c)] due to the photoconductive effect of ZnO NWs. Almost no degradation in the transfer curve under light is observed under bending states [Fig. 6(d)].

It is easy for electrolyte-gate FOSTs to develop multiple gates and achieve the functionality of dendritic integrations by using multiple inputs of voltage pulses, which shows great potential in the mimicking of complex synaptic behaviors.³⁹ 

Due to the reversible polarization tuned by gate voltage, ferroelectric materials are ideal candidates for the dielectric layer to modulate the carrier concentration in synaptic transistors.^49,50 Due to the Coulomb interaction between the carriers in the channel and polarization in the ferroelectric insulator, the channel conductance can be modulated and the synaptic behaviors can be mimicked. The ferroelectric polymers, poly(vinylidene fluoride-trifluoroethylene) [P(VDF-TrFE)] and poly[(1-vinylpyrrolidone)-co-(2-ethyldimethyl ammonioethyl meth-acrylate ethyl sulfate)] [P(VP-EDMAEMAES)], have been used by Wang et al. to fabricate ferroelectric modulated FOST on a polyacrylonitrile (PAN) substrate [Fig. 7(a)].⁵¹ The SIDP [Fig. 7(b)] and SRDP [Fig. 7(c)] of the biological synapse were mimicked by the transistor under the stimulation of 550 and 850 nm light. The transistor’s long-term potentiation (LTP) can be reset by gate voltage to regain its resting state. The hole mobility of the transistor remained unchanged under the 100 µm bending radius against flat ones.

Ferroelectric FOSTs exhibit advantages such as multilevel conductance and high stability; however, they still face scaling-up problems due to the difficulty in obtaining large-area ferroelectric films with high quality.³⁰ 

The difference between the floating-gate FOSTs and the conventional transistors is that an additional charge trapping layer (floating gate layer) is embedded in the dielectric layer in floating-gate FOSTs [Fig. 2(d)].⁵² Different materials in the dielectric layer play the role of charge blocking, charge trapping, and charge tunneling.^53,54 When the light is absorbed by the channel layer, some photogenerated electrons/holes in this layer will move across the tunneling layer into the floating gate layer, and they will be stored in the floating gate layer, resulting in non-volatile photocurrent in the channel.²⁹ 

Sun’s group reported a floating-gate FOST with 0D black phosphorus (BP) QDs⁵⁵ inserted between the Al₂O₃ layers as the floating gate [Fig. 8(a)]. The thick and thin Al₂O₃ film behaved as the blocking layer and the tunneling layer, respectively, while BP QDs acted as the floating gate layer to tailor the channel conductance. Under the 473 nm light stimulation, electrons and holes were generated in the MoSSe channel, and the holes moved across the Al₂O₃ tunneling layer and were stored in the BP QDs floating gate layer [Fig. 8(b)]. The accumulated electrons in MoSSe would lead to an increase current in the channel. The transistor mimicked the synaptic behaviors, such as SIDP [Fig. 8(c)] and SDDP [Fig. 8(d)]. Comparing with pure electrical or optical modulation, photoelectric synergistic modulation can strengthen LTP for higher order correlations [Fig. 8(e)]. The electrically modulated LTP/long-term depression (LTD) of the transistor could still be emulated under the 5 mm bending radius [Fig. 8(f)].

ZrO₂,²⁸ AlO_x nanoparticles,⁵⁶ as well as some polar dielectric materials, such as polyacrylonitrile (PAN) and poly(lactic acid) (PLA),⁵⁷ were also utilized as the floating gates of the FOSTs. Similar to the other FOSTs, floating-gate FOSTs need optical signals and electrical gate bias as programming resources for excitatory and inhibitory conductance modulation in the neuromorphic visual systems.

To successfully achieve the flexibility of the optoelectronic synaptic devices, all the related materials need to be flexible, including dielectric materials, substrates, and active materials. Flexible dielectric materials have been introduced above, such as inorganic solid electrolytes, ionic liquid, and polymer electrolytes. In terms of commonly used flexible substrates, there are mainly included ionic conductive cellulose nanopaper (ICCN),¹⁷ mica,⁴⁰ chitosan,⁴⁸ and polymer films, including polyethylene naphthalate (PEN),³¹ polyethylene terephthalate (PET),³⁹ PI,⁴¹ and PAN.⁵⁷ In the past few years, various types of active materials have been used to fabricate FOSTs, including organic semiconductors, metal oxides, perovskites, and 2D materials, which will be discussed in detail below.

Organic semiconductors dominate the studies of flexible devices due to their low-cost manufacturing, flexibility,^58–62 and biological compatibility.⁶³ Both organic small molecular semiconductors and polymer semiconductors have been used as the active materials in FOSTs.

Dai et al. reported a flexible C8-BTBT-C8 synaptic transistor on a PAN substrate.⁵⁷ The trapping and detrapping of the photogenerated charges at the C8-BTBT-C8/PAN interface provided the transistor with synaptic behaviors, such as EPSC and PPF. An array of 10 × 10 devices emulated the memory and learning behaviors similar to the human brain. Jie’s group fabricated a FOST on PEN¹⁵ and PI⁴¹ substrates, respectively, by using a p-type small molecular semiconductor, 5,11-bis(triethylsilylethynyl) anthra-dithiophene (Dif-TES-ADT). The transistors were capable of directly transmitting, memorizing, recognizing, and learning light stimulus like a biological neural system. The image recognition functionality was examined under both a flat state and a bending radius of 11 mm.¹⁵ 

Wang et al. demonstrated a FOST on a PAN substrate with a p-type copolymer, poly(isoindigo-co-bithiophene) [P(IID-BT)], as the active material.⁵¹ Based on the incident light intensity, frequency, and wavelength, the transistor could transform light stimulus into volatile and non-volatile synaptic signals. The hole mobility of the transistor did not change when it was wrapped around a capillary tube with the radius of 100 µm. Other polymer semiconductors, such as poly[2,5-(2-octyldodecyl)-3,6-diketopyrrolopyrrole-alt-5,5-(2,5-di(thien-2-yl)thieno [3,2-b]thiophene)] (PDPP-DTT),⁴³ chlorophyll-a,⁴⁷ and indacenodithiophene–benzothiadiazole (IDTBT),⁶⁴ have also been utilized as the active materials in the FOSTs.

It is worth noting that organic materials suffer from dissolution in water or organic solvents^63,65 and easy photoaging under UV irradiation, which limit their further applications in some extreme environment, such as high-temperature and strong UV irradiated conditions.^66–68 

Metal oxide was one of the materials that were earliest used in the flexible synaptic transistors due to its high charge carrier mobility, large-area uniformity, mechanical flexibility, transparency, and photosensitivity.^35,69,70 In recent years, IGZO,³⁵ VO₂,⁴⁰ ZnO,⁴⁴ and SnO₂⁴⁶ have been applied as the active materials in the FOSTs. For instance, the IGZO transistor reported by Wang et al. exhibited typical light-stimulated synaptic behaviors, including EPSC, PPF, and memory functionality.³⁵ The bent configuration and cycling (1000 bending times) showed negligible effect on the mobility and light to dark current ratio, indicating excellent mechanical flexibility of this FOST. Wei et al. proposed a FOST with the thin film of SnO₂ nanoparticles as the active layer and PEN film as the substrate.⁴⁶ The FOST realized essential synaptic responses with low power dissipation under UV light pulses of different intensities and durations. The electricity-stimulated postsynaptic current retained 41.8% of the initial value after 10 000 bends, showing the excellent mechanical flexibility of the device.

Intrinsically high mobility, excellent optoelectronic response performance, and good flexibility endow halide perovskites with great promise for high-performance optoelectronic synaptic devices.^71–74 Zhu et al. presented a FOST array of 32 × 32 devices using PEN as the substrate and CsPbBr₃-QDs and CNT as active materials.³¹ The sensor array demonstrated neuromorphic reinforcement learning after training with a weak light pulse. The transfer curves of the transistor were almost identical under a flat state and a bending strain of 0.4%. Recently, another perovskite material, PEA₂SnI₄, which exhibits lead free, nontoxic, and low-temperature solution processability, has been reported as the active layer in the FOST.⁴² The device could convert light signals into postsynaptic current and simultaneously perform learning and memory like a biological visual system.

The emerging 2D materials are considered to be ideal candidates for FOSTs because they are compatible with flexible substrates due to their ultra-thin nature.⁷⁵ In addition, they show high optical response due to lower dimensions and longer photocarrier lifetime because of spatial confinement effects.^76,77 In recent years, FOSTs with MoS₂ and MoSSe as the active materials are reported successively. Zhang’s group proposed MoS₂ FOSTs on both polydimethylsiloxane (PDMS)⁷⁶ and PET⁷⁸ substrates. The LTP synaptic behaviors of the MoS₂ synaptic transistor on a PDMS substrate were barely unchanged under the stretching strain of 6.5%.⁷⁶ Both the optical potentiation and electrical depression characteristics of the MoS₂ synaptic transistor on a PET substrate could be emulated even at a bending state (bending strain: 0.8%) without any degradation.⁷⁸ Sun’s group also conducted a series of innovative work about 2D material-based FOSTs. The flexible MoS₂²⁸ (MoSSe⁵⁵) synaptic transistor could emulate the electrically irritated LTP and LTD of the biological synapse under a bending state with a bending radius of 10 mm (5 mm) without any degradation.

FOSTs have shown great potential for applications in neuromorphic visual systems, including associative learning,⁵⁵ MNIST handwritten digit recognition,⁷⁶ color recognition,^35,42,51 image recognition and memory,^15,43,64 motion detection,⁴⁰ and pain perception.⁴⁶ Next, we will discuss the four latters in detail since they are relatively new.

Large-scale optoelectronic synaptic devices that can detect multi-wavelength light are highly desirable for neuromorphic visual systems to distinguish different objects and perceive the world like the human visual system. Generally, researchers always exploited narrow-bandgap or composite materials to achieve color distinction from visible light to the near-infrared (NIR) light. For example, Wang et al. generated subgap states in the bandgap of a-IGZO and integrated 7 × 7 a-IGZO transistors in an array.³⁵ The array was able to detect a multicolor circular pattern of light with wavelengths of 300, 350, 400, 450, and 500 nm [Fig. 9(a)]. Huang et al. combined the organic semiconductors and perovskite to broad the absorption and photodetection of the synaptic transistor [Fig. 9(b)] from visible to NIR light.⁴² The four 3 × 5 FOST arrays were able to distinguish the colors of blue (450 nm), green (520 nm), red (650 nm), and NIR light (808 nm) and display four images of different color patterns.

When external objects emit or reflect visible light into the human eyes, they are imaged on the retina. Neuromorphic visual systems based on non-volatile optoelectronic synaptic devices integrate the functionalities of sensing, memory, and processing, which can mimic the functionalities of image recognition and memory of the human visual system. For instance, an array of 8 × 8 FOSTs that could be tightly attached on a human eyeball model [Fig. 10(a)] achieved the pattern “8” recognition and learning under the bent state after light stimulation [Fig. 10(b)].¹⁵ Another array of 4 × 4 FOSTs adhering to the concave hemisphere was also able to recognize and memorize the letter “N” via light spikes irradiation and synaptic weight modulation [Fig. 10(c)].⁶⁴ 

The human visual system is highly efficient in acquiring moving objects.⁷⁹ However, the targets that current neuromorphic visual systems detect are mostly limited in static patterns and images. With the advent of the artificial intelligence era, the detection and recognition of moving objects are becoming increasingly important.^3,80 The motion detection has been preliminarily tried in FOSTs recently. Ge et al. conceptually demonstrated the motion detection of a polar bear in snow by using the UV-responsive flexible devices and a system with an in-sensor computing architecture (Fig. 11).⁴⁰ In this real-time application, two device arrays were used to perceive images at two different moments. Subsequently, the memorized graphs were passed into a differential amplifier for frame difference calculation. The prototype demonstration of the FOST offered a promising solution for efficient motion detection in neuromorphic visual systems.

Pain may occur when human eyes are exposed to UV light, so detection of intensities of UV light and perception of the corresponding pain are quite important to protect the human eyes from being injured. The FOSTs presented by Wei et al. could act as nociceptors to realize the pain perception.⁴⁶ A simulated cornea was exposed to UV light with different intensities and durations to achieve the three nociceptive features. When the postsynaptic current increased as the light intensity increased and surpassed the current threshold (15 pA), the pain was assumed to occur [Figs. 12(a) and 12(b)]. This result is similar to the physiological pain perception, in which an increase in harmful stimulus increases the degree of damage. Furthermore, by modulating the intensity and the interval between two successive UV pulses, the pain facilitation [Figs. 12(c) and 12(d)] and pain de-facilitation [Figs. 12(e) and 12(f)] were mimicked by the transistor, respectively.

The working mechanism, substrate, dielectric material, active material, array, stimulated light, energy consumption, bending test conditions, working temperature, and applications in the neuromorphic visual systems of each FOST mentioned above are summarized in Table I.

We have reviewed state-of-the-art FOSTs and their working mechanisms and device structures, including heterogeneous channel FOSTs, electrolyte-gate FOSTs, ferroelectric FOSTs, and floating-gate FOSTs. Various active materials, such as organic semiconductors, metal oxides, perovskites, and 2D materials, have been used to develop the FOSTs. These FOSTs can well mimic the synaptic behaviors, such as EPSC/IPSC, PPF/PPD, SDDP, SNDP, SRDP, STDP, and LTP/LTD. Furthermore, the advanced applications of FOSTs in neuromorphic visual systems have been demonstrated, including color recognition, image recognition and memory, motion detection, and pain perception. Despite significant progress of the FOSTs, there are some challenging issues to implement FOST-based neuromorphic visual systems for robot visual perception, visual prosthetics, and artificial intelligence.⁸¹ 

1. Low energy consumption. Energy consumption is of great importance to neuromorphic visual systems because vast information and multiple complex tasks need to be processed in parallel to accomplish visual perception like humans. To date, few FOSTs report the energy consumption, which should be as low as that (1–100 fJ) of a biological synapse per synaptic event.⁶³ The reduction of the device size and the decrease of optical spike width would be an effective approach to lower the energy consumption.⁸² The energy consumption per synaptic event of a FOST was 0.07–34 fJ when the optical spike width was less than 100 ms.¹⁵ Moreover, self-powered devices with photovoltaic effect may be the goal of FOSTs in the background of decreasing the energy consumption.

2. Fully optical modulation. Although Hou et al. proposed a 2T flexible optical synaptic device that can emulate both the excitatory and inhibitory synaptic behaviors in optical means,⁸³ most flexible optoelectronic synaptic devices seldom exhibit negative photoresponse, and the electrical gate bias is essential for inhibitory conductance modulation, restricting the bandwidth, processing speed, and integration density of the devices.^24,84 The optoelectronic devices that can achieve bipolar synaptic behaviors in optical means^85–88 and the flexibility should be further developed.

3. Excellent anti-fatigue performance. The anti-fatigue performances of current FOSTs were preliminarily evaluated by measuring their source–drain current⁴⁰ or electric-stimulated EPSC⁴⁶ after 10⁴ times of bending. The stability of light-stimulated synaptic behaviors after sufficient fatigue measurements was hardly evaluated, while the fatigue life of >10⁵ cycles is needed to meet the immediate applications in skin-attachable, wearable, and implantable devices.⁸⁹ Corresponding anti-fatigue strategies to achieve the excellent anti-fatigue performance of the FOSTs should be further designed.

4. 3D image recognition and motion detection. The human visual system is highly efficient in acquiring 3D moving objects. However, the targets that current neuromorphic visual systems detect are mostly limited in 2D plane and static images, while the extraction and detection of motion targets are hardly explored. There is an urgent need to accelerate the 3D image recognition and motion detection of the FOST-based neuromorphic visual system.³ The motion detection was preliminarily realized through an in-sensor computing architecture by Ge et al., offering a promising solution for efficient pattern recognition and motion detection systems.⁴⁰ 

5. Multimodal perception. The FOSTs with additional external sensory elements (such as mechanical displacement,⁹⁰ touch,⁹¹ or gas) should be developed to realize human-like integrated sensory platforms and intelligent soft robotics. A universal system that could process a variety of sensing signals and realize multimodal perception is of great importance for the future development of neuromorphic systems with low energy consumption.³⁰ 

This work was mainly supported by the National Key Research and Development Program of China (Grant No. 2018YFB2200101) and the Natural Science Foundation of China (Grant Nos. U20A20209, U22A2075, and 62004078). Partial support was provided by the Natural Science Foundation of China for Innovative Research Groups (Grant No. 61721005), the Fundamental Research Funds for the Central Universities (Grant No. 226-2022-00200), as well as the leading Innovative and Entrepreneur Team Introduction Program of Hangzhou (Grant No. TD2022012).

The authors have no conflicts to disclose.

Xiao Liu: Writing – original draft (lead); Writing – review & editing (equal). Dongke Li: Writing – review & editing (lead). Yue Wang: Writing – review & editing (equal). Deren Yang: Funding acquisition (lead); Project administration (lead); Supervision (lead). Xiaodong Pi: Funding acquisition (lead); Project administration (lead); Supervision (lead); Writing – review & editing (equal).

Data sharing is not applicable to this article as no new data were created or analyzed in this study.