The velocity of data generation has enormously increased in consequence of the rapid development of the Internet of Things, artificial intelligence, autonomous driving, and other related technologies^[1−6]. In the conventional computational architecture, processing unit and memory unit are typically separated. Therefore, the conventional computational architecture is faced with relatively high energy consumption and data latency when processing large amounts of redundant and unstructured data^[7]. Data-centric computing architecture offers an efficient computing paradigm that is particularly suitable for modern data-intensive applications, including big data analytics, artificial intelligence, machine learning, and cloud computing^[8]. In comparison to the conventional computational architecture, the data-centric computing architecture owns more integrated processing and storage elements, which results in the reduction in latency, bandwidth usage, and energy consumption^{[9, 10]}. The implementation of multiple cores at the hardware level to enhance the computational parallelism requires the allocation of considerable resources, including memory bandwidth. Consequently, the perceiving mechanism of the human brain attracts intensive attention to deal with the parallel computing fundamentally. The human brain is comprised of a highly intricate neural network, which is estimated to contain approximately 10¹¹ neurons and 10¹⁵ synapses^[11]. Neurons establish connections through synapses and synapses serve to facilitate communication and learning within neural networks. As information is transmitted between neurons in the form of spikes, the synaptic weights change simultaneously, thereby allowing for processing and storage in parallel^{[12, 13]}. Therefore, designing artificial synaptic devices is one type of potential approach to implementing the data-centric paradigm^[14]. Synaptic devices are expected to emulate the functions of biological synapses, such as signal transmission and plasticity^[15]. In the early stage, researchers have focused on the synaptic devices triggered by electrical inputs, including transistors^[16−18], memristors^[19−21], and atomic switches^[22]. However, the bandwidth-connection density and interconnect-related resistance-capacitance delay confine the capability of neural networks consisting of synaptic devices. In recent years, there has been a notable advancement in the field of optogenetics neuroscience^[11]. Following this advancement, optoelectronic synaptic devices have emerged and attracted considerable attention^[23]. In terms of optoelectronic synaptic devices, light is employed as the input to modulate synaptic characteristics. The utilization of light signals presents a number of advantages, including the elimination of resistance−capacitance delay, high bandwidth, and low crosstalk in comparison to electrical inputs^[13]. Furthermore, optoelectronic synaptic devices possess the capacity to emulate the functions of visual systems, comprising visual sensing, processing, and memory. Therefore, it is appealing to develop optoelectronic synaptic devices for the modern data-intensive applications^[6].

In recent developments, a diverse range of materials have been employed in the construction of optoelectronic synaptic devices^[24]. Among the various materials employed, organic materials exhibit desirable characteristics, including low cost, mechanical flexibility, the potential for large-area fabrication, and diverse processing methods^[25]. Moreover, the bandgap of organic materials can be modulated by modifying the molecular structure, leading to the desired optoelectronic properties. In addition, in comparison to two-terminal devices, three-terminal devices exhibit relatively reduced crosstalk, enhanced signal modulation capability, and greater fault tolerance characteristics. This endows optoelectronic synaptic transistors based on organic materials with enormous potential for the application in neuromorphic systems. Although a single organic optoelectronic synaptic transistor is capable of effectively emulating the fundamental behaviors of biological synapses, including potentiation, inhibition, and synaptic plasticity, the sophisticated functions that a single device can achieve are limited. In order to achieve image sensing or vector-matrix multiplication (VMM)^[26], neuromorphic visual systems and neuromorphic computing systems require device arrays that can be considered the essential pixels. As a result, the fabrication of organic optoelectronic synaptic transistor arrays (OOSTAs) is a necessary step for the practical utilization of the optoelectronic synaptic transistors in the real world.

In this article, the recent progress of OOTSAs is reviewed. The strategies for constructing OOSTAs are introduced in the second Section, with a focus on the manufacturing process of functional material films and the application potential enabled by the strategies. Several technologies, including coating and casting, physical vapor deposition, printing, and photolithography, are discussed. Subsequently, innovative applications of the OOSTA system integration are overviewed, including the neuromorphic visual systems and neuromorphic computing systems realized by the OOSTAs. In the final Section, the challenges and perspectives for the real-world utilization of OOSTAs in future are discussed, which mainly focuses on materials design, neuromorphic computing systems, and integration technology.

For constructing OOSTAs, a variety of strategies have been employed in the preparation of organic functional films (Fig. 1 and Table 1). It is noteworthy that, in recent studies, the research on OOSTAs has not emphasized the patterning of functional films. By employing suitable processing strategies, homogenous films based on functional materials can be manufactured, and subsequently patterned electrodes are deposited to form OOSTAs. As the demand for high-performance OOSTAs increases, the necessity for OOSTAs based on patterned functional materials becomes apparent^{[27, 28]}. By patterning functional materials, several benefits can be demonstrated, such as eliminating the crosstalk between adjacent devices, improving the consistency of photoelectric performance, and achieving circuit integration. This Section provides a comprehensive overview of the mainstream fabrication strategies for functional materials. These approaches enable the formation of uniform thin films, serving as a foundation for the subsequent preparation of OOSTAs.

For organic semiconductors, solution-processing techniques are highly viable and widely used for fabricating thin-film transistors, including organic optoelectronic synaptic transistors. In specific, casting and coating are particularly effective methods for the preparation of large-area films with controllable thickness and uniform surface. The film area and thickness mainly depend on the solution-processing methods employed, the volume of the solution, the viscosity of the solution, the volatility of the solvent, and other factors. Once large-area functional films with suitable thickness have been prepared, matching patterned metal electrodes can be prepared following the conventional techniques. Consequently, the fabrication of transistor arrays, including OOSTAs, is a relatively straightforward process. The techniques facilitate the processing of functional material without the necessity for sophisticated equipment, thereby reducing the cost of the fabrication. Furthermore, the functional material can be processed without the use of high temperatures, thus preventing the damage from harsh processing conditions.

Spin coating is currently a prevalent technique utilized for the preparation of uniform films of organic functional material. Prior to 1958, Emslie et al. conducted a preliminary analysis of this technique^[29]. In many scenarios, the material utilized in spin coating is polymeric^[30]. As the solvent evaporates, a film forms on the appropriate substrate. The most evident advantage is that the thickness of film can be readily adjusted by changing the spin speed or altering the concentration of the solution. Additionally, the resulting film exhibits relative uniformity. In regard to the carrier transportation in OOSTAs, a uniform film could lead to better consistency in their electrical characteristics. Hao et al. constructed OOSTAs comprising 6 × 6 pixels. The functional material, poly[2,5-(2-octyldodecyl)-3,6-diketopyrrolopyrrole-alt-5,5-(2,5-di(thien-2-yl) thieno [3,2-b] thiophene)] (DPPDTT) mixed with CsPbBr₃ quantum dots, was dissolved in an appropriate solvent and spin-coated onto the Si/SiO₂ substrate^[31]. Atomic force microscope analysis was employed to ascertain the surface quality, which was found to be smooth and uniform. In this OOSTA configuration, the source and drain electrodes were evaporated with shadow masks. Although the functional film of each OOSTA unit is not separate, the units demonstrated comparable photoresponsivity with decent consistency. Liu et al. also selected spin coating to construct humidity- and oxygen-insensitive functional film of the OOSTAs^[32]. In their work, the 10 × 10 OOSTAs based on DPPDTT/Cl-HABI functional film exhibited consistent transfer characteristic curves (Fig. 2(a)). It is noteworthy that spin coating enables the fabrication of a functional film based on organic semiconductors that are compatible with flexible substrates. Jiang et al. fabricated the flexible OOSTAs based on poly(3-hexylthiophene)-block-poly(phenyl isocyanide) (P3HT-b-PPI(5F))^[33]. The flexible OOSTAs could be transferred on various substrate and showed reliable EPSC across a range of bending radii (Fig. 2(b)). Furthermore, stretchable OOSTAs with the stretchable semiconductor layer have been constructed by Xu et al.^[34]. Initially, the indacenodithiophe-benzothiadiazole (IDTBT) solution was spin-coated on the octadecyltrichlorosilane-treatment rigid SiO₂ substrate. Subsequently, the IDTBT thin film with the polydimethylsiloxane (PDMS) dielectric was peel off by the PDMS embedding with the gate electrodes and laminated on the stretchable PDMS substrate with source and drain electrodes (Fig. 2(c)). The stretchable OOSTAs showed uniform photoelectric characteristics. By meticulous selection of the solvent and materials, a high-quality film can also be formed with small molecules. Wang et al. selected 6,13-bis(triisopropylsilylethynyl)-pentacene (TIPS-pentacene) by direct spin coating on a hydrophobic surface to prepare the functional film of the 9 × 8 organic transistor array (Fig. 2(d)), thereby demonstrating the potential of the spin coating strategy for the fabrication of OOSTAs based on small molecular semiconductors^[35]. The spin coating technique displayed the remarkable capability in the fabrication of large-scale OOSTAs, even at wafer scale. However, the process also exhibited certain limitations, which is wasteful of functional materials. In addition, uneven shear stress may also cause the performance variation of the OOSTAs from the center to the edge of the substrate.

Dip coating represents another technique of the preparation of functional material films or ordered organic semiconductor crystal arrays. In this process, the appropriate substrate is submerged in the solution of functional material. Subsequently, the substrate is pulled up with a specific velocity. The functional film or ordered crystal array is formed, depending on the type of material utilized. The quality of the resulting film is dependent upon a number of factors, including the viscosity of the solution, the pulling speed and the temperature of the process. In comparison to spin coating, dip coating allows for the creation of single crystal arrays with more precise patterns^{[36, 37]}. Some efforts have been made to construct functional layers of wafer-scale transistor arrays, including OOSTAs, using the dip coating technique. Significant research has been conducted on the utilization of the dip coating process for the preparation of organic transistor arrays based on the single crystal arrays. They commonly pre-patterning the microstructures based on photoresist, such as SU-8^[36−39]. Subsequently, the pull-up speed of the surface with surface modification in the dip coating process was controlled. Fig. 3(a) illustrates this process in detail^[38]. Under appropriate conditions, single crystal semiconductor nanowire arrays could grow on the specific substrate uniformly, as shown in Fig. 3(b). As illustrated in Figs. 3(c) and 3(d), similar work also was conducted, showing the reliability of the dip coating^[39]. High-performance organic transistor arrays could be further fabricated. Notably, little study was conducted on the optoelectronic performance of their constructed organic transistor arrays, including the simulation of synaptic behavior. Further research into the optoelectronic performance of OOSTAs fabricated through dip coating could yield valuable insights and enhance this field’s comprehensive capabilities. Xu et al. employed the dip coating technique to prepare CH₃NH₃PbI₃ nanoparticles/2,7-dioctyl[1]-benzothieno[3,2-b][1] benzothiophene (C8-BTBT) crystal array (Fig. 3(e)). Based on the crystal array as the functional layer, 10 × 10 phototransistor arrays were subsequently constructed (Fig. 3(f))^[40]. Nevertheless, there is a lack of research on the fabrication of OOSTAs through dip coating, which represents a promising avenue for further investigation.

Blade coating represents an attractive technique, offering the additional benefits of low material wastage and compatibility with roll-to-roll printing processes^[41]. The blade is an indispensable component of the blade coating process. In the initial stage of the process, the blade is placed over the substrate. Subsequently, the optimal solution is positioned within the gap between the substrate and the blade. The solution will then be distributed evenly within the gap. In this case, the blade moves and drags the solution, spreading it onto the substrate to form a film of the desired thickness and consistency. The thickness and quality of the film can be modulated by controlling the speed of blading, the viscosity of the solution, and the surface characteristics. In a recent study, Zhang et al. fabricated a complete and pinhole-free wafer-scale (2-inch) bis(triethylsilylethynyl) anthradithiophene (Dif-TES-ADT) crystalline film at a high blade coating speed (Fig. 4(a))^[42]. It is essential to implement a physical separation between the crystals in order to effectively mitigate the current leakage between the neighboring devices of OOSTAs. Kim and his colleagues reported the formation of transistor arrays by blade coating^[43]. This method allows for the isolation of 6,13-bis(triisopropylsilylethynyl) pentacene (TIPS-PEN) crystal patterns could form by controlling the blade speed, and the crystals were highly oriented (Fig. 4(b)). Furthermore, it is noteworthy that flexible and stretchable OOSTAs based on blade coating preparation have also been reported in the literature. In a related development, Shi et al. developed a flexible energy-saving OOSTA^[44]. The utilization of superhydrophobic fluoropolymer (CYTOP) patterns prepared through inkjet printing enabled the isolation of the units of OOSTAs, thereby eliminating the cross-talk effect between the adjacent devices. Subsequently, Dif-TES-ADT/PS solution was blade-coated to form the functional layer with defined patterns. Fig. 4(c) provides the detailed illustration of the aforementioned process. Fig. 4(d) shows the optical image of the OOSTAs. The OOSTAs demonstrated the image recognition and reinforcement capabilities, which will be discussed in more detail later. This strategy also permits the realization of stretchable phototransistor arrays. Li et al. selected poly [2,2’-[(2,5-bis(2-octyldodecyl)-3,6-dioxo-2,3,5,6-tetrahydropyrrolo [3,4-c] pyrrole 1,4diyl)] dithiophene-5,5’-diylalt-thieno [3,2 b] thiophen-2,5-diyl] (PDBT-co-TT)/styrene-ethylene-butylenestyrene (SEBS) as the solution for the blade coating process^[45]. A functional layer with a large area (10 × 10 cm²) was fabricated through proper blade coating speed (Fig. 4(e)). In contrast with the study of Shi et al., the semiconductor films were patterned by the laser ablation procedure. As illustrated in Fig. 4(f), the fabrication of isolated stretchable units in the array was successfully completed.

Drop casting is another feasible method for preparing functional material films of OOSTAs. In the process, a specific solution is dropped on the substrate. When the conditions are optimal, the solvent is naturally volatilized, facilitating the formation of a suitable film on the substrate. However, the quality of functional films is typically variable due to fluctuations in the surrounding environment. Chen’s group aimed to employ the drop casting method for fabricating the functional film of OOSTAs with minimal device-to-device variation and consistent performance. To prepare a high-quality, well-patterned TIPS-PEN crystalline array, they utilized a combination of nanoimprint lithography and pre-patterned guided crystal growth with a drop coating strategy^{[46, 47]}. The OOSTAs schematic diagrams are illustrated in Figs. 4(g) and 4(h). The AFM and SEM images demonstrated the isolated crystalline array (Fig. 4(i)). Their work provided a straightforward path for fabricating the OOSTAs.

The processes of coating and casting are attractive due to their feasibility, cost efficiency, and simplicity. Nevertheless, there are still limitations. The density of fabricated OOSTAs is typically relatively low. Moreover, these methods are inadequate for precisely patterning the functional material layer. The absence of isolation in the fabrication process would result in the cross-talk effect and current leakage between the continuous units of OOSTAs, thereby constraining the practical applications of the device^[48]. It is imperative to adopt more sophisticated methodologies for the fabrication of higher-density OOSTAs with discrete units.

Physical vapor deposition (PVD) represents an effective method for the preparation of functional films based on small molecules for OOSTAs. In this procedure, the evaporation of molecules occurs within a high-vacuum chamber when the thermal source attains a specific, optimal temperature value. Once the material vapor has diffused across the substrate, the molecules are able to condense and undergo crystallization on the substrate. The quality of the films can be regulated in a number of ways, including the temperature of the thermal source and the substrate. In particular, this strategy can be employed to achieve multi-layer deposition without concern for the damage to the subsequent layer caused by the solvent. Although this method needs expensive and complicated equipment, PVD is a prevalent technique in the fabrication of OOSTAs based on small molecules. In the early period, similar to solution processing functional materials, most research on OOSTAs has not paid attention to the segmentation and patterning of the organic active layer manufactured through the PVD (Figs. 5(a) and 5(b))^[49−51]. Once a complete film has been prepared by PVD, small-scale OOSTAs were fabricated by evaporating electrode patterns onto the functional film. Large-scale OOSTAs are more susceptible to crosstalk than small-scale arrays. As technology advances and the demand for high performance OOSTAs increases, well-patterned functional film of OOSTAs is essential. Proper shadow masks are integral to the small molecule patterning process during the PVD procedure^[52−54] (Figs. 5(c)−5(e)). However, the precision of shadow masks is also constrained, and the alignment of electrodes and functional layers is dependent on human intervention. In order to realize the precise isolation of functional film of OOSTAs, Kim et al. have taken steps to utilize the SU-8 with a thickness of 3 μm as the isolator for pentacene film (Fig. 5(f)). Hemispherical OOSTAs can be prepared through the developed scalable fabrication process^[55]. Nevertheless, it is exceedingly challenging to prepare a well-patterned film through the PVD process for polymer semiconductors. Consequently, there is a clear need for additional techniques to be developed for the fabrication of OOSTAs with precisely patterned polymer semiconductors.

As discussed above, it is difficult to produce OOSTAs with well-patterned functional materials by relying solely on coating or casting techniques. PVD is a good way to prepare the well-patterned functional materials. However, PVD is usually suitable for small molecules. Considering this, printing techniques are viable options for reproducing OOSTAs based on various materials, including polymeric material^[56−59]. Printing techniques can be divided into two types, contact and non-contact. Screen printing is a representative of contact printing. It relies on the screen mask to form the patterned functional films. The resolution of the patterns can be affected by many factors, including the quality of the mask, the viscosity of the solution, and the substrate. In recent studies, the screen-printing technique has been employed for the preparation of organic transistor arrays. Duan et al. demonstrated the preparation of flexible organic transistor arrays based on the benzothieno[3,2-b]benzothiophene (C8-BTBT)/poly(methyl methacrylate) (PMMA) or 2,6-bis(4-hexylphenyl)anthracene (C6-DPA)^[60] (Fig. 6(a)). Both transistor arrays exhibited high mobility and consistency (Fig. 6(b)). Furthermore, the screen-printing process can be employed for the fabrication of ion gels and electrode materials (Fig. 6(c)). Nevertheless, there is a lack of research on the optoelectronic performance of these materials^[61]. The screen-printing procedure represents a promising avenue for the fabrication of OOSTAs.

Inkjet printing is a prevalent non-contact printing method. Typically, inkjet printing employs either the piezoelectric or thermal inkjet to deposit functional material in specific patterns on the substrate. The spray position can be readily adjusted through digital control. This fabrication method offers a compelling approach to producing OOSTAs, offering advantages including high resolution, the absence of a mask, and the potential for material savings. A variety of materials are suitable for use as the ink. More importantly, the integration circuit can be fabricated through the technique of inkjet printing^[62] (Fig. 6(d)). Furthermore, the inkjet printing procedure allows for the production of complex structures of organic synaptic transistor arrays. Fang et al. fabricated vertical phototransistor arrays, with Ag nanowire electrodes and a functional layer, poly [2,5-bis(alkyl)pyrrolo[3,4-c] pyrrole-1,4-(2H,5H)-dione-alt-5,5’-di(thiophen-2-yl)-2,2’-(E)-2-(2-(thiophen-2-yl)vinyl)-thiophene] (PDVT-8), prepared by inkjet printing^[63] (Fig. 6(e)). The array demonstrated decent photoresponsivity. Wang et al. conducted subsequent work^[64]. They fabricated vertical ferroelectric OOSTAs based on poly(vinylidenefluoride-co-trifluoroethylene) (PVDF-TrFE) and dioctyl substituted perylene tetracarboxylic diimide (PVDT-10) (Fig. 6(f)). However, the photoelectric performance of the arrays was not extensively studied in either case, leaving scope for further investigation. Printing technology plays an indispensable role in the fabrication of OOSTAs, due to its high composition controllability, high throughput, compatibility with various large-scale substrates, and rapid prototyping at a low cost.

Photolithography is an efficient technique for the fabrication of OOSTAs. It is a high-throughput manufacturing process. The precision of the photolithography technique can be superior to that of printing techniques^[65]. Through the photolithography process, functional material can be patterned to reduce the device size, improve the consistency of the device in OOSTAs, and eliminate crosstalk in circuits. However, the majority of functional materials, organic semiconductors, have a low tolerance to solvent immersion and washing, rendering them incompatible with the photolithography process. Many subtle attempts have been made to address this issue.

Recent research has concentrated on the following areas: (ⅰ) The development of novel crosslinkers for semiconductors. The research group led by Cho has developed several universal crosslinkers, including 2Bx and 4Bx^[66−68]. Upon exposure to UV light, the azide groups of the crosslinkers are capable of generating the active singlet nitrene. Subsequently, the singlet nitrene inserts into the C−H bond within the alkyl chains of the adjacent polymers (Fig. 7(a)). Once crosslinked, the functional material is capable of tolerating solution environments and exhibiting compatibility with photolithography techniques. The Cho’s group employed the universal crosslinker to pattern multiple materials and fabricate high-performance organic transistor arrays (Fig. 7(b))^[66]. In line with the findings of the preceding study, Wu et al. patterned functional film employed the 4Bx crosslinker and fabricated the OOSTAs (Fig. 7(c)), which exhibited consistent photoresponsivity and synaptic-like behavior^[69]. Moreover, Zheng et al. employed the carbene insertion reaction as a universal strategy for the patterning of multiple materials. They fabricated high-performance organic transistor arrays comprising 42 000 units per square centimeter and an elastic functional circuit^[70]. The crosslinker may be applicable in the preparation of OOSTAs. (ⅱ) Mixing functional materials with photoresist. Wang et al. conducted a study in which semiconductors were mixed with the photoresist^[71]. During the photolithographic procedure, the blend film was precisely patterned, and subsequently, high-density organic transistor arrays were successfully fabricated. The p-type DPPDTT and n-type N2200 transistor arrays exhibited the desired performance characteristics. The use of SU-8 or PCell allowed for the feasible patterning of functional materials. In light of these findings, this method could also be used for the fabrication of OOSTAs. (ⅲ) The construction of nano-interpenetrating structures between semiconductors and photoactive materials^{[72, 73]}. This strategy is universal for polymeric semiconductors. Wei’s group employed this strategy to develop a semiconducting photoresist^[74−76]. The researchers combined a photocrosslinkable monomer, a photoinitiator, a thiol additive, and polymeric semiconductors to create a semiconductor photoresist. Upon exposure to ultraviolet radiation, the crosslinked monomers and semiconductors underwent a chemical reaction, forming a nano-interpenetrating structure. The solubility of the exposed film was found to be significantly diminished owing to its nano-interpenetrating structure. An organic transistor array with a density of 1.1 × 10⁵ units cm⁻² was fabricated (Fig. 7(d)). Moreover, they recently fabricated high-density (3.1 × 10⁶ units cm⁻²) organic phototransistor arrays based on a similar strategy. The arrays exhibited remarkable photoresponsivity. Imaging chips with an integrated scale of over 2²¹ units were manufactured using photolithography (Fig. 7(e)). (ⅳ)The employment of pre-patterned templates to facilitate the process. Guo et al. presented a method for patterning polymeric semiconductors^[48]. The photoresist was prepatterned to form the desired templates through the photolithographic procedure. The polymeric semiconductors were spin coated in the template. Subsequently, the photoresist template was stripped by the orthogonal solvents of polymeric semiconductors. Functional films with the requisite patterns were isolated and remained on the substrate. This approach was used to manufacture OOSTAs with a density of 6500 units per cm².

An increasing number of suitable strategies have been reported for the fabrication of OOSTAs, with the data-centric computing paradigm also being promoted based on these fabricated OOSTAs. The integration of the unit within OOSTAs paves the way for the realization of sophisticated systems, including neuromorphic visual systems and neuromorphic computing systems. These systems hold intriguing prospects and have attracted increased research interest. In this Section, cutting-edge neuromorphic visual systems and neuromorphic computing systems based on OOSTAs are summarized.

The retina is capable of perceiving the surrounding environment and acquiring visual information that is multidimensional in nature. Moreover, it is capable of preprocessing a multitude of visual data. Through this procedure, massive and redundant visual information can be simplified. Furthermore, important information is extracted. Inspired by the biological vision system, more effective methods for processing complicated visual data have been developed. The units of OOSTAs are organic optoelectronic synaptic transistors. They are capable of sensing light stimulation, acquiring visual information, and emulating the behaviors of biological synapses. By combining these functions, organic optoelectronic synaptic transistors have the potential to serve as the fundamental units of neuromorphic visual systems. However, it is challenging to develop a neuromorphic visual system based on a single organic optoelectronic synaptic transistor. Therefore, the demonstrated capabilities of OOSTAs in constructing neuromorphic visual systems are noteworthy.

In the early stage, OOSTAs were employed to emulate the learning and forgetting behaviors of biological synapses. The pixels of specific images were represented by light pulses. The number and intensity of the pulses were determined in accordance with the extent of the learning degree. Upon the application of light pulses, the units of OOSTAs exhibited distinct photocurrents in accordance with the number and intensity of the light pulses. When the heatmap of currents was demonstrated, specific images could be displayed. After a period of oblivion, the image did not disappear (Fig. 8(a))^[49]. Recently, there has been a notable increase in the demonstration of sophisticated OOSTAs that are capable of processing more complex tasks. A number of research projects have been conducted with the objective of developing flexible and stretchable OOSTAs with the intention of creating a functional emulation of the human eye. Shi et al. presented a demonstration of the OOSTAs on a PEN substrate, wherein the 8 × 8 synaptic transistors exhibited uniform photoelectrical characteristics. In order to emulate the human eyeball, the OOSTAs were affixed to an eyeball model, which served as the retina. Regardless of whether the OOSTAs are in a flat or bent state, they are able to clearly discern the number "8". Furthermore, pattern reinforcement can be achieved through the use of flexible OOSTAs, as demonstrated by the heatmap of the 64 pixels photocurrent (Fig. 8(b))^[44]. In consideration of the actual retina, it can be regulated by the ciliary muscle. The development of stretchable OOSTAs is a crucial step forward. Xu et al. have reported the development of intrinsically stretchable OOSTAs^[34]. The units of OOSTAs are composed of stretchable components, including the photoactive layer, the dielectric layer, electrodes, and the substrate. Upon stretching to 100%, the EPSC of the units triggered by light pulses exhibited minimal alteration. The letter "N" constituted by 4 × 4 pixels, was identified with high clarity (Fig. 8(c)). Furthermore, a sophisticated strategy for the construction of high-performance stretchable OOSTAs was reported by Kim et al.^[55]. Each pixel of the OOSTAs was established on a separate rigid SiN_x substrate. Subsequently, a polyimide film was used as the substrate. The rigid SiN_x island is beneficial for maintaining the flat state of the unit in the OOSTAs. A hemispherical image sensor was established. The letter "X" image was focused onto the OOSTAs and successfully recognized (Fig. 8(d)).

It is insufficient to merely handle these image recognition tasks. In addition to this, there is a need to address a number of other image processing tasks that are more versatile in nature. These include tasks such as associative learning, color filtering, motion perception, and vision adaptation. The research group led by Huang produced a series of high-density OOSTAs with a device density of 6500 devices cm⁻². In one of their studies, the degree of photoresponsivity to light with different wavelengths was employed to enhance the contrast between the useful information and the background noise (Fig. 8(e)). Furthermore, memory recall was achieved through the OOSTA, based on the memory behavior of the units^[48]. In another instance, the capacity for motion perception was also demonstrated by the constructed OOSTAs. The motion of items was encoded into a series of light pulses, which were then applied to the pixels of the OOSTAs. Following stimulation with a variety of light pulses, the photocurrents exhibited distinct responses at the pixel level. The trajectory of the motion could be readily discerned based on the heat map of the currents of the OOSTA, which reflected the temporary memory of the photocurrent. Another significant application of OOSTAs is in the field of vision adaptation. Wang et al. constructed a 5 × 5 stretchable OOSTA for the realization of scotopic and photopic adaptation. Moreover, the currents of the OOSTA remained stable even when the device was subjected to stress^[77]. Machine vision plays an indispensable role in the field of OOSTAs. It is anticipated that more sophisticated functionalities will be required in the future.

The application of artificial intelligence and deep learning algorithms is becoming increasingly prevalent across a multitude of fields. However, the predominant algorithms rely on software rather than hardware^[2]. Massive data are generated during the utilization of these algorithms. These operations are conducted on computers that are based on the Von Neumann architecture. It is essential to facilitate the transfer of redundant data between the computing unit and the memory unit. Accordingly, the computing speed is likely to be constrained by the latency, which is also known as the Von Neumann bottleneck. The brain’s computing and memory operations are conducted in parallel within a densely interconnected network of neurons. Synaptic weights are responsible for the memorization and processing of information. In light of these considerations, neuromorphic computation represents a promising avenue for addressing the limitations of the Von Neumann bottleneck. Analogous VMM serves as the foundation for artificial neural network (ANN) algorithms, which can be operated in hardware. In accordance with Ohm’s law and Kirchhoff’s law, the current value can be easily determined, which represents the VMM of the ANN algorithms (I_m = ∑G_n,mV_n). During the procedure, G_n,m is the conductance of the neuromorphic devices of the crossbar array, which is regulable.

In order to perform the VMM using hardware, a number of attempts have been made, including the development of multiple types of devices and the simulation of different algorithms^{[78, 79]}. The organic synaptic transistors display synaptic functionality and analog regulation of conductance comparable to those observed in biological systems, which suggests the potential for facilitating neuromorphic computation. The conductance modulating process typically manifests as long-term potentiation or long-term depression (LTP/LTD) (Fig. 9(a))^[79]. Furthermore, considering the need to accurately regulate synaptic weights between different neurons in the ANN, OOSTA with consistent units is crucial for future applications. In previous research, light pulses were employed for conductance modulation in the LTP process of organic optical synaptic transistors, while electrical pulses were used for conductance modulation in the LTD process. Huang’s group employed organic optoelectronic synaptic transistors to emulate a single-layer perceptron^{[59, 80]} (Fig. 9(b)). In these experiments, the weight update was represented by a change in conductance, which followed a specific formula. The results of these simulations offer preliminary evidence that organic photoelectric synaptic transistors can be employed as element units in hardware implementations of ANN algorithms. Nevertheless, the single-layer perceptron represents merely the initial stage of ANN algorithms. Additionally, Huang’s group emulated supervised and unsupervised learning algorithms based on single-hidden-layer ANNs. It is noteworthy that the weight update processing was formed through six synaptic transistors, including three positive synaptic weight states and three negative synaptic weight states in order to facilitate convergence during the training process^[81]. Six synaptic transistors can form more synaptic weight states, which is feasible for hardware. In another work of Huang’s group, both rigid and flexible OOSTAs were fabricated. The authors demonstrated that the ANN could be conducted based on the LTP and LTD of both rigid and flexible OOSTAs (Fig. 9(c)). Notably, the difference from the previous work is that units of the OOSTAs are employed to denoise the dataset based on the differences in photoresponses before the process of number recognition in the ANN. Noisy pixels in the image intended for recognition are the useless information. It has been demonstrated that the presence of noisy pixels in the images has a detrimental effect on the accuracy of image recognition during the ANN process. By utilizing differentiated photoresponses of the OOSTAs to diverse light wavelengths, the strength of noise dots in the handwriting number was attenuated. This approach resulted in a notable improvement in the recognitive accuracy for colored number images (Fig. 9(d)).

The development of OOSTAs provides a possible way to emulate the perceiving mechanism of human brains at the hardware level. The OOSTAs have shown the intriguing prospects in constructing neuromorphic visual systems and neuromorphic computing systems. However, the development of OOSTAs in practical applications is still in its early stages. There are a few factors that can advance the development of OOSTAs and drive their practical application.

The development of high-quality materials exhibiting wafer-scale growth and mature fabrication technologies represents a crucial step for the scalable preparation of large-scale OOSTAs with high throughput. Furthermore, intrinsic properties of these materials, such as high absorption, eco-efficiency, and quantum efficiency, are also essential for the development of high-performance OOSTAs. Notably, the bandgap of organic materials can be regulated, resulting in the desired change to their optoelectronic properties. Optimized functional materials facilitate the achievement of low dark current and high signal-to-noise ratio in OOSTAs, thereby enhancing their optoelectronic performance and facilitating the efficient processing, transmission, and storage of information.

The stability of functional materials employed also has an impact on the overall stability and reliability of OOSTAs. It is imperative that OOSTAs demonstrate consistent performance across devices and cycles in order to meet the requirements of their intended applications. Meanwhile, photolithography is a critical strategy to fabricate the high-throughput OOSTAs.

To be compatible with photolithography, novel materials need to be developed. The development of photo crosslinkers is required to enable the universal crosslinking of a range of functional materials without compromising their performance. Additionally, the utilization of organic materials with photo-crosslinkable functional groups is imperative for the direct patterning of functional materials during the photolithography procedure.

Energy consumption is another criterion for OOSTAs. Reducing the size of devices, decreasing the duration, and lessening the intensity of programming spikes are effective ways to diminish the energy consumption during the employment of OOSTAs^[82]. Moreover, OOSTAs based on optimized functional materials can realize high photosensitivity, leading to the further reduction of the energy consumption of OOSTAs.

Moreover, it is vital to evaluate the balance between the photoelectric performance of OOSTAs and the associated manufacturing cost. Wafer-scale candidate materials should be developed and validated to ultimately transfer them from lab environments to production facilities.

In the context of neuromorphic computing utilizing the ANN, the presence of both positive and negative synaptic weights is a fundamental necessity. In order to generate negative synaptic weights, it is possible to employ subtraction operations to create a set of "fake negative weights". In order to achieve compact integration, it is essential to pursue the integration of positive and negative weights through a single synaptic device. For OOSTAs, a single unit with both negative and positive optical response is an attractive option for future applications and should be developed accordingly. Additionally, LTP represents the consolidation of neural connections. After each pulse is applied to the synapse, the LTP of the synapse can be strengthened. However, the LTP may dissipate with the removal of stimuli and the passage of time. The retention of synaptic weight, including LTP/LTD, is required over a range of time periods, from a few seconds to several hours, in various neuromorphic computing applications. Although online training has less rigorous requirements for retention, batch training requires high levels of retention. It is therefore essential to develop OOSTAs with suitable retention properties of units. Concurrently, the development of algorithms specifically designed for the use with OOSTAs is also a necessity. In contrast to the majority of ANN algorithms based on linear operations, OOSTAs are capable of non-linear responsivity. The incorporation of nonlinear operations into sensor computation can enhance performance and fully leverage the potential of optoelectronic technology.

With regard to the units in OOSTAs, the integration density of the devices in OOSTAs is dependent upon the manufacturing technologies employed. Photolithography represents an effective approach to the fabrication of OOSTAs with high-density devices. However, organic functional materials are typically incompatible with the photolithography process due to its low tolerance to organic solvents, which involves during the manufacturing process. It is of the great importance to pursue meticulous methods in order to establish a connection between the photolithography process and the preparation of large-scale OOSTAs.

From the perspective of OOSTA integration, the signal-to-noise ratio and the integration level of OOSTAs may be influenced by the integration of processing circuits and readout circuits in the plane. The monolithic-3D architecture offers considerable promise as a strategy for enhancing the performance of OOSTAs, accommodating the substantial bandwidth requirements, and implementing a data-centric architecture. In the context of monolithic-3D architecture, the processing units and readout circuits can be situated in a position below OOSTAs, thereby increasing the level of integration. Moreover, high-performance selective components are also required to eliminate sneak currents. In addition, the high cost and thermal effects caused by compact interconnection in monolithic-3D architecture represent significant obstacles. It is therefore imperative to pursue optimal interconnection strategies for high-performance OOSTAs in monolithic-3D systems in the future.