Recent progress in optoelectronic memristors for neuromorphic and in-memory computation

Neuromorphic computing has been gaining momentum for the past decades and has been appointed as the replacer of the outworn technology in conventional computing systems. Artificial neural networks (ANNs) can be composed by memristor crossbars in hardware and perform in-memory computing and storage, in a power, cost and area efficient way. In optoelectronic memristors (OEMs), resistive switching (RS) can be controlled by both optical and electronic signals. Using light as synaptic weigh modulator provides a high-speed non-destructive method, not dependent on electrical wires, that solves crosstalk issues. In particular, in artificial visual systems, OEMs can act as the artificial retina and combine optical sensing and high-level image processing. Therefore, several efforts have been made by the scientific community into developing OEMs that can meet the demands of each specific application. In this review, the recent advances in inorganic OEMs are summarized and discussed. The engineering of the device structure provides the means to manipulate RS performance and, thus, a comprehensive analysis is performed regarding the already proposed memristor materials structure and their specific characteristics. Moreover, their potential applications in logic gates, ANNs and, in more detail, on artificial visual systems are also assessed, taking into account the figures of merit described so far.


Introduction
It is becoming increasingly urgent to find an alternative to the long-time used Von Neuman architecture, in which the bottleneck of transferring data between the processor and the memory unit remains a severe issue and blocks the realization of complex problems with high speed and accuracy. Moreover, in extensively data-centric applications like artificial intelligence (AI), novel computational paradigms are vital to bring down power consumption. Neuromorphic computing based on resistive switching (RS) devices, or memristors, is a brain-inspired technology that has emerged as a viable ultra-low power consumption alternative. In this technology, information processing is performed directly at the memory element, preventing data shuffling and, therefore, enabling cost and power-efficient, real-time in-memory computation.
The memristor, discovered by Chua [1], is a non-linear resistor, usually composed of a metal-oxide-metal structure in which the metal layers compose the electrodes and an oxide film is referred to as active layer. Typically, these devices are fabricated in a vertical stack which allows for further miniaturization when compared with technologies such as Static random-access memory (SRAM), dynamic random-access memory (DRAM) or flash, as evidenced by the smallest 4 nm 2 device reported up to now [2]. The most interesting feature of the memristor is the fact that its resistance state can be tuned by an external electric field into two or more distinct states. The switching process occurs at the active layer and the transition from the highest-resistance state (HRS) to the lowest-resistance state (LRS) is known as set operation and from LRS to HRS, as reset operation. An exceedingly fast 50 ps switching speed has been obtained which is, naturally, of extreme importance to ensure power efficiency [3,4]. The retention of the programmed states is related to the material structure of the device and can be translated into short-or long-term memory behavior, suited for different neuromorphic applications.
The first fabricated memristor was reported in 2008 with TiO 2 as active layer and Pt as electrodes [5]. To date, the most popular materials for oxide-based memristive devices are still the transition metal oxides, for which the switching occurs, most commonly, in a filamentary form, such as tantalum oxide (Ta 2 O 5 ), hafnium oxide (HfO 2 ) or aluminum oxide (Al 2 O 3 ) [6][7][8]. Typically, the transition between different resistance states is an abrupt one, as exemplified in figure 1(a). An initial step, called electroforming, is required to create conductive filaments (CFs), usually composed of oxygen vacancies, within the insulating layer. This is, usually, accomplished with a once electric bias application of higher voltage than the operating one. Once these filaments are created, current can pass through which decreases the resistance of the device, now at the LRS. Then an operating voltage will cause the dissolution and consequent restoration of CFs resulting in set and reset operations. The CFs are typically robust and can only be dissolved by the operating voltage which results in long retention of programmed states. In the case of valence change memories, the dominant mechanism is determined by the balance between thermochemical and electrochemical redox reactions, which is mostly defined by the combination of the materials used in the device stack [9][10][11].
As opposed to abrupt-RS, analog switching behavior, displayed in figure 1(b), can be achieved using, for example, oxide semiconductors such as zinc oxide (ZnO), indium-gallium-zinc oxide (IGZO) or zinc-tin oxide (ZTO) as active layers [12][13][14][15][16]. Typically, this type of switching occurs at the interface and is attributed to an interchange or trapping of ions between a metal electrode and the oxide [17]. It, usually, results in a lower device-to-device (D2D) and cycle-to-cycle (C2C) variability, since an area-dependent behavior with a uniform contribution of defects is in place. Additionally, the programmed resistance state is, usually, not retained for a long period of time and is, instead, gradually replaced by the equilibrium state of the oxide. This retention time can be manipulated with the amplitude or number of programming electric pulses [18]. This behavior is unsuited for inference-only applications in which a stable retention of at least two well-separated states is required. However, it can be advantageous in the realization of synaptic functions.
Moreover, the manipulation of several distinct resistance states can be accomplished by both the abrupt and the analog RS and is preferred for in-memory computation tasks in stateful logic and/or artificial neural networks (ANN) loosely inspired by the human brain. An ANN hardware composed by interconnected and adaptive electronic elements was first envisioned in 1962 [19] and can be accomplished with RS devices organized in crossbars to perform parallel data processing. The human brain has 10 11 neurons that communicate between each other through more than 10 14 synapses, in which their connection strength varies [20]. In deep neural networks (DNNs) composed by memristor crossbars, synaptic behavior is mimicked when electric short pulses are continuously applied to the device to gradually decrease or increase its resistance state, in processes called potentiation and depression, respectively. In fact, several synaptic characteristics have already been emulated by RS devices [13,21,22] proving its overwhelming potential to substitute the classical computing system.
Nonetheless, memristor crossbars that operate under only electric stimuli often suffer from heat generation and crosstalk issues, related to the interference of neighboring devices, which is blocking high-scale implementation [23]. Crosstalk can be eliminated with the addition of a selector transistor in each crossbar cell. However, this approach undoubtedly increases cell area and energy consumption [24]. In fact, the biological system cannot be matched with purely voltage-assisted neuromorphic platforms. The human brain itself is stimulated by a number of sensing inputs such as light, sound, smells or touch, with about 80% of this information being acquired through visual perception [20]. Hence, an ideal AI system should merge the capabilities offered by electrical and optical domains [25].
Optoelectronic memristors (OEMs) use both light and electronic signals as inputs and can behave as sensory artificial synapses with high energy-efficiency, low crosstalk and fast data processing and are, thus, suitable to link artificial visual systems with its processing on hardware [26]. In detail, OEM technology offers the possibility of a non-electric programming method, and can, thus, be an enhanced solution to challenges such as excessive Joule heating and crosstalk that occur in high-density passive crossbar architectures. Beyond that, learning functions and signal transmissions cannot be simultaneously performed on devices operated solely with electric field. Adding light as a third terminal facilitates parallel data transmission and processing. Hence, the applications of such a system can be extended to all mobile computing applications where the near/in sensor computation concept is developed such as wearables, visual information processing, invisible touch screens, electronic eyes, and smart processors.
To build an OEM, firstly one must consider the targeted application, since a particular set of requirements related to its performance will be in place. Figure 1(c) illustrates the most relevant applications of reported OEMs in literature. As a photodetector, the optical induced change in resistance should only be detected during light input since the purpose is to reveal when there is, in fact, light irradiation. Contrary, in DNNs a quasi-full state retention after light pulse is mandatory for the highest accuracies in any task. In its turn, spiking neural networks (SNNs) is the paradigm closer to accurately simulating brain functionality and demands a dependent state retention on several factors such as number or frequency of applied input light pulses to mimic learning and forgetting capabilities of the human brain.
The retention of the light provoked resistance state is, usually, ascribed to a persistent photoconductivity (PPC) effect that is, often, inherent to the active layer material [27][28][29][30]. The PPC effect is, typically, attributed to (i) macroscopic physical barriers at the interfaces or doping defects that decrease the rate of recombination of photogenerated electron/hole pairs [31] or (ii) atomic scale barriers located at the large lattice relaxation sites in which the energy level of the empty defects is above the conduction band minimum and the occupied energy level is located in the band gap, preventing recombination [29,32].
Consequently, the choice of memristor material structure is of the most importance. There should be at least one transparent electrode, such as indium-tin oxide (ITO) or aluminum-doped ZnO (AZO) [33][34][35][36], or a nm thick semi-transparent electrode, such Au, to allow light to reach the active layer. Moreover, the bandgap of the active layer material should be considered for proper light wavelength irradiation, as depicted in figure 1(d). Naturally, a lower bandgap material is desired for wide sensitivity to multiple wavelengths. However, these materials often suffer from inferior retention stability.
As evidenced in the graph of figure 1(e), where the number of publications in optoelectronic and photonic RS devices is presented pear year, there is an obvious growing interest in researching OEMs, as their potential becomes more and more clear. However, as it is still a novel topic, only few review articles can be efficiently consulted for a comprehensive analysis on the recent advances.
This review paper aims to fill in this gap, providing a summary on the developments on inorganic-based OEMs fabricated in a vertical stack, with focus on metal and semiconducting-oxides, 2D-based materials and 1D structures as active layer. The planar structure has also been widely explored [37][38][39][40][41][42], however for high-scale integration and power-efficiency, the vertical stack is the favored one. Organic materials have also been widely investigated for optoelectronic features [43][44][45][46] due to their advantages in low temperature and simple solution processing techniques. In addition, organic molecular tuning is possible and desirable which leads to certain optical and electrical properties [47]. Nevertheless, organic materials are not, usually, as stable as inorganic ones mainly due to oxidation issues. In optoelectronic applications, metal oxide semiconductors are the ones which show high mobility and sufficient mechanical and electrical stability, besides transparency in some cases [48]. Furthermore, recent works of inorganic materials show immense progress in low-cost fabrication techniques using solution and printing methods [49]. Other classes of materials explored as active layer for optoelectronic features are ferroelectric or magnetic [50][51][52][53] and organic/inorganic perovskites materials [54][55][56][57][58][59][60][61][62][63][64]. However, most of the available studies are at the device level and, therefore, only a small sum have demonstrated synaptic functions [40,[65][66][67] or logic gates implementation [35,[68][69][70][71].
The potential applications of OEMs based on the targeted materials in logic gates, ANNs and, in more detail, on artificial visual systems are also evaluated here, with the assist of the reported individual features. Additionally, the current challenges and future perspectives in this field are also discussed.

Metal oxide OEMs
The first report on controlling the resistance state in a memristor using light in combination with voltage signals was in 2012 by Ungureanu et al [72]. An active layer of Al 2 O 3 on a Si/SiO 2 substrate was used with patterned Pd as top electrode. Under dark, there was barely any RS and the device could be considered a resistor, but combining −10/10 V electric pulses and a 6 V read pulse with ultraviolet (UV) or infrared (IR) light pulses, multiple resistance states could be achieved. As soon as light was switched off, the dark current value immediately returned, showing no PPC effect. However, for data storage and short/long-term memory applications, PPC effect is required. This long-lived effect after light exposure is beneficial to modulate a non-volatile multi-resistance state.
Later, HfO x was proposed using Si substrates as bottom contact and Au [73] or Pt [74] as top electrodes. In these preliminary studies, PPC effect was not shown or discussed. In fact, the authors show RS behavior only under light irradiation. The former showed RS behavior under a broadband halogen-tungsten light source with multiple resistance states dependent on light intensity. The latter showed visible light sensitivity, with blue irradiation enabling RS characteristics. In this case, the optically boosted carriers transfer from Si together with electrically redistributed oxygen vacancies within the oxide, result in the switching behavior under light. Since these devices show memristive switching characteristics only under light, logic gates could be targeted as application, in which light is irradiated constantly from surroundings.
Light can be used as a third input to control set or reset operations in OEMs for a more power-efficient and fast approach. For instance, light irradiation was found to perform reset in a HfO x -based memristor [75]. The proposed structure, illustrated in figure 2(a), presented a typical bipolar and abrupt switching behavior in dark, as shown in figure 2(b), with a high I LRS /I HRS ratio of more than 10 3 . When the device was set in the LRS, visible light irradiation forced a transition to a higher resistance state and depending on its intensity, could perform a full reset. Figure 2(c) presents a graph where multiple resistance states can be distinguished that correspond to different light intensities for multilevel cell (MLC) operation. The different resistance states were proven to be non-volatile, which is related to the light-sensing mechanism. As explained before, this abrupt-type of switching is, usually, associated with the formation and rupture of CFs with different dimensions located within the oxide layer bandgap. A small CF is composed of a low amount of oxygen vacancies and, therefore, requires low light intensity for recombination with the surrounding oxygen interstitials, which causes filament rupture. A large CF is formed by a higher amount of oxygen vacancies and, thus, requires light of higher intensity to be annihilated. Additionally, it was observed that blue light triggers the same reset effect as white light while the longer, less energetic, wavelength of red light did not incite any reaction. The dependency on light wavelength was suggested to be associated with the position of oxygen vacancies inside the HfO x bandgap after set. In fact, a wide-bandgap oxide such as HfO x (E g = 5.7 eV) is by nature not photo-responsive. However, photoexcitation of defects located at a level of ≈2-3 eV below the conduction band minimum is possible and would bring the device back to its initial HRS [75].
Similarly, More et al proposed a titanium oxide (TiO x )-based memristor in which illumination with a tungsten lamp could decrease V Reset from 3.5 V to 0.8 V, reducing energy consumption [76]. More recently, a Pt/TiO x /Ti memristor was reported to respond to multiple irradiation wavelengths, ranging from UV to red light, in a completely different manner [77]. In this work, an analog-type of switching was observed for Pt/TiO x /Ti devices. The current state increased as the irradiated wavelength shortened, particularly for blue and UV illumination, which is in agreement with TiO x bandgap of ≈3.2 eV. Photons with shorter wavelengths provide enough energy to excite more electron-hole pairs within the TiO x layer.
A different mechanism for photoexcitation was proposed by Zhou et al, related to a Pd/molybdenum oxide (MoO x )/ITO memristor structure [78]. In this study, a light induced manipulation of Mo valence states in the MoO x thin film explained the photonic set. UV light irradiation caused the generation of electrons and holes in the MoO x layer. As photogenerated electrons transferred to the MoO x conduction band, a reaction would take place between photogenerated holes and absorbed water molecules that produced protons (H + ). Under surface science studies, authors revealed that the photogenerated electrons together with the protons would trigger the creation of H y MoO x and incite a shift from Mo 6+ to Mo 5+ , which ultimately resulted in the increase of current to LRS. When light was switched off, the current state did not return to HRS and, instead, suffered a slow decay (PPC effect). During electronic reset, an electric field forced the protons away from the MoO x layer towards the Pd electrode and Mo 5+ changed back to Mo 6+ , causing a decrease in current state to HRS.
A bi-layer structure of Ta 2 O 5-x /Ta 2 O 5 , sandwiched between an Si substrate and an Ag top electrode, demonstrated significant potential as an OEM [79]. This device showed a dependent current increase on UV and visible light irradiation with full retention of the photo-produced states, proving light information can be stored in situ. The bandgap of Ta 2 O 5 is ≈4.0 eV, and therefore, it responds mostly to UV light. However, due to oxygen vacancies in the Ta 2 O 5-x film, the electrons at the defect energy level can still be forced to the conduction band by lower energy wavelengths.
Therefore, a bi-layer switching medium can be used to enhance the performance of optoelectronic features, similar to electric field-induced counterpart. The AlO y /Cerium oxide (CeO x ) memristor with Al as bottom and ITO as top contacts-figure 2(d)-was reported by Tan et al [80][81][82]. A native AlO y layer was created upon deposition of Al to form the Schottky junction with the CeO x , that resulted in analog switching behavior, shown in figure 2(e). The photo-sensitive mechanism was explained by electrons trapped in the oxygen vacancies at the AlO y /CeO x interface. These electrons could be excited by photons and leave positively charged oxygen vacancies. This would lower the effective Schottky barrier by reducing its width, which would cause a decrease in the resistance of the junction. This memristor structure presented photocurrents dependent both on light intensity and wavelength, as shown in figure 2(f). Naturally, shorter wavelengths and higher light intensity triggered the highest current change. Moreover, PPC effect could be noticed as, even after light pulse, the current state remained unaltered. A voltage reset was then required to fully erase the resistance state.
Additionally, metal-oxide materials have been used in double-layer strategies in combination with semiconducting-oxides for switching medium in OEMs and will be discussed in the following section. In table 1, the comparison of the most important characteristics in metal-oxide OEMs can be consulted and it is noticeable some devices show optical set/reset coupled with electric reset/set, respectively. Naturally, photo-induced switching can be abrupt or analog depending on material stacks. The working principal of analog optical memristors is mainly related to the change of barrier profile at the interface of the electrode and the switching layer, which resembles to analog electric field-induced RS devices. Abrupt optical RS shows higher retention stability, however, inferior C2C and D2D reproducibility is expected due to the switching dependent on CFs. When reset is induced by light, the resistance state obtained is always maintained after light pulse. However, this is not considered PPC since, in this case, light is used to partially/fully break CFs and not to increase the conductance state.
No fully photonic switching behavior has been demonstrated within the same device, which hinders their possible application in wire-free crossbars. Moreover, as demonstrated in figure 2(g), most studies report on UV or blue irradiation which makes the discrimination between signals of different wavelengths challenging. Future technology will rely on visible light communication, where the application of abundant light sourcing devices (lamps, mobile displays, TV, etc) will be very important in the concept of ambient intelligence.

Oxide semiconductors OEMs
Oxide semiconductors can provide additional advantages as active layer on memristors. In particular, amorphous oxide semiconductors (AOS), such as IGZO or ZTO, have high flexibility, due to a lack of grain boundaries, which enables low-processing temperatures. This facilitates the use of flexible substrates (paper, polyimide, polyethylene naphthalate (poly(ethylene 2,6-naphthalate) (PEN) / polyethylene terephthalate (PET) [24,83,84], providing the means for embedded flexible technologies in wearables and internet-of-things applications. Furthermore, AOS are used in flat-panel display technology for active-matrix driver circuits, making them ideal candidates for future integrated memristor-based hardware for AI.
ZnO has been the most popular material as switching medium for OE devices. Using Pt and ITO as bottom and top contacts, Xie et al showed a decrease in V Set and V Reset upon UV light for more efficient power consumption [85]. Contrary, Shih et al showed an increase in V Set and V Reset which is not very appealing [86]. However, PPC effect was present in these devices, as a change in RS behavior was observed even after UV light was turned off, which means ANNs could be targeted as applications. Moreover, the memristor showed improved I LRS /I HRS ratio after UV illumination. The authors explained that the conduction mechanism changed from Schottky emission to Poole-Frenkel after UV illumination, which induced the creation of extra oxygen ions and radicals due to an accelerated breaking of Zn-O bonds.
The analog behavior of the ZnO memristor was also studied using different electrodes for OE applications [87,88]. UV light was proven to increase the current state with a PPC effect. Interestedly, one of the reported OE structures for analog-type of switching, ITO/ZnO/Ag [88], was also described by another group yet with an abrupt-type of RS [89]. On both reports, the sputtering technique is employed to deposit ZnO and similar thicknesses for all composing memristor layers are used. However, ZnO deposition conditions are different which can explain the observed results. Generally, electrochemical mechanism (ECM) is responsible for the RS in devices with active electrodes such as Ag and Cu. In this respect, depending on the electric field, switching can be tuned to be abrupt for higher field or analog, depending on the size of metallic filament. The analog RS behavior could be also related to interface properties of ZnO with the contacts which is directly related to deposition conditions. In fact, in the original works, the reported analog switching for this memristor structure is associated with oxygen vacancies filaments whereas the abrupt switching is related to ECM with Ag filaments [88,89].
A semi-transparent ZnO OEM, with Pt and Au as electrodes, that could perform both set and reset with optical inputs has been developed, with the structure illustrated in figure 3(a) [90]. In detail, Yang and co-workers showed set could be induced by UV, blue or green irradiation with PPC characteristics, showing MLC operation. UV illumination resulted in the highest current increase with an I LRS /I HRS ratio of ≈10 4 . Nonetheless, less energetic wavelengths such as green light could also induce a sufficiently high I LRS /I HRS ratio of ≈10 2 , as can be seen in figure 3(b). ZnO has a wide bandgap of 3.2 eV. However, deposition in pure Ar atmosphere can result in a larger amount of oxygen vacancies along with material heterogeneity with different energy levels [91]. Visible light irradiation can, thus, induce an ionization of oxygen vacancies, converting them into positively charged oxygen vacancies and free electrons which leads to an increase of the current state, as commonly accepted for AOS materials [92]. Interestingly, the retention time (or PPC effect) was also impacted by light exposure. The retention decay was fastened by irradiation with longer wavelengths such as red and near-IR (NIR) illumination, leading to an optical reset, in the scale of seconds (see figure 3(b)). This photonic reset was explained with an injection of electrons from the metal to the ZnO layer, by internal photoemission or photo-assisted tunneling. Part of these electrons neutralized ionized oxygen vacancies, decreasing its density, and fastening the PPC decay.  Similarly, the same research group showed photonic set and reset in a double layer IGZO memristor, illustrated in figure 3(c) [26]. The analog RS behavior was attributed to the formation of a potential barrier and well at the interface of a layer with low oxygen (LO) content (IGZO (LO) ) and a layer with high oxygen (HO) content (IGZO (HO) ). The width of this interfacial barrier would be dependent on the ionized oxygen vacancies density. The lower this density, the smaller the barrier, which facilitates electron tunneling and increases the current state of the device. On the other hand, during reset, the neutralization of ionized oxygen vacancies would increase the barrier width and result in a decrease of current state. Set could be induced by light in wavelengths from 420 to 1000 nm. Naturally, the longer the wavelength used for irradiation, the less change the current state would suffer. It was found that the PPC decay time could be decreased (reset) by red or NIR (800/900 nm) light, only if the previous set process was performed by short wavelengths illumination (350/420 nm), as can be seen in figure 3(d). To explain this phenomenon, the authors proposed that during illumination, the photocurrent would depend on a dynamic equilibrium between ionization and neutralization of oxygen vacancies (figure 3(e)). When the memristor was exposed to visible light of higher energy to perform set, a small amount of oxygen vacancies was left that could be further ionized by subsequent irradiation with light of lower energy. In this case, a neutralization of oxygen vacancies would, therefore, be dominant and enable a decrease in the current state-reset.
Later, a double-layer ZTO OEM was described, following the above explained strategy of oxygen rich and deficient layers [93]. Set was performed electrically with a ≈ 10 2 I LRS /I HRS ratio. Reset could be performed with MLC properties via blue, green and red light irradiation and the I HRS was not only dependent on wavelength, but also on light power. It was clarified that light would provide the necessary energy for interstitial ions to recombine with electrically produced oxygen vacancies within the ZTO layer, causing CFs to disrupt. Interestedly, analyzing the DC sweep displaying RS behavior, set and reset transitions do not seem abrupt. However, the CF formation/annihilation explanation for the RS behavior is aligned with the full 10 5 s retention of states obtained for this memristor structure. The same behavior was observed in a bilayer memristor with ZTO/TaO x (structure illustrated in figure 3(f)), in which MLC could be achieved by different power of blue light irradiation, as can be seen in figure 3(g).
In table 2, a list is provided with the most relevant features on the, so far reported, OEMs based on oxide semiconductor materials and bi-layers. To our knowledge, this class of materials was the only one so far not relying on nanostructured materials to demonstrate full photonic RS behavior with set and reset being performed by light with different wavelengths. However, as evidenced by the statistical percentages on different light-induced RS presented in figure 3(h), the research is, still, mostly on UV and blue light irradiation. Since IR light is the transmission method for data in wireless systems, it is highly desirable that further research on longer wavelengths is realized to allow conversion and storage of IR data [94]. High-speed process is one of the main challenges for further development of OEMs, relying on the working principal as mentioned in most of the reports. In this respect, there are several solutions such as increase the built-in potential by engineering the memristive device layer stack, either via defect engineering of the RS layers or via modification of the effective work function at the contacts. The second main challenge is the bandgap tunability of oxide thin films for a broader spectrum. Here, the same approaches reported for visible light AOS-based photodetectors can be adopted. For instance, an absorption layer composed of hydrogen-doped IGZO revealed visible light detection [95]. RGB-color detection is also possible with strategies such as using a system combining an AOS film with colloidal quantum dots (QDs) [96], as will be later discussed.

2D-materials OEMs
Two-dimensional-materials such as graphene (G), transition metal dichalcogenides (MoS 2 , WS 2 , WSe 2 , MoSe 2 , MoTe 2 , etc) or black phosphorus (BP), have also been widely explored for OEMs due their excellent photoresponsivity [104], related to a large surface-to-volume ratio, and RS switching performance [105,106]. A review paper from 2019 can be consulted, where mainly planar 2D-based structures for OEMs are described in detail, to that date [94]. Here, as previously explained, the focus is on vertically stacked devices, which were mostly reported more recently.
In this regard, a monolayer of MoS 2 was used as switching medium in Si substrates and W top electrodes [30]. UV illumination was used to perform set with one order of magnitude of I LRS /I HRS ratio and PPC effect was observed related to a spontaneous trapping/detrapping of electrons. However, the necessary voltage to perform reset on these devices was high (−8 V). Since this monolayer has a bandgap of 1.8 eV, it would be interesting to test these devices under IR illumination. In another study, WO 3 OEMs showed enhanced performance under light input [107]. A combination of a 600 • C annealing and 45 min of white light irradiation resulted in an improved I LRS /I HRS ratio that went from 1.7 (as-prepared device) to 25.
Later, Jaafar and Kemp demonstrated a current increase induced by UV light in a GO based-memristor [108]. Here, a planar structure was compared to a vertical one. The latter showed superior features with an abrupt RS of ≈10 4 ratio. Moreover, during UV irradiation, both V Set and V Reset were reduced. This behavior was attributed to photogenerated electron-hole pairs which increased the current state in both LRS and HRS. A PPC decay was also shown, that can be useful for ANNs applications.
Reset induced by light was also reported on 2D OEMs. For instance, an heterostructure of MoSe 2 /Bi 2 Se 3 nanosheets was embedded in a polymethyl methacrylate (PMMA) film to form the active layer on a device that showed reset under NIR light [109]. PMMA was included in the active layer to passivate the trap states on MoSe 2 /Bi 2 Se 3 . The full device structure is illustrated in figure 4(a). The memristor showed abrupt RS behavior assigned to Joule-heating provoked formation/disruption of metallic CFs. The authors explained NIR light would facilitate the rupture and prevent the creation of CFs since the Ag clusters would react with photogenerated holes and oxidize back to Ag + . In the graph of figure 4(b), the current state of the device with multiple stimulus can be evaluated in time. It can be seen that the current decreases to the full-HRS only under both NIR illumination and electric reset.
A different mechanism for reset by light was proposed by Zhou et al regarding a memristor based on BP nanosheets coated with polystyrene (PS) (BP@PS) [110], with the structure illustrated in figure 4(c). The DC sweep displaying set and reset under different light wavelengths can be consulted in figure 4(d). It is clear that a higher energy illumination, further decreased V Reset from 2 V (dark) to 1.17 V (UV), as summarily presented in figure 4(e), and achieved a lower current state in the HRS, enabling higher I LRS /I HRS ratio. The authors explained that during electric reset, the external electric field would change the direction of the internal electrical field (induced by a transfer of electrons from the BP@PS to the ITO layer, due to the different material work functions), forming a Schottky barrier. When the memristor was exposed to light, the trapping sites would capture the photogenerated electrons and the surface potential would increase, further increasing the Schottky barrier, and, therefore, increasing I LRS /I HRS ratio. Additionally, as the scattering centers were negatively charged by the captured photogenerated electrons, the internal electric field would increase and a lower V Reset would be necessary to switch off the device.  The research on 2D-materials for vertically stacked OEMs is still at its infancy and requires further investigation. A complete review on 2D-based memristors and its challenges and future perspectives can be consulted in [111], in which the need for the development of proper 2D material films, that can be employed in high performance devices in a large-scale, is emphasized. Once this is accomplished, it is expected that this type of OEMs will offer a faster and more energy efficient switching, due to the atomically thin characteristics of 2D materials [112]. In fact, the switching speed is not yet widely investigated in the small amount of available literature on vertically stacked OEMs based on 2D materials. As can be seen in table 3, so far, mostly abrupt RS related to trapping and detrapping of electrons was described, which is more advantageous for in-memory computation and DNNs applications. NIR light irradiation has been explored in a higher percentage when compared with other classes of materials, as evidenced by figure 4(f), mostly for reset operation. In fact, it has been previously observed that 2D-based photodetectors can be built for Terahertz detection [113,114] which can be very promising for memristors. To our knowledge, no fully-photonic 2D-based memristor has been proposed.

1D-structured OEMs
Another strategy to build an OEM is to take advantage of 1D structured materials, such as NRs or NPs ( figure 5(a)), in order to achieve enhanced performance. Compared with thin-films, nanostructures provide higher surface-to-volume ratio and larger photon harvesting, that ultimately results in improved photoconductivity, with highly localized RS [117]. Another major advantage is the potential for miniaturization, as these nanostructures can usually be produced by bottom-up approaches allowing for a rigorous control of structures dimensions [118,119].
The first report on a 1D-structured OEM dates back to 2012. Park et al reported on a ZnO NR-based device with FTO and Au as electrodes [120]. Showing an abrupt-behavior, RS was only observed when the device was illuminated by a xenon light source and was ascribed to oxygen vacancies filament formation/rupture. In the dark, oxygen molecules would be chemisorbed onto the surface of ZnO NRs acting as an oxygen source and preventing the oxygen vacancies filament formation. Upon light irradiation, electron-hole pairs would be created, and the holes would combine with oxygen ions adsorbed at the ZnO, allowing for CFs creation. Moreover, the modulation of light incident angle was shown to regulate RS in ZnO NRs [121]. In this work, a surface treatment was carried out to induce hydrophobic properties on ZnO NRs, inducing the formation of an air bubble protecting the memristor when emerged into water. The device  behaved as a resistor with no RS properties when the light beam incidence surpassed a critical angle at the water/air interface, because total internal reflection would occur and light would not reach the ZnO NRs. When the light beam incidence was below the critical angle, RS was observed. In another study, the photoconductivity of a ZnO NRs-based device was calculated to be about 10 4 and compared to the much lower 70 achieved in a ZTO thin-film based memristor [122]. Furthermore, an heterojunction composed of ZnO and phosphorene (ZP) NPs showed a decreased V Set under irradiation of UV, green, red and IR light, with the possibility of MLC [123]. In another work, ZP NPs were sandwiched between two PMMA layers and an analog RS was observed [124], in which the current state could be increased by UV, green and red light. However, no PPC effect could be noticed.
Li et al proposed an heterostructure composed of PbS QDs crammed between ZnO thin-films [20]. The authors were able to design a photonic memristor making use on the bandgap of the ZnO film (3.37 eV) and of the PbS QDs (1.15 eV). An optical induced set was demonstrated by UV light, whereas photonic reset was performed by IR light. Upon UV light exposure the same explanation in the concept of ionized oxygen vacancies was given by authors. However, for lower light energies such as IR light irradiation, electrons from QDs are excited, neutralizing ionized oxygen vacancies, resulting in the decrease of the current state of the device.
Red light was used to achieve multiple states for MLC characteristics in a memristor composed of Sb 2 Se 3 /CdS NRs, with PPC effect [125]. The RS behavior was attributed to the formation/rupture of CFs. It was clarified that under exposure, part of the irradiated light would be directly absorbed by NRs and other part would be reflected towards adjacent NRs, enhancing energy harvesting. NRs would act as trapping sites for photogenerated holes on the Se 2 Sb 3 surface, which would increase the current state and the lifetime of photogenerated electrons, resulting in PPC effect.
Another interesting approach for MLC operation was proposed by Guo et al. The memristor structure, had an active layer composed of Ag NPs embedded in Mxene-TiO 2 nanosheets. In this study, both set and reset were accomplished by electrical means and an additional state, medium resistance state (MRS) was realized by exposure to UV light, as exemplified in figure 5(b). MRS could be 2 orders of magnitude higher than HRS with fair retention of states (figures 5(c) and (d)) [126]. The physical mechanism of switching was explained by oxygen vacancies CFs facilitated by Ag NPs. A missing peak of atom signals around Ag NPs was found on the TEM analysis, which the authors related to an oxygen vacancy where the conductive path on LRS would originate. UV light could create electron-hole pairs in the TiO 2 , increasing the content of Ti 3+ and, thus, facilitating the creation of oxygen vacancies. Therefore, Ag NPs do not take part of the optical induced reaction in this case. This strategy can be used to decrease D2D and C2C variabilities by controlling the CFs position and minimizing randomness.
On the contrary, UV light was used to perform reset on an Ag NPs@TiO 2 -based OEMs device [127] and, in this work, Ag NPs provided the means for an optical set and reset. An additional absorption peak in the visible region, not observed in the TiO 2 film, was found with the introduction of Ag NPs. Both visible light and UV irradiation caused an immediate current increase when the device was on the HRS. As can be seen in figure 5(e), when visible light was turned off, the device gradually decreased its current and stabilized in a state higher than the initial state (HRS) -indication of PPC effect. During UV irradiation, the current would also increase similarly to the previous described study. However, after UV was turned off, the current decreased to a lower conductance state than the HRS (figure 5(f)). This light induced set and reset was related to an oxidation and reduction of Ag NPs, respectively. As schematically illustrated in figure 5(g), under visible illumination, hot electrons would be excited within the Ag NPs through the localized surface plasmon resonance effect and would hop to the conduction band of TiO 2 , leaving positively charged Ag + . This charge separation would induce an oxidation of Ag NPs and decrease the Schottky barrier at the Ag/TiO 2 interface, increasing the current state. On the other hand, UV light irradiation would promote the creation of electron-holes pairs in TiO 2 . The resulting free electrons would then react with Ag, reducing the Ag + to Ag 0 , increasing the Schottky barrier and decreasing the current state.
One-dimensional structures were also coupled with perovskites materials [128][129][130][131] for enhanced OE features. In most cases, an abrupt behavior on RS was reported and UV light was used to decrease V Set and V Reset .
Undoubtedly, meaningful attempts have been made in understanding how one can use 1D-structures coupled with other materials to manipulate light-induced RS and achieve photonic set and reset operations. A summary of the most relevant works, and its figures of merit, on this topic is provided in table 4. Once again, IR light provoked RS has the smallest percentage on the works reported, as can be seen in figure 5(h).
It is noticeable, however, that the physical mechanism of RS behavior under light on nanostructured materials still requires clarification for further progress. One NR/NW/NRi can be considered as a single device as a means to fully understand the light interaction with the material in these nanostructures. However, manipulating the RS behavior still poses as a challenge since this type of memristors usually suffer from poor retention and endurance related to Joule heating, high C2C variability and high operating voltages [118,132]. As an alternative, a network of the nanostructures can be considered as a single device or as an array of devices. The second approach could offer incredible high-density integration since the nanostructured network could be organized in a crossbar configuration. As a prospective example, core-shell NRs could be employed as both bottom and top contacts and single memristor cells would be represented in each crosspoint. The main challenge here is the ordered positioning of the nanostructures [132]. Additionally, the access to single cells would be hard to accomplish due to such small features [118]. Moreover, high D2D variability has been reported for memristors with the smallest areas [2] which can be an indication of how challenging decreasing variability is in single nanostructures. At this point of the research and while effective solutions do not arise to all the above challenges, NPs embedded in thin-films as active layer appears to be a more straight-forward approach to take advantage of nanostructured materials to enhance photo-harvesting in OEMs.

Remarks
All the works analyzed have provided significant developments in the realization of OEMs using light as an additional input to modulate RS behavior. Both analog and abrupt-RS can be obtained by engineering all sort of material structures. Therefore, it is important to keep in mind what is the application envisioned for the development of an OEM with the required features, as will be discussed in the next section. Metal oxide-based OEMs, in their majority, show abrupt RS with sufficient retention time well-suited for digital processing in-memory architecture. For such application, the PPC decay should be suppressed for large enough time, hence, any charge/ion drift destructing the stored state can arise an issue. As shown in the tables above, especially for 2D-materials based OE RS devices, IR light has been demonstrated to induce reset in some cases. IR induced RS is very appealing for wireless systems to enable conversion and storage of IR data. However, for the realization of completely wire-free systems, both set and reset operations need to be performed by light.
Semiconductor oxides and 1D-structured materials as switching medium in OEMs have already shown desirable photonic behavior. Here, strategies such as double-layered thin films with different oxygen contents to arrange a dynamic equilibrium between ionization and neutralization of defects can be applied. In this case, a high energy wavelength can be used to set the device into a saturated photocurrent and then a low energy wavelength can promote the neutralization of defects performing reset. Moreover, the combination of thin films with nanostructured materials such as QDs or NPs can also be employed for a photonic approach. 2D-materials such as graphene can also be coupled with nanostructures such Au NPs for an enhancement of the photocurrent as already shown for a three-terminal device [139].
The switching speed was expected to be enhanced in optical induced transitions, especially for 2D-based or nanostructured OEMs. However, from the available literature this is not yet noticeable. One of the reasons is the immaturity of the technological concept, which leads to most of the reported devices having common bottom contact and not ready for further investigations. However, the switching speed improvement should be a top priority in future developments.
To tackle high performance, OE devices based on metals of nanometer dimension should be also investigated, which has not been explored yet. As can be seen in [140], the performance in terms of switching speed and retention is superior than bulky systems where slow-driving physical mechanisms are dominant. However, one restriction of the pure metallic device concept relies on miniaturization, where maintaining compositional integrity is difficult at reduced dimensions.
If the PPC effect is suppressed in favor of sufficient retention time, DNN applications can be targeted. On the other hand, PPC decay can be further investigated in the scope of SNNs systems. In this regard, short term and long-term memory and time notion should be realized. Time encoding based on optical stimulus turns the system into an event-driven network which favors power efficiency and allows the computational machine to become a truly scalable SNN hardware with yet unexplored computational capabilities [140].

OEMs beyond the device level
OEMs have tremendous potential for the development of AI hardware. Since these devices can realize computing in different forms depending on their single-cell characteristics, new possible applications are constantly being trialed.
To be integrated in circuits, novel architecture models should be envisioned as a matrix of OEMs, forming as many synapses as possible, processing element inputs, with built-in auxiliary electronics inspired by pixel circuit, forming an artificial optical neuron.
The advantage of dense circuits and high-speed optics can be realized if the program/write operation takes place in the optical domain and followed by an electrical read-out signal. However, for different applications such as DNN or SNN, various electrical components are required to be considered such as photovoltaic devices and photodetectors to convert/revert electrical/optical signals.
For instance, plasmonic OEMs with the possibility of optical read-out have been proposed. For this purpose, conductive-bridge memristors were developed in which metallic CFs are responsible for RS behavior. In such, when the device is at LRS, metallic CFs block the passage of light whereas at the HRS, light can easily reach a detector [141,142]. Other studies have shown an optical read using the photovoltaic effect [143,144].
Light emitting memristors (LEMs) have been developed that can be used for an optical read and write-strategy [145]. Moreover, these light emitting devices can be used as optical communication between layers in ANNs without the need of external light sources. Zhu et al proposed a OE artificial nerve composed of LEM layers in which the optical output of one layer was transmitted to the next layer, while simultaneously realizing dynamic adjustable transmissions [146].
Therefore, light stimulus can be used in different strategies to propel the development of an efficient ANNs hardware. On the neural network itself, light input can modulate the synaptic weight update, which corresponds to the conductance state of the memristor. In this case, OEMs can accomplish a variety of synaptic functions bringing us closer to a brain-inspired AI technology. In the next sections of this paper, a review on the already demonstrated capabilities of inorganic OEMs regarding logic gates, neuromorphic functions and, in more detail, artificial visual systems is presented.  [147], with the permission of AIP Publishing. (d) Two images were inputted through visible light in the form of letters Y and X when the devices in a 5 × 5 matrix were in (e) the HRS for an AND gate. Reprinted with permission from [81]. Copyright (2017) American Chemical Society.

Arithmetic units and Boolean logic
Boolean logic is a form of algebra based on the truth possibilities of true or false that correspond to logic 1 or 0. OEMs can exhibit different non-volatile resistance states at different light wavelengths irradiation, which allows for logic gates and arithmetic functions to be employed. This approach enables in-memory computation which is a significant enhancement from the conventional logic gates based on volatile switches such as transistors, where extra power consumption and time are required due to the physical separation of storage and processing units.
AlO y /CeO 2 OEM in which four different resistance states could be realized using blue and green irradiation with different intensities (4 and 6 pW µm 2 ), enabling information storage of 8-bit codes, was reported by Tan et al [80]. Broadband light irradiation also showed an increase of current state dependent on light intensity with a retention of state demonstrated for 10 4 s. Moreover, a linear relation between resistance state and number of light pulses was used to employ counter and adder arithmetic functions.
Moreover, an OR gate can be accomplished by OEMs [61,70,71], in which a light pulse is used as input A, with 0 corresponding to no light and 1 to irradiation, and an electric pulse as input B, with 0 corresponding to 0 V (or to the read voltage) and 1 corresponding to an input voltage value of X V, in which X > 0 V (or the read voltage) and depends on the device's material structure and its set voltage, as schematically illustrated in figure 6(a)). The output should be 1 when A or/and B is at 1 state. In figures 6(b) and (c), an experimentally demonstration of an OR gate is presented, relating to an OEM of structure FTO/BiVO 4 /TiN [147]. The authors defined the optical input as 1 when the device was irradiated with laser of 405 nm and 11 pW µm −2 and the electric input 1 was not described in the original work. A threshold value of current (10 −6 A) was decided and, therefore, a current output below 10 −6 A is logic 0 and above 10 −6 A is logic 1. PPC effect was not shown, which significantly hinders applications where light-induced state storage is necessary. OEMs that can respond to multiple light wavelengths with some degree of retention are preferred. In this regard, an AND and OR logic gate has been employed in a single cell by OEMs [79,89]. Light and electric pulses are inputs A and B with 0 and 1 possibilities. Contrary to the OR gate, in the AND gate the output should be 1 only when both inputs A and B are 1. Wang et al demonstrated an AND logic gate in which red light was used as input A with a low energy consumption of 4.503 nJ [79]. Red light irradiation by itself could not increase the current state to the defined output logic 1 (I output > 400 nA). Therefore, the output logic value was 1 only when the electric input was 1 (4 V with current compliance (CC) set to 1 µA) and the optical input was 1 (red light irradiation). Additionally, using the same device, blue light irradiation could be used for an OR gate with an energy consumption of 4.526 nJ. Blue light illumination supplied enough energy to increase the current state to above the threshold value defined to the output 1 (I output > 400 nA). Moreover, the electric input (input B) for the OR gate was defined as 1 for a 4 V input with higher CC than the AND gate (1 mA) to ensure that only the combination of both logic inputs 0 would result in the logic 0 output.
In a different approach, a 5 × 5 array was constructed to perform logic optoelectronic AND and OR gates. In this experiment, schematically illustrated in figure 6(d), two images were introduced by visible light (λ = 600 nm for 4 s) in the form of letter Y and X for both AND and OR gates. In this case, both logic inputs are optical (letter Y and X). The AND gate could be realized when all OEMs in the crossbar were in HRS before light irradiation, since only the cells irradiated twice would show an increase of current state to above the threshold value, as can be seen in figure 6(e). The OR gate was in place when all devices in the crossbar were in LRS before irradiation, since the output current of all cells irradiated at least once would increase to above the threshold defined to output logic 1. The results showed an accurate output, proving these crossbar logic OE operations could be applied in larger scales for image processing and storage [81].
NOT operation is also indispensable to the spanning set of a complete logic vector space. Tan et al showed NOT operation with an electric pulse that performed reset [81]. In this work, a complex logic design was proposed that comprised an AND gate when input A (light) and input B (electric set) were applied. This AND gate could be reconfigured to an OR gate by applying an additional light pulse before inputs A and B. To perform NOT, an additional electric pulse for reset was applied after inputs A and B, deleting the information encoded, as illustrated in figure 7(a). The experimental results can be consulted in figure 7(b) and show that this strategy can accurately perform these operations. Additionally, optical adder and digital-to-analog converter functions were realized by employing two light pulses as inputs. These light inputs can be combined into binary digital inputs in 00, 01, 10, 11 forms, in which 0 corresponds to dark and 1 corresponds to light irradiation. The current output will be the analog addition of the equivalent of the inputs.
More complex logic gates were also demonstrated such as IMP using two OEMs in series with a conventional resistor in a more complex strategy. NAND could also be observed using blue light [74]. In a recent work, all possible 16 logic functions were realized in a single cell ZnO OEM [90]. However, except for the OR and NIMP functions, a control light had to employed to experimentally demonstrate the other 14 logic functions which ultimately means two light inputs in different sequences for each specific function.
Although these works provide efficient methods for complicated operations using OEMs, the proposed system strategies are somewhat complex utilizing several inputs as part of the operations procedure. In table 5, a summary of the demonstrated logic gates and arithmetic functions by OEMs can be consulted. It is important to note that Boolean logic is not the only approach to construct integrated circuits that perform mathematical operations. A programmable fuzzy-logic gate approach was proposed as a simpler alternative, in which negative pulses would not be necessary for NOT operation [148], using an OEM that performed material nonimplication and logical true operations using light and electric stimulus.

In-memory and neuromorphic computation
A range of synaptic functions can be emulated by memristors using their resistance levels as synaptic weights. A non-volatile gradual increase in the memristor conductance state is usually called potentiation, which translates into a higher synaptic weight and simulates the connection between two neurons becoming stronger. On the contrary, depression is when the conductance states decrease related to a lower synaptic weight, which happens as the connection of two neurons become weaker. Therefore, memristive crossbars have been proposed as an alternative to realize ANNs, on hardware in a highly energy, cost and fast efficient manner. OEMs, with their additional light stimulus, can further enhance the advantages reported for neuromorphic computing. Since it is physically separated from the memristor's electrodes, light can act as non-invasive write input to train synaptic weights in crossbars [149,150].
DNNs are used in several tasks such as pattern recognition, speech recognition or machine translation [151]. In memristor-based DNNs, crossbars perform vector-matrix-multiplication operations (VMMs) as the result of Ohm's and Kirchhoff 's summation laws [152], as schematically illustrated in figure 8(a)). The input data is applied as a voltage/current vector at the rows and the output can be read at the columns [153]. These crossbars usually suffer from sneak-path current issues, which are related to electric interference between cells, since all the memristors are connected to each other through conducting wires [154]. A transistor in each cell can be added as a selector element [24,155,156], however, it increases energy consumption and cell area. In photonic crossbars, light as an alternative strategy could solve this problem in an energy efficient manner.
A DNN composed of ZnO-based OEMs to perform pattern recognition [89] was proposed by Wang et al. A simulation was carried out using potentiation and depression experimental results using input, hidden and output layers composed by 1024, 256 and 15 neurons, respectively. The potentiation and depression were demonstrated using visible light pulses and electric pulses, correspondingly, with a ratio of 25. Based on the back-propagation algorithm simulation, 15 input images were trained in the DNN and after 1000 epochs, the recognition rate was of 86.7%, demonstrating excellent potential. To our knowledge, this work reports on the highest POT/DEP ratio achieved by an OEM so far. However, it still does not quite reach the requirement of 10 2 ratio to ensure less impact of noise in the measurement [157]. Another ZnO-based OEM suited for DNNs was proposed and simulated for pattern recognition as depicted in figure 8(b). All-photonic potentiation and depression was demonstrated using visible light inputs and a pattern recognition accuracy of more than 90% was achieved (figures 8(c) and (d)) [90]. However, potentiation/depression ratio was only 3.2 which is very low for real implementation.
In fact, very few works related to OEMs report on DNNs applications. Not only because of the high ratio requirement but also due to the unwavering criterion of long retention data [157]. As evidenced by the previous sections of this review, this long light-induced current retention is possible with suppression of PPC decay. However, a controlled PPC decay effect by temporal light stimulus is much more useful for SNNs.
The exploration of the PPC effect to emulate SNN related synaptic functions was first proposed by Lee et al using planar AOS-based memristors [158]. Since then, several works have followed the trend using vertical OEMs. For instance, using different long enough irradiation time (not light pulses <1 s), Wang et al showed the transition from short-term plasticity (STP) to long-term plasticity (LTP) as can be seen in figure 9(a) [125]. The irradiation time affected the PPC dynamic. In the biological system, STP usually last from seconds to dozens of minutes whereas LTP can be determined by hours-weeks or even a life-time of no memory loss [159]. Another way to realize the transition from STP to LTP is by increasing light power, as can be seen in figure 9(b) [160]. A schematic illustration of the multistore memory model is presented in figure 9(c) [161]. As can be perceived, the transition from STP to LTP should be performed by repetition (rehearsal) which can be accomplished by increasing the frequency of light pulses. An example is presented in figure 9(d) [30].
In fact, by continuously irradiating an OEM, the device will suffer a learning process [116,138]. In figure 9(e), a learning and forgetting test is presented from a double layer ZnO/HfO x based memristor [160]. During the first learning/forgetting stage, the synaptic weight increases during a 7 s long UV illumination and then, when the light is turned off, gradually decreases (PPC effect) to an intermediate state. Once the light is turned on again, only 4 s are required for the device to return to the synaptic weigh achieved with the first leaning, which shows faster re-learning. Moreover, the retention ratio also increases from 52% to 74% in the second process, showing harder forgetting. The learning and forgetting synaptic function can also be  demonstrated under light pulses which is more realistic for future energy-efficient applications, instead of a constant illumination method [99]. Another common synaptic function is paired-pulse facilitation/depression (PPF/PPD). The PPF/PPD index is calculated as the relative current state difference between single and double spikes ((I2-I1)/I1) with the same power and width, yet different intervals between double pulses [77,87,101]. An experimental demonstration of PPF can be consulted in figure 9(f), related to a Mo/Sb 2 Se 3 /CdS/ITO OEM and 650 nm light pulses [125]. PPD was also reported regarding an OEM in which potentiation was performed electrically and depression in a combination of electric pulses and constant optical input [109].
A more complex synaptic function is spike-time dependent plasticity (STDP), key performance applied in SNN models. Here, the synaptic weight change (∆W) is modulated by the time difference between pre and post-synaptic spikes. As an example, Song et al designed a pulse scheme strategy for STDP using two OEMs [103]. The pre-synaptic light spike was applied to one of the devices and the post-synaptic light spike was applied to the other one. In another work, the pre-synaptic pulse was applied with light and post-synaptic pulse was electrical [125]. The time interval between the two spikes was defined as ∆t = t pre − t post and the ∆W= (∆W 2 -∆W 1 )/∆W 1 . In figure 9(g), the experimental results can be seen. ∆W decreases with longer ∆t, as expected, which is typical STDP characteristics of biological synapses [21].
Furthermore, brain-inspired synaptic functions were realized in all-photonic memristors [20,26]. STDP was emulated in an IGZO device using blue light for potentiation and IR light for depression [26]. The transition from STP to LTP and PPF/PPD functions were demonstrated using UV and IR light on a ITO/ZnO/PbS(QDs)/ZnO/Al photonic memristor [20].
The few studies reporting on all-photonic realized synaptic functions uncover the infinite potential of OEM for brain-inspired neuromorphic computation. In table 6, a summary on the synaptic functions of inorganic OEMs is presented. The scaling possibilities of RS devices, together with an all optical controlled strategy for the implementation of different weights in ANNs, suggest that OEMs crossbar implementation could offer an abysmal efficiency improvement over silicon-based technologies [162].

Artificial visual systems
In the human visual system, eyes collect information through light signals and convert it to electrical pulses, which are then detected by visual neurons and analyzed and memorized by the visual cortex in the brain. Several kinds of artificial visual systems have been proposed that try to emulate this behavior [163,164]. Usually, in the conventional image processing based on deep learning-based computer visions, an array of  photodetectors and image sensors are employed, to act as the retina, that gathers light information and converts it to electric signals, transferred to software-based ANN for further processing and storage [165][166][167][168]. To improve computer visions, novel algorithms are required to be investigated, especially in the context of human vision-inspired models and SNN approach to synthesize the most realistic image. Furthermore, better sensor performance for long distance and low-light resolution is another research priority in this regard. OEMs can directly respond to light stimulus and perform light induced synaptic functions, which can simplify the peripheral circuit while offering a more energy and area efficient alternative [169,170]. Two layers ANNs can easily be realized using crossbars of OEMs, composed of input and output layers, and has already been simulated several times with high pattern recognition accuracies [102,116,160,171], proving the outstanding potential of this approach. However, using photocurrent and light intensity as synaptic weigh modulator in a more complex ANN, with several hidden layers, is a hard to surpass challenge [66]. The light irradiation that inputs on the first layer of the neural network cannot be cascaded into the following layers, unless the already reported strategy on LEM is in place. In such, the first layer OEMs emit light to the second layer, as illustrated in figure 10(a) [146]. Even then, VMM operations would be extremely hard to be accomplished with high accuracy, since photonic I LRS /I HRS ratio is usually low (<10). More studies are demanded to increase the memory window.
A combination of OEMs with ANNs electric-field assisted memristors crossbar looks closer to the reality of today's technology development and can be considered as an effective way to construct an artificial visual system, as illustrated in figure 10(b)). An OE memristive crossbar can be used for a primary in-situ processing and image sensing, acting as a neuron that then transfers this information to an electronic-based memristive crossbar, where the processing and storage of this data would take place. In this way, high accuracy recognition tasks can be performed and the overall performance of artificial visual systems would be greatly enhanced [66,78]. Moreover, using PPC decay on the photonic crossbar can be an efficient way to, on one hand, store the input image for some time (sufficient for inference), and, on the other hand, maintain low power operation due to auto-resetting of the system following relaxation time. Synaptic functions could also be applied and the whole system could undergo learning processes, in which the more times an image is inputted, the harder it gets for the photonic crossbar to 'forget it' .

Conclusion
In this review, the state of the art in inorganic OEMs is described and analyzed. Since the development of these devices should be carried out with a specific application in mind, due to different performance requirements, both memristor materials and structures with its figures of merit and different applications are covered.
On the materials point of view, metal oxides and 2D-materials based OEMs mostly show abrupt RS, suitable for in-memory computation in logic gates or DNNs. IR irradiation has been used to induce reset, which is very appealing for wireless systems, to enable conversion and storage of IR data. Semiconducting-oxides and 1D-structured materials employed as active layer enabled the realization of fully-photonic devices, in which both set and reset were induced by different light wavelengths. However, in the future, it would be important that further studies on the physical mechanism of light-induced RS, and how to control it, be developed.
In regard to possible applications, the PPC effect induced by light can be used to realize all sort of neuromorphic computing requirements. The logic gates already accomplished using inorganic OEMs are discussed in detail. Moreover, in some cases, the PPC decay was shown to be dependent on irradiation time, intensity, frequency and wavelength which was explored to emulate several synaptic functions, also described in this review. In particular, these synaptic functions can be applied for the construction of a more energy, cost and area efficient artificial visual system. Several approaches have been proposed and are analyzed here.
Since the photonic ratio has demonstrated to be low (<10) and due to a lack of an efficient way to optically transfer photo-inputs from one to the next layer in an ANN, a combination of photonic and electric crossbars is suggested to efficiently solve these issues.

Data availability statement
No new data were created or analysed in this study.