All-Optically Controlled Memristor

Memristors have emerged as key candidates for beyond-von-Neumann neuromorphic or in-memory computing owing to the feasibility of their ultrahigh-density three-dimensional integration and their ultralow energy consumption. A memristor is generally a two-terminal electronic element with conductance that varies nonlinearly with external electric stimuli and can be remembered when the electric power is turned off. As an alternative, light can be used to tune the memconductance and endow a memristor with a combination of the advantages of both photonics and electronics. Both increases and decreases in optically induced memconductance have been realized in different memristors; however, the reversible tuning of memconductance with light in the same device remains a considerable challenge that severely restricts the development of optoelectronic memristors. Here we describe an all-optically controlled (AOC) analog memristor with memconductance that is reversibly tunable over a continuous range by varying only the wavelength of the controlling light. Our memristor is based on the relatively mature semiconductor material InGaZnO (IGZO) and a memconductance tuning mechanism of light-induced electron trapping and detrapping. We demonstrate that spike-timing-dependent plasticity (STDP) learning can be realized in our device, indicating its potential applications in AOC spiking neural networks (SNNs) for highly efficient optoelectronic neuromorphic computing.


Introduction
In the past few years, artificial intelligence (AI) has been getting more and more practical applications to human society, which based on electron trapping [34,49] or proton intercalation, [35] a light-induced persistent photocurrent [34,49] or photocatalytic [35] effect was used to increase the memconductance; however, a decrease in memconductance could be achieved only via electric stimuli [34,35] or with strong dependence on an external electric field. [50] In the case of memristors based on conducting nanofilaments (belonging to the class of ion-migration-driven memristors), light illumination can cause filament breakage, resulting in decreased memconductance; [51,52] however, to increase the memconductance via filament formation or rejuvenation, electric excitation is still necessary. [51,52] In addition, a photomechanical switching effect can be used to modulate memconductance, [36] although this causes severe expansion or contraction of the device. To the best of our knowledge, the reversible tuning of memconductance by applying only optical excitation in the same device has not yet been realized.
In this work, we fill this gap by proposing a bilayered oxide memristor based on a memconductance tuning mechanism of electron trapping and detrapping. The memconductance of our device can be reversibly tuned by varying only the wavelength of the controlling light. The light-induced memconductance states are nonvolatile. Our all-optically controlled (AOC) memristor can serve as an excellent synaptic emulator, as we demonstrate by mimicking spike-timing-dependent plasticity (STDP). The realization of AOC neuromorphic devices makes possible real optoelectronic NC that will significantly promote the development of the new generation of AI.

Memristive Switching Behavior
In our device, a wide bandgap amorphous oxide material, InGaZnO (IGZO), was used as the active layer. IGZO is widely used as a key thin-film transistor material for high-performance display production. The device is based on an oxygen-deficient IGZO (O D -IGZO)/oxygen-rich IGZO (O R -IGZO) homojunction (see the inset of Figure 1a, Experimental Section, and Figure S1 in the Supporting Information). Such a bilayered structure is crucial for achieving an AOC memristor, as will be discussed in detail later. This device demonstrates typical memristive behavior [53] when measured in the dark (Figure 1a). Positive voltage sweeping converts the device from a low memconductance state (LMS) to a high memconductance state (HMS), referred to as the SET operation; the reverse process of memconductance switching from the HMS to the LMS under negative voltage sweeping is called the RESET operation. The LMS and HMS exhibit nonvolatility (see Figure S2a in the Supporting Information). The memconductance can also be reversibly modulated by alternately applying positive and negative voltage pulses (see Figure S2b in the Supporting Information).
To determine the mechanism of memristive switching, we analyzed the band structures of O D -IGZO and O R -IGZO (for details, see the Experimental Section and Figure S3 in the Supporting Information). The analysis indicates that the electrons in O D -IGZO tend to diffuse into O R -IGZO, resulting in the formation of a built-in electric field at the O D -IGZO/O R -IGZO interface and, thus, a potential barrier on the O D -IGZO side and a potential well on the O R -IGZO side. The width of this interfacial barrier, which depends on the density of ionized oxygen vacancies (V O 2+ s) and determines the tunneling current, [54,55] plays a key role in memristive switching (for details, see the Experimental Section and Figure S4 in the Supporting Information). Specifically, the SET behavior originates from electron detrapping at neutral oxygen vacancies (V O s) located in the interfacial barrier region; that is, ionization of the V O s causes an increase in the density of V O 2+ s, thus leading to a decrease in the width of the barrier, which facilitates electron tunneling across the junction (see Figure S4e in the Supporting Information). In contrast, for the RESET behavior, electron trapping at V O 2+ s, i.e., the neutralization of V O 2+ s, gives rise to an increase in the width of the barrier, thus resulting in a lowered tunneling current (see Figure S4f in the Supporting Information).

Optical SET Behavior
Having demonstrated memristive behavior under electrical stimulation, we investigated the performance under light exposure. The Au/O D -IGZO/O R -IGZO structure shows a transmittance of >55% for light wavelengths from 400 to 1000 nm (see the inset of Figure S5a in the Supporting Information). Therefore, the device can be irradiated with visible (e.g., 420-650 nm) and nearinfrared (e.g., 800-1000 nm) light for memconductance modulation. When the device was measured soon after blue light (420 nm) exposure, we observed memristive behavior similar to that observed in the dark, but with a significant increase in current ( Figure 1a). To better understand the influence of irradiation on the memconductance, the effects of irradiating the device using light of various wavelengths (420, 530, 650, 800, 900, and 1000 nm) were investigated (Figure 1b). The current gradually increases under these sub-bandgap illumination conditions. After irradiation, the device exhibits a strong persistent photocurrent. This shows that both visible and near-infrared light can convert the device to a nonvolatile HMS. The light-induced HMS can be reset to the LMS by means of electric stimuli. Thus, the memconductance can be reversibly modulated through a combination of visible (or near-infrared) light and negative voltage pulses (see Figure S5 in the Supporting Information).
Single-layered O D -IGZO or O R -IGZO devices exhibit markedly different behavior, which provides insight into the mechanism of optical memconductance. No photocurrent can be observed upon visible light irradiation of the O D -IGZO device (see Figure S6a in the Supporting Information). For the O R -IGZO device, a volatile photocurrent is generated upon visible light illumination, but no photocurrent appears upon nearinfrared light irradiation (see Figure S6b in the Supporting Information). We therefore deduce that for the bilayered O D -IGZO/O R -IGZO device, the interfacial barrier region plays a key role in the optical SET operation (Figure 1c Figure S7 in the Supporting Information. In the figure, we can see a broad deep emission band ranging from 500 to >1000 nm centered at 780 nm for O D -IGZO. It is reported that a high oxygen vacancy density can broaden the energy level distribution of the oxygen vacancies. [56] When IGZO is deposited in the low oxygen partial pressure environment, the optoelectronic response wavelength can be extended beyond 750 nm. [57] Given that our O D -IGZO was deposited in pure argon, we can reasonably deduce that the density of oxygen vacancies in the O D -IGZO is higher than that in the IGZO deposited in the low oxygen partial pressure environment. Therefore, the O D -IGZO has a wider energy level distribution of the oxygen vacancies, thus resulting in a longer optoelectronic response wavelength of 1000 nm. The unique long-wavelength response capability of the O D -IGZO/O R -IGZO device enables the realization of an AOC memristor.

Optical RESET Behavior
The key to achieving an AOC memristor is to realize an optical RESET operation. We found that for the blue-light-irradiated device, subsequent irradiation with green (530 nm) and red (650 nm) light causes an increase in the photocurrent after an initial decrease, whereas near-infrared light irradiation (800 and 900 nm) induces a stronger decrease in the current compared to the case without light irradiation (Figure 2a). It indicates that near-infrared light may be used to reset the device irradiated with blue light. To confirm this hypothesis, the device was first exposed to blue light (420 nm) for 15 s to set it to an HMS. Then, it was irradiated with light of various wavelengths from 530 to 900 nm at 10 min after the initial blue light exposure ( Figure 2b). Green light (530 nm) converted the device to even higher HMSs, whereas red light (650 nm) resulted in a weak   in the photocurrent after an initial increase. Upon nearinfrared light irradiation (800 and 900 nm), a strong decrease in the current, i.e., RESET behavior, was observed following a relatively weak increase; near-infrared light at 800 nm was more efficient than 900 nm light for reducing the device current. Therefore, we selected 800 nm for subsequent RESET operations in pulse mode ( Figure 2c). As demonstrated in Figure 2c, the RESET efficiency increased with the light power density.
We also found that the wavelength range for the RESET operation strongly depends on the light used for the initial SET operation. After ultraviolet light illumination (350 nm), the device can be reset to an LMS by both red and near-infrared light (see Figure S8a in the Supporting Information). However, a device irradiated with green or red light cannot be restored to the LMS even by near-infrared light (see Figure S8b,c in the Supporting Information). Moreover, the ultraviolet-lightilluminated device (see Figure S8a in the Supporting Information) exhibits a current reduction of a larger magnitude upon near-infrared light irradiation than the blue-light-exposed device in Figure 2b. These findings suggest that setting the device with short-wavelength light facilitates the subsequent RESET operation. To further confirm this conclusion, the initial SET was performed with 350, 380, 420, and 450 nm light, followed by a RESET operation with 800 nm light (see Figure S9 in the Supporting Information). Figure 2d illustrates the RESET indexes for different SET operations, which quantitatively represent the RESET efficiency. Obviously, RESET index decreases with increasing wavelength of the light for the SET operation. Then, we can safely conclude that the shorter the light wavelength during the SET operation, the better the subsequent RESET operation. The RESET rate, which represents the instantaneous rate of memconductance decrease, was found to decrease with irradiation time (Figure 2e).
Because the optical SET mechanism depends on barrier narrowing (Figure 1c), we expected that widening of the barrier region due to a reduced V O 2+ density would result in an optical RESET. Upon near-infrared light irradiation, electrons in the potential well formed by band bending of O R -IGZO are apt to tunnel through [58] or jump over [59] the barrier and enter the O D -IGZO conduction band (Figure 2f). Some electrons are captured by V O 2+ s, which then transform into V O s. This explanation for the RESET mechanism can be supported by the strong dependence of the RESET efficiency on the light power density in Figure 2c; that is, light with a higher power density excites more electrons into the conduction band of O D -IGZO, leading to a larger probability of V O 2+ neutralization and thus widening the interfacial barrier.
We further propose that upon illumination, these two opposite reactions, namely, the ionization of V O s and the neutralization of V O 2+ s, occur simultaneously. The photocurrent depends on the dynamic equilibrium between these two reactions. In the ionization-dominated (or neutralization-dominated) case, the device shows an increase (or decrease) in photocurrent. For a device in the LMS, the ionization of V O s is dominant, thus resulting in an increase in memconductance upon light irradiation (Figure 1b). However, when the device is initially exposed to light of a relatively short wavelength, only a small number of V O s can be ionized by subsequent irradiation at a longer wavelength owing to the low density of V O s with relatively shallow energy levels. In this case, the neutralization of , leading to a higher RESET efficiency. In the latter case, the V O density in O D -IGZO increases with irradiation time, resulting in an increasing probability of V O ionization and thus decreasing the RESET rate.

AOC Memristor
We implemented our bilayered O D -IGZO/O R -IGZO memristor using blue and near-infrared light pulses for the SET and RESET processes, respectively, as schematically illustrated in Figure 4a. The top panel in Figure 4b shows a continuous increase and decrease in the memconductance under a series of successive light pulses. Our AOC memristor exhibits good operation endurance (Figure 4b, bottom panel). To verify the nonvolatility of the light-induced memconductance states, we present retention measurements of ten states obtained after both the SET and RESET operations (Figure 4c). The memconductance exhibits an initial slow decay, which is an intrinsic feature of optoelectronic devices, [23][24][25][26][27][28][29][30][31][32][33][34][35]49,55] and then remains stable, with all states being clearly distinguishable even after 10 4 s. The retention measurement results indicate that the lightinduced memconductance states are nonvolatile regardless of the operation mode (SET or RESET).
As mentioned above, the persistent photocurrent decay is related to the tunneling of electrons through the interfacial barrier, which is why the decay was weaker for lower memconductance. More specifically, a lower memconductance implies a wider barrier, resulting in fewer electrons tunneling through it. Subsequently, fewer V O 2+ s transform into V O s, thus leading to less barrier widening.
Our AOC memristor is structure-changeable.  Figure S11 in the Supporting Information). Given that O D -IGZO is highly conductive and thus Ohmic contact forms in the top electrode/ O D -IGZO interface (see Figure S4a in the Supporting Information), similar electrical and optoelectronic behaviors could Top panel: The device was initially set to a relatively low memconductance of about 100 nS, and then ten SET operations were performed with different numbers of 420 nm light pulses. After each SET operation and subsequent retention measurement, the device was set to 100 nS again. Bottom panel: The device was initially set to a relatively high memconductance of about 400 nS, and then ten RESET operations were conducted with different numbers of 800 nm light pulses. After each RESET operation and subsequent retention measurement, the device was set to 400 nS again. In (b) and (c), the memconductance values were measured at 10 mV. We have also tried employing other oxides, such as ZnO, In 2 O 3 and SnO 2 , as the active layer of memristive devices with the same structure as that of IGZO, i.e., Au/oxygen-deficient oxide (O D -oxide)/oxygen-rich oxide (O R -oxide)/Pt. However, no optical RESET behavior could be realized in these memristors. It may be due to 1) a weak persistent photocurrent and 2) a poor long-wavelength response capability, for example, no optoelectronic response was observed upon the light irradiation with wavelength >700 nm. We will try other ternary and quaternary oxides or modifying the deposition parameters of binary oxides. We want to check if optical RESET behavior exists in memristive devices with other materials (other than IGZO) as the active layer.

AOC Memristive Synapse
The multiple reversibly tunable memconductance states of our AOC memristor (Figure 4b) make it an excellent synaptic emulator. Specifically, we found that our memristor can mimic STDP-an important learning rule in the brain that requires variation in the synaptic weight (W) to be a strong function of the pre-and postneuron spike timing (∆t) [60] (Figure 5a). A single blue light pulse and a train of ten near-infrared light pulses serve as pre-and postsynaptic spikes, respectively (Figure 5a, inset). The memconductance is denoted as W. We see that synaptic potentiation (∆W > 0) occurs when the presynaptic spike arrives before the postsynaptic spike (∆t > 0); by contrast, synaptic depression (∆W < 0) occurs when the postsynaptic spike arrives first (∆t < 0). In addition, |∆W| increases with decreasing |∆t|. These observations are consistent with the typical STDP characteristics of biological synapses. [60] Given that STDP is the basic learning rule of spiking neural networks (SNNs) that reflect the information processing style in the human brain, [61] our AOC memristor is promising for applications in optoelectronic SNNs.
The mechanism of STDP emulation is explained in detail in Figure 5b,c. For the emulation of synaptic potentiation, the device was initially set to a relatively low memconductance of M C0 ≈ 100 nS. Figure 5b presents two W modulation events with ∆t 2 > ∆t 1 > 0. First, the blue light pulse induces a positive persistent photoconductance, that is, a light-induced memconductance that is higher than M C0 . The subsequent near-infrared light pulses induce a negative persistent photoconductance, that is, a light-induced memconductance that is lower than that before near-infrared irradiation. We can see from Figure 5b that the updated memconductance (M C1 or M C2 ) is higher than M C0 , indicating the realization of synaptic potentiation. Given that ∆t 1 < ∆t 2 , the M C values before and after near-infrared light irradiation for ∆t 1 are higher than those for ∆t 2 . It follows that M C1 > M C2 , and thus, ∆W 1 > ∆W 2 . To mimic synaptic depression, the device was initially set to a relatively high memconductance of M C0 ≈ 250 nS. Figure 5c shows two W modulation events with ∆t 2 < ∆t 1 < 0. First, the near-infrared light pulses induce a negative persistent photoconductance. The subsequent blue light pulse induces a positive persistent photoconductance. The updated memconductance (M C1 or M C2 ) is lower than M C0 , indicating the realization of synaptic depression. Given that |∆t 1 | < |∆t 2 |, the memconductance before blue light irradiation for ∆t 1 is higher than that for ∆t 2 . However, because of the relatively large decay amplitude of the memconductance after blue light irradiation, the subsequent memconductance for ∆t 1 is lower than that for ∆t 2 . Therefore, M C1 < M C2 and |∆W 1 | > |∆W 2 |.
It deserves mention that choosing an appropriate number of near-infrared light pulses is crucial for mimicking STDP. It is found that ten is the optimal choice for achieving both synaptic potentiation and depression. When too few/too many near-infrared light pulses are used, only synaptic potentiation/ depression can be obtained regardless of ∆t. Figure S12 in the Supporting Information illustrates the effect of the number of near-infrared light pulses on the STDP emulation. It can be seen that the application of five near-infrared light pulses results in synaptic potentiation even when ∆t < 0, whereas the application of twenty near-infrared light pulses leads to synaptic depression even when ∆t > 0.

Conclusions
We believe that the realization of AOC memristors is an important milestone toward real optoelectronic NC that possesses the advantages of both photonics and electronics. For traditional electrical memristors, a high voltage or current is generally required to tune their memconductance. Such strong electrical stimuli may result in 1) high power consumption, 2) a large amount of Joule heat, 3) microstructural change accelerated by the Joule heat, and 4) high crosstalk in memristor crossbars. These issues can be addressed with our AOC memristor given the very low light power densities (≈20 µW cm −2 ) required to operate it. On the other hand, optoelectronic computing using our AOC memristor is more practically feasible than purely optical computing [18][19][20] owing to the simple structure and easy fabrication of this device. Future research might explore neuronal functions (e.g., integrate and fire) using our AOC memristor to enable optoelectronic SNNs.

Experimental Section
Material Growth: Amorphous IGZO thin films with a diameter of 100 µm were deposited on Pt/Ti/SiO 2 /Si, SiO 2 /Si, and quartz substrates at room temperature (RT) via RF magnetron sputtering of an InGaZnO 4 (In 2 O 3 :Ga 2 O 3 :ZnO = 1:1:2, molar ratio) ceramic target of 99.99% purity with in situ metal shadow masks. The sputtering power and pressure were 60 W and 0. Electrical and Optoelectronic Characterization: Electrical and optoelectronic measurements were performed at RT in air using a Keithley 4200 semiconductor parameter analyzer equipped with a monochromatic light source (Omni-λ 3007). The bias voltage was applied to the top electrode (Au or ITO) with the bottom electrode (Pt) grounded, and the light entered into the device through the top electrode (Au or ITO) (Figure 1a).
It is found that the as-fabricated bilayered O D -IGZO/O R -IGZO device was not in an LMS but in an HMS with a memconductance of ≈10 2 -10 3 nS (at −10 mV). Given that the device was highly sensitive to visible light and that exposure to environmental light was inevitable, such an HMS should be induced by environmental light exposure. To restore the device to the initial LMS before electrical or optoelectronic measurements, an electrical or optical RESET operation was first performed. After the RESET operation, the device presented a low memconductance of approximately 10 −1 -10 2 nS (at −10 mV) depending on the RESET parameters used.
Material Characterization: The film thickness was measured via variable angle spectroscopic ellipsometry (M-2000 DI, J. A. Woollam Co., Inc.). The carrier density and resistivity were determined with a Hall effect measurement system (HP-5500C, Nanometrics) using the van der Pauw method. Absorption and transmittance spectra were collected using both a UV-vis-IR spectrophotometer (Lambda 950, PerkinElmer) and spectroscopic ellipsometry. Photoluminescence spectra were measured with a confocal microscopic Raman spectrometer (Renishaw inVia Reflex, 325 nm). All of the above measurements were performed at RT in air. Cross-sectional microstructural observations were performed by means of high-resolution transmission electron microscopy (HRTEM, Talos F200X, Thermo Fisher). Cross-sectional specimens were fabricated by means of focused ion beam etching. Elemental and energyband structure analyses were conducted using X-ray photoelectron spectroscopy (XPS, Kratos Axis Ultra DLD). The work function was determined via ultraviolet photoelectron spectroscopy (UPS). O D -IGZO and O R -IGZO films with a thickness of 100 nm were used for the XPS and UPS measurements. To eliminate the influence of surface contamination (e.g., moisture from the atmosphere) on the measurements, an ≈30 nm thick surface layer was etched with Ar plasma before the optoelectronic signals were collected.
The As determined by Hall effect measurements, O D -IGZO showed an electron concentration of 10 19 cm −3 and a resistivity of 10 −2 Ω cm. O R -IGZO had a much higher resistivity, which exceeded the measurement limit (≈10 5 Ω cm) of the Hall measurement system. Given that oxygen vacancies are the main n-type dopants in IGZO, [62] the low resistivity of the O D -IGZO indicates a high density of oxygen vacancies. The lower conductivity of O R -IGZO indicates that it contains much fewer oxygen vacancies than O D -IGZO. The oxygen vacancies in IGZO have a wide distribution of energy levels in the bandgap. [63][64][65] Band Structure Analysis: The optical bandgaps (E g ) of O D -IGZO and O R -IGZO were determined to be 3.7 and 3.6 eV, respectively, via a modified Kubelka-Munk function ( Figure S3a in the Supporting Information). The E g values were obtained from the abscissa intercepts of the straight lines fitted to the linear portions of the plotted data points in the high energy region.
The differences between the Fermi energy (E F ) and the valence band maximum (E V ) for O D -IGZO and O R -IGZO were determined to be 3.1 and 2.5 eV, respectively, by measuring the valence band XPS spectra ( Figure S3b  and hν is the photon energy (21.22 eV). The differences between E F and E V for O D -IGZO and O R -IGZO were determined to be 3.2 and 2.6 eV, respectively, which are consistent with the values obtained from the XPS spectra ( Figure S3b in the Supporting Information).
Then, the energy band diagrams of O D -IGZO and O R -IGZO can be plotted before contact, as shown in Figure S3d (Supporting Information). The differences between the conduction band minimum (E C ) and E F for O D -IGZO and O R -IGZO were calculated to be 0.6 and 1.1 eV, respectively, according to E g − (E F − E V ). Subsequently, the differences between the vacuum energy (E VAC ) and E C for O D -IGZO and O R -IGZO were determined to be 3.6 and 3.8 eV, respectively, according After  Figure S4a,b in the Supporting Information), it is deduced that the O D -IGZO/O R -IGZO interfacial region plays a key role in memristive switching. Moreover, the strong dependence of the memconductance on the device size (i.e., the diameter of the top electrode or IGZO thin film) for both HMSs and LMSs indicates radially homogeneous switching behavior [66] (Figure S4c  More specifically, it is proposed that the memconductance of the device mainly depends on the width of the O D -IGZO/O R -IGZO interfacial barrier, which determines the tunneling current. [54,55] The barrier width is determined by the density of positive ion space charges (i.e., ionized oxygen vacancies): a higher density of ionized oxygen vacancies results in a narrower barrier width. [54] As mentioned above, there are abundant V O s with a wide distribution of energy levels in O D -IGZO. The V O s in relatively shallow energy states tend to be doubly ionized, [67,68] transforming into V O 2+ s with shallow energy levels ( Figure S4d in the Supporting Information). When a positive bias is applied to the device, some of the V O s in relatively deep energy states ionize into V O 2+ s [69] ( Figure S4e in the Supporting Information). These additional V O 2+ s generated in the barrier region cause the barrier to narrow, which facilitates electron tunneling across the junction. Thus, the device is set to an HMS. The HMS is nonvolatile because free electrons produced in the barrier region are swiftly pulled into O D -IGZO by either the built-in or the external electric field and therefore cannot recombine with V O 2+ s. The existence of an energy barrier hindering V O 2+ neutralization also contributes to the nonvolatility. [67] When a negative voltage is applied, the interfacial barrier is forward biased, and many electrons are injected into the barrier region. Some electrons will be captured by V O 2+ s generated during the electrical SET operation, which then transform back into V O s ( Figure S4f in the Supporting Information). As a result, the barrier becomes wide again and the device is reset to the LMS.
The current decay of the HMS in Figure S2a (Supporting Information) can therefore be explained by the spontaneous neutralization of ionized oxygen vacancies: some electrons tunneling through the interfacial barrier from O R -IGZO to O D -IGZO are captured by ionized oxygen vacancies, which then transform into neutral oxygen vacancies; a decrease in the number of ionized oxygen vacancies results in widening of the interfacial barrier and, thus, a reduction in current.

Supporting Information
Supporting Information is available from the Wiley Online Library or from the author.