Optically driven intelligent computing with ZnO memristor

Artificial vision is crucial for most artificial intelligence applications. Conventional artificial visual systems have been facing challenges in terms of real-time information processing due to the physical separation of sensors, memories, and processors, which results in the production of a large amount of redundant data as well as the data conversion and transfer between these three components consuming most of the time and energy. Emergent optoelectronic memristors with the ability to realize integrated sensing-computing-memory (ISCM) are key candidates for solving such challenges and therefore attract increasing attention. At present, the memristive ISCM devices can only perform primary-level computing with external light signals due to the fact that only monotonic increase of memconductance upon light irradiation is achieved in most of these devices. Here, we propose an all-optically controlled memristive ISCM device based on a simple structure of Au/ZnO/Pt with the ZnO thin film sputtered at pure Ar atmosphere. This device can perform advanced computing tasks such as nonvolatile neuromorphic computing and complete Boolean logic functions only by light irradiation, owing to its ability to reversibly tune the memconductance with light. Moreover, the device shows excellent operation stability ascribed to a purely electronic memconductance tuning mechanism. Hence, this study is an important step towards the next generation of artificial visual systems.


SUMMARY
Artificial intelligence is widely used in everyday life. However, an insufficient computing efficiency due to the so-called von Neumann bottleneck cannot satisfy the demand for realtime processing of rapidly growing data. Memristive in-memory computing is a promising candidate for highly efficient data processing. However, performance of memristors varies significantly because of microstructure change induced by electric-driven matter migration.
Here, we propose an all-optically controlled (AOC) memristor with a simple Au/ZnO/Pt sandwich structure based on a purely electronic tuning mechanism of memconductance. The memconductance can be reversibly tuned only by light irradiation with different wavelengths.
The device can be used to perform in-memory computation such as nonvolatile neuromorphic computing and Boolean logic functions. Moreover, no microstructure change is involved during the operation of our AOC memristor which demonstrates superior operation stability. Based on this and its structural simplicity, the device has attractive application prospects for the next generation of computing systems.

INTRODUCTION
Due to increased use of artificial intelligence (AI) in applications such as face recognition, self-driving vehicles, internet search, Go game and internet of things, increasing demands (especially high speed and low energy) are being placed on the supporting computation hardware. Based on the von Neumann architecture, the central processing unit (CPU) (for data processing) and the main memory (for data storage) of modern computers are physically separated and the data transfer between them consumes most of the time and energy. This is known as the von Neumann bottleneck. [1,2] The reason for this bottleneck is that CPUs are advancing at a much faster pace than memories, causing the computer performance to be increasingly limited by the speed of memoriesa phenomenon known as the memory wall. [3] Unlike the von Neumann architecture, in-memory computing can perform calculations in situ, i.e., computations can be carried out in the memory. [4][5][6] Hence, the latter approach seems promising to eliminate the von Neumann bottleneck or the memory wall. The concept of in-memory computing was pioneered by Kautz in the late 1960s. [7] He proposed cellular logic-in-memory arrays where a computational memory cell is composed of several electronic elements such as transistors, diodes and resistors. Since the late 2000s, due to the slowing of Moore's law, wide applications of AI, and rapid development of nonvolatile memories, in-memory computing, which can be performed in a single electronic element, has been receiving increasing attention. These nonvolatile memories include floating gate transistors, [8][9][10] memristors, [11][12][13][14][15][16][17][18][19][20][21][22][23][24][25][26][27] phase change memories, [28,29] magnetoresistive memories, [30,31] and ferroelectric memories. [32,33] As illustrated in Figure 1, in-memory computing can generally be classified into three categories: nonvolatile neuromorphic computing, [9,[13][14][15][16][17]30,31,33] logic-in-memory, [10,[34][35][36][37][38][39] and arithmetic-in-memory. [28,40] Note that nonvolatile neuromorphic computing includes matrix-vector multiplication (according to universal circuit laws such as Ohm's law and Kirchhoff's law) [9,[41][42][43][44][45] and spike-timingdependent plasticity (STDP). [9,[13][14][15][16]30,31,46] Among nonvolatile memories, memristors are considered as one of the ideal candidates for in-memory computing due to their simple two-terminal structure, high operation speed, and low energy consumption. [4][5][6] Generally, memristors are controlled (in full or in part) by electrical stimuli; that is, memconductance is tuned via electric-field-driven or current-driven ion (or atom) migration within the active layer or at the electrode/active layer interface, [47,48] resulting in microstructure change. However, the microstructure change induces significant

In-memory computing
Arithmetic-in-memory STDP

PPF & PPD
Integrate-and-fire cycle-to-cycle and device-to-device performance variations, which are a major obstacle to their practical application. [49,50] In this work, we propose an all-optically controlled (AOC) memristor which can be operated at low light power densities. This device consists of a simple Au/ZnO/Pt structure. It demonstrates nonvolatile memconductance switching due to light-induced electron trapping and detrapping at intrinsic defects in the oxide, thereby no microstructure change occurs.
This device shows excellent operation stability and is used to realize in-memory computing such as nonvolatile neuromorphic computing and logic-in-memory. Therefore, this AOC memristor has a great application potential for designing highly efficient computing architectures.

Memristive Switching Behavior
We used polycrystalline ZnO with a wide band gap of 3.2 eV as the active layer (see Figure S1). ZnO fabricated by various methods has been widely used as the active layer of memristive devices. [51][52][53][54] Herein, the ZnO thin film was deposited at room temperature via RF magnetron sputtering in pure Ar gas. The device demonstrated a nonpolar memristive switching behavior, i.e., from a high memconductance state (HMS) to a low memconductance state (LMS) when measured in the dark ( Figure S2a). To investigate the switching mechanism, the metal electrodes were replaced by a transparent conducting oxide (Sn-doped  [59] that is, a higher/lower density of V O 2+ s led to a narrower/wider width. In an equilibrium state without bias voltage, the numbers of electrons tunneling through the junction along both directions were equal and the net current was zero ( Figure S2c). When applying a positive bias, there was a flow of electrons into ZnO ( Figure S2d) ZnO decreased, which also contributed to the lowered memconductance.

Optical SET and RESET Behaviors
The Au/ZnO structure has a mean transmittance of > 60% for light wavelengths from 350 to 1000 nm ( Figure S4). This suggests that its memconductance could be modulated via light irradiation. Light was injected into the device through the top electrode (Au). Upon illumination with short-wavelength lights (350 and 420 nm), the device current instantaneously increased and reached to a saturation value (Figure 2a, top and middle panels). After illumination, the device showed a persistent photocurrent (PPC) phenomenon (see the black curve). Green light (530 nm) irradiation caused a gradual increase in the current followed by a PPC (Figure 2a, bottom panel, black curve). PPC is known to be an intrinsic phenomenon in most optoelectronic devices, which can be understood as follows: irradiation with light of appropriate wavelengths results in an increase in the current; after irradiation, although a current decay occurs, the device cannot be restored to its initial conductance state before irradiation, i.e., the conductance state after irradiation is nonvolatile. Therefore, relatively short-wavelength light (350, 420, and 530 nm) could switch the device from LMS to HMS, referred to as the SET operation. On the other hand, when the device was irradiated with long-wavelength light (650, 725, and 800 nm), very weak or even no current change was observed ( Figure S5).  For the RESET operation, the device was first set to an HMS with 530 nm light, followed by irradiating it with light of long wavelengths 30 minutes after the initial short-wavelength light illumination. d) Dependence of the RESET index on wavelength of the light used for the initial SET and the subsequent RESET operations. e) Equilibrium energy band diagram of the ZnO/Pt Schottky junction after irradiation with short-wavelength light (λ S ). The V O ionization reaction is also schematically illustrated (blue arrows). The black dotted line indicates the positions of E C before irradiation. f) Equilibrium energy band diagram after irradiation with long wavelength light (λ L ). The electron tunneling and jumping processes as well as the subsequent V O 2+ neutralization reaction are also schematically illustrated (blue arrows). The black dotted line indicates the positions of E C before irradiation. g) Capacitance-frequency characteristics of the AOC memristor. The device was first at the initial LMS (black curve), and then exposed to 530 nm light (red curve), and afterwards kept in the dark for 30 minutes (blue curve); subsequently, the device was

2+
ZnO Pt exposed to 650 nm light (pink curve), and finally kept in the dark for 30 minutes (green curve). In (a-d, g), the light power densities were maintained at 36 μW/cm 2 and the current values were measured at 10 mV. Note that both the V O ionization and V O 2+ neutralization reactions and electron motion actually occurred under nonequilibrium conditions. SET and RESET indexes were used to quantitatively study the optical SET and RESET efficiencies. The index values were calculated by the formula (I 1 -I 2 /I 1 )  100%, where I 1 and I 2 are the device currents before and after irradiation, respectively. As demonstrated in  The nonvolatility of the light-induced memconductance states is likely due to the following reasons: i) the free electrons generated in the Schottky barrier region were pulled into ZnO by the built-in electric field and therefore could not recombine with V O 2+ s, and ii) an energy barrier originating from the outward relaxation of bonds around the oxygen vacancy sites must be overcome to neutralize V O 2+ s. [60] The visible light (e.g., 530 nm) response of such wide band gap ZnO is due to an abundance of V O s with a wide distribution of energy levels given that ZnO was deposited in pure Ar. [22,61] It is also supported by photoluminescence measurement (Figure S6), in which broad deep emission bands ranging from 480 to 900 nm was observed. In addition, it has been reported that V O s in ZnO could have energy levels ranging from 0.2 to 1.3 eV below the conduction band minimum. [62] Contrary to the optical SET, an optical RESET is expected to result from widening of the Schottky junction due to a reduced density of V O 2+ . As mentioned previously, no significant photocurrent was generated by irradiation with 650, 725, and 800 nm light for the device in LMS ( Figure S5). Therefore, we can deduce that for the device after the optical SET, i.e., in an HMS, ionization of V O s could be ignored under irradiation with such long-wavelength light given a lower density of V O s in ZnO compared to the case without the optical SET (in LMS).
It has been reported that in a metal/oxide/metal junction, [63,64]  To verify the proposed memconductance tuning mechanism, i.e., optical SET and RESET originated from the respective light-induced narrowing and widening of the Schottky junction, we measured the device capacitance at different irradiation conditions (Figure 2g).
Capacitance of the device initially increased when irradiated with 530 nm light, but decreased upon a second irradiation with 650 nm light. As mentioned previously, the ZnO/Pt Schottky junction dominated the electrical properties under a positive voltage. This implies that the device capacitance could be controlled by the ZnO/Pt Schottky junction. It has been reported that the capacitance of a Schottky junction increased with decreasing junction width. [65] Hence, we could conclude that the 530 nm irradiation resulted in a narrowing of the Schottky junction (optical SET), whereas the 650 nm irradiation led to a junction widening (optical RESET).
We believe that during the optical SET process, short-wavelength light with relatively high photon energy also induced internal photoemission or photoassisted tunneling of

AOC Memristor
Based on the optical SET and RESET operations, we can realize the reversible tuning of memconductance by applying only optical excitation.  To compare the memconductance tuning performance between optically and electrically controlled memristors, we prepared Ti/ZnO/Pt and Cu/ZnO/Pt devices in which the ZnO layer was deposited at the same parameters as the AOC memristor (Au/ZnO/Pt). It has been reported that Ti/ZnO/Pt and Cu/ZnO/Pt demonstrated memristive switching due to electrically controlled electron trapping/detrapping and Cu nanofilament rupture/rejuvenation, respectively. [15,54] After an electroforming process, the Au/ZnO/Pt showed memristive switching based on electrically controlled rupture and rejuvenation of conducting nanofilaments composed of oxygen vacancies. [66] Figures S9-S11 illustrate electric-induced memristive behaviors of the Ti/ZnO/Pt, Cu/ZnO/Pt and electroformed Au/ZnO/Pt devices, respectively. By comparing the single memconductance increase/decrease cycles in Figure 3 and Figures S9-S11, we could observe that the memconductance increase/decrease curves of the AOC memristor and Ti/ZnO/Pt were much smoother than those of the Cu/ZnO/Pt and electroformed Au/ZnO/Pt. This could be attributed to the purely electronic memristive switching mechanism of the AOC memristor and Ti/ZnO/Pt. Moreover, we found that the cycle-to-cycle and device-to-device variations for the AOC memristor were much smaller than those for the other three devices (Figure 3 and S8-S11). Given that the functional layers of these four memristive devices were deposited at the same parameters, it could be deduced that the AOC memristor has a better memconductance tuning performance than the electrically controlled memristors based on switching mechanisms of both nanofilament rupture/rejuvenation and carrier trapping/detrapping. As for the Ti/ZnO/Pt, the global increase in memconductance during successive cycling likely resulted from the generated Joule heat given the rather high programming voltage and current (Figure S9a,c), whereas large variations in the initial memconductance among the devices might be due to various extents to which Ti reacted with ZnO ( Figure S9d). [15] We propose that the superior memconductance tuning performance of the AOC memristor could be attributed to its purely electronic memconductance tuning mechanism as well as an extremely small amount of heat generated during the programming process given a very low power density (≈ 30 μW/cm 2 ) of the programming light. To verify that such weak irradiation did not induce microstructure change in the AOC memristor, we exposed the device to 350 nm light with a power density of 36 μW/cm 2 for a period of time as long as one hour, followed by measuring its memconductance increase/decrease cycles ( Figure S12). By comparing Figure 3b and Figure S12, we were unable to observe any obvious performance deterioration after this long-term irradiation. Generally, optoelectronic characteristics of a semiconductor device are extremely sensitive to its microstructure change. Hence, the microstructure change could be excluded in the programming process of this AOC memristor given that the shortest wavelength used in our experiments was 350 nm.

Nonvolatile Neuromorphic Computing
Our AOC memristor could be used to perform in-memory computing, e.g., nonvolatile neuromorphic computing. As schematically illustrated in Figure 4a, a three-layer artificial neural network (ANN) was constructed with the CrossSim [67] simulator using the measured memconductance values in Figure 3b as synaptic weights. The ANN could be utilized to recognize handwritten digits. The following two datasets were employed to train the ANN: small images (8 × 8 pixels) of handwritten digits from the "Optical Recognition of Handwritten Digits" dataset [68] and large images (28 × 28 pixels) of handwritten digits from the "Modified National Institute of Standards and Technology" dataset. [69] During the training process, synaptic weights were updated based on a back-propagation algorithm. Figure 4b,c shows the cumulative distribution functions (CDFs) for optical SET and optical RESET processes in Figure 3b. Herein, CDF is the probability that ∆M C takes a value less than or equal to the ∆M C plotted. During training, the CDF was randomly sampled for weight update of each synapse. The training results for two datasets are illustrated in Figure 4d,e (blue curves). We could observe that after three epochs, the recognition accuracy for both small and large images of handwritten digits exceeded 92%. The yellow curves in Figure 4d,e show the simulation results of the ideal floating-point-based ANN (theoretical limit for the algorithm).
The recognition accuracy of the ideal ANN exceeded 98% after 40 epochs, indicating promising application prospects of our AOC memristor in image recognition.

Logic-in-Memory
Another type of in-memory computing, logic-in-memory, could also be performed in our AOC memristor. Boolean logic is a form of algebra in which the variable's values are the truth values, i.e., true and false (generally denoted as "1" and "0", respectively). It fits well with the binary numbering system used by modern computers, where each bit has a value of either 1 or 0. There are 16 Boolean logic functions in two-input (e.g., p and q) systems. Compared to traditional logic gates based on transistors, logic-in-memory computing architectures provide a more efficient way to store and process information. [10,[34][35][36][37][38][39] Nonvolatile Boolean logic could be demonstrated in our AOC memristor, in which the computing results were in situ stored as the memconductance states. As schematically illustrated in to the input light, the device was exposed to 530 and 650 nm light and set to an HMS and LMS, respectively; wavelength of the control light was selected according to the logic operation type. As for the NAND, IMP, RNIMP, p, and NOT p functions, the schemes of applying the control light strongly depend on the initial memconductance state (p = 1 or 0), which should be determined first, as well as the logic operation type and the input light. In the case of NOT q and NOR functions, the optoforming operations were necessary; the schemes of applying the control light were determined by the initial memconductance state  and logic operation type. Other than the above 14 logic functions, additional steps were needed to achieve the XOR and NXOR functions. Apart from initial state-dependent optoforming processes, it was necessary to determine the intermediate memconductance states prior to applying the subsequent control light. Furthermore, the control light applied before and after the input light was composed of both 530 and 650 nm irradiation.
As mentioned previously, memconductance decay occurs after light irradiation. In the case of pʹ = 0, such decay did not affect the device state since the memconductance gradually deviated away from the baseline value (100 nS). On the other hand, for pʹ = 1, the memconductance gradually approached the baseline value as a result of the decay. To confirm the nonvolatility of the output, the decay curves were fitted using exponential functions ( Figure S14). The fitting results show that the memconductance could maintain its relatively large value above 100 nS over time, thus indicating the nonvolatile output.

Conclusions
An AOC memristor with a simple Au/ZnO/Pt structure was fabricated. The memconductance could be reversibly tuned over a continuous range via varying only the wavelength of the controlling light. The device could be operated at light power densities as low as ≈ 30 μW/cm 2 and the light-induced memconductance states were found to be nonvolatile. The observed memconductance switching behavior most likely stemmed from a reversible width variation of the Schottky junction at the metal/ZnO interface due to electron trapping and detrapping at oxygen vacancies. The device showed excellent operation stability.
In-memory computing, including nonvolatile neuromorphic computing and logic-in-memory, was demonstrated using this AOC memristor, indicating its great potential as a candidate for high-performance computing.
It deserves mention that before AOC memristors could be practically used for in-memory computing, an answer to the following question must be found, i.e., how can light be controllably introduced into each device to tune its memconductance in a memristor crossbar array? Usage of thin-film optical waveguides based on Si/SiO 2 , Si 3 N 4 /SiO 2 , etc., might be a feasible and effective method to overcome this challenge. These waveguide materials are easy to be integrated into AOC memristors, making high-density integration of AOC memristors possible.

EXPERIMENTAL PROCEDURES
Resource Availability

Lead Contact
Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Fei Zhuge (zhugefei@nimte.ac.cn).

Materials Availability
This study did not generate any new unique materials.

Data and Code Availability
The published article includes all data analyzed during this study. of the Hall measurement system. Transmittance spectra were measured by a UV-visible-IR spectrophotometer (Lambda 950, PerkinElmer). Photoluminescence spectra were measured via a confocal microscopic Raman spectrometer (Renishaw inVia Reflex, 325 nm). These measurements were performed at room temperature in air.

Device Fabrication and Characterization
Au