CONTENTS

Introduction. Modern Information Processing

1. Memristive Systems

 1.1. Memristor Electronic Devices

 1.2. Two-Dimensional Crystals

 1.3. Memristors Based on 2D Crystals

2. Photomemristor Systems

2.1. Photomemristors Based on 2D MoS2 Crystals

3. Image Sensors and Photonic Storage Devices

 3.1. Arrays of Image Sensors Based on Graphene–    MoS2

 3.2. Photon Chip with Photon Synapses

Conclusions and promising lines of research

INTRODUCTION. MODERN INFORMATION PROCESSING

The need for processing large unstructured volumes of information imposes a requirement on modern digital electronics to develop efficient information technologies similar to natural forms, which use neuromorphic architecture and neural networks for fast and energy-efficient calculations and information storage. Digital computers, based on very-large-scale silicon integrated circuits, which were developed for solving problems with a fixed data type, for the recognition of patterns, sound, and motion are inefficient without using artificial neural networks. The physical separation of memory and processor in a digital computer (Fig. 1) limits performance in the case of operations with large volumes of information and leads to immense power consumption. For example, simulating the work of 1% of neurons of the human brain using the Fujitsu “K Computer” modern supercomputer (10 petaFLOPS (1016 billion operations per second) with a random-access memory of 1 petabyte) consumes a power equal to ~10 MW. Simulating a second of operation of the brain takes ~40 min. Simulating the operation of the cerebral cortex of a mouse with 104 neurons and 1010 synapses using the “Blue Gene” supercomputer with 109 transistors (~1 laptop per neuron) consumes ~0.1 MW. It is known that an increase in the performance of modern computers is limited by the scaling limit of the silicon complementary metal–oxide–semiconductor (CMOS) technology. A decrease in the size of transistors leads to the heating of integrated circuits and enormous power consumption upon the processing of large volumes of data. In 2020, Samsung manufactured a 3-nm chip, based on the Gate-all-around (GAA) transistor architecture [1], which comes close to the limit of the low-power technology [2]. This indicates the importance of developing new materials and new approaches to the storage and processing of large bodies of information.

Fig. 1.
figure 1

von Neumann architecture of a digital computer.

Full size image

The smart processing of digital information using artificial neural networks and advanced algorithms of brain-like large parallel computations makes it possible to increase significantly the efficiency and rate of unstructured-data processing [3]. Recently, various types of neural networks (e.g., convolutional [4], recurrence [5, 6], oscillator [7–9]) and reservoir computing [10–12] yielded rapid development and good prospects for efficient solution of the problems of synthetic vision and the recognition of speech and object motion. Various new materials and devices (e.g., spintronic and memristive oscillators [7, 9, 13], photon moduli [14–16], memristors [10, 17–19]) have been investigated to facilitate the hardware realization and high efficiency of reservoir computational systems. Significant progress has been achieved in the implementation of artificial neural networks based on memristors, due to the use of their analog resistive switching properties [20–25]. The main attention in this review is paid to the consideration of optoelectronic memristor devices based on two-dimensional crystals for neuromorphic sensor systems of artificial vision.

1 MEMRISTIVE SYSTEMS

1.1 Memristor Electronic Devices

Memristor electronic devices are similar to biological synapses in neural networks. Memristors are a new type of electronic logical gate and memory with extremely low power consumption and footprint [26]. These new electronic components may solve the problem of the physical and technological limitations of modern CMOS technology and become the element base for efficient big data processing and the development of artificial intelligence and synthetic vision.

The term “memristor” as a nonlinear resistive element was coined by L.O. Chua in 1971 to describe the missing fourth basic element of an electric circuit [27]. Along with the other known elements of electric circuits (capacitor, resistor, inductor), the memristor could describe nonlinear effects in solid-state electronics, which were already well known. In 1922, O. Losev observed a new phenomenon of negative differential resistance in a two-electrode point device (cristadyne) [28, 29], which was then used for the generation and detection of signals in radio broadcasting. Losev’s cristadyne made it possible to use frequencies up to volumes 100 MHz, which were unattainable and incomprehensible for application at the time. In 1957, L. Esaki independently demonstrated a similar nonlinear device—tunnel diode—and was awarded the Nobel Prize in physics in 1973 for discovery of this effect. In 2007, the first memristor from a ZnO nanorod was demonstrated experimentally [30]. The resistance switching in an electric field in ZnO:Fe,Li nanorods (Fig. 2) was explained by the influence of Fe, Mn, and Li impurities: they could both locally change their electrical properties and form magnetic domains [31, 32] (Fig. 2, inset), which provided multiferroic switching in these nanostructures. A nonlinear two-electrode memristive device, based on a TiO2 film, was demonstrated in 2008 by a group from Hewlett Packard [33]. This device consisted of two (doped and undoped) nanoscale regions, the relative shift of which controlled the ON and OFF states. The first memristor array was made using TiO2 on a CMOS crystal at the Hewlett-Packard laboratory in 2012. The memristor with two platinum electrodes was a nonlinear dynamic structure, the resistance of which depended on the electric field and current. This nonlinear device made it possible to form energy-independent states for information storage for the case of a switched-off power supply and to obtain an ultrahigh writing density, low switching energy, high operating speed, long storage time, and multilevel storage using discrete or continuous states. Some complex computational problems (e.g., image processing and pattern recognition) can be solved efficiently using the nonlinear dynamic approach. In particular, a smart photodetector by Toshiba with a universal chip, based on a convolutional neural network, can process images similarly to the human brain, which makes it possible to calculate elementary image-recognition problems for several nanoseconds. Memristors, i.e., electronic analogues of synapses in biological neural networks, can be used to form high-efficiency smart machines and computers with brain-like neuromorphic architecture. It is known that the human brain, which contains 1011 neurons and 1015 synapses (Fig. 3), processes analog information and consumes only about 20 W [34, 35]. The calculations, which are performed in the brain, require large parallel operations with a high degree of random connections between neurons for the analog processing of signals and memory states and the storage of changed memory states in synapses. Analog data processing is efficient for processing unstructured sensor data (e.g., image, video, sound, motion, etc.).

Fig. 2.
figure 2

I–V characteristics of a ZnO:Fe,Li nanowire 70 nm in diameter, recorded at a change in the voltage from 0 to 3 V and back and from 0 to –3 V and back. SEM image of the nanowire with aluminum electrodes, obtained using electron-beam lithography, and AFM image of a ZnO:Mn nanowire in an external magnetic field are shown in the top right and bottom left insets, respectively. The other magnetic-field direction leads to the switching of single magnetic nanodomains from spin up (upper part, bright region) to spin down (lower part, dark region).

Full size image
Fig. 3.
figure 3

Neural network.

Full size image

Electronic devices can have an architecture, similar to the living nervous system of the cerebral cortex, which contains ~0.15 quadrillion (1015) synapses (a trillion (1012/cm3) synapses per cm3). For comparison, a 7-nm FinFET chip (Samsung) has 40 billion gates (4 × 1010/cm2). At 20 years of age, the brain contains ~150 000–180 000 km of nerve fibers, which connect all these neural elements. Each neuron can connect to any other neuron by no more than six bonds (six degrees of separation). We note that about 85 000 neurons die in one day (about one neuron per second) [36]. Understanding the mechanisms of the self-organization and self-healing of neuromorphic systems will make it possible to form artificial electronic networks, similar to living neural ones, and to use not only error correction and element redundancy in computers similar to the natural world but also to form self-assembling and self-healing electronic components.

1.2 Two-Dimensional Crystals

Due to their unique electronic and optical properties, atomically thin two-dimensional (2D) materials (e.g., graphene, graphene oxide, molybdenum disulfide, etc.) exhibit great potential for the formation of nanoelectronic and bioelectronic devices of ultrahigh density for cutting-edge information systems. Graphene is a crystalline 2D layer of carbon with a thickness of one atom. The unit cell in a hexagonal graphene lattice is described by vectors a1 and a2, which contain two atoms from two equivalent carbon sublattices (Fig. 4). Great interest in this material emerged in 2004 after publications of researchers from the Institute of Microelectronics Technology and High-Purity Materials, Russian Academy of Sciences, and Manchester University about the influence of an electric field on atomically thin carbon films [37]. In 2010, A. Geim and K. Novoselov were awarded the Nobel Prize in physics “for groundbreaking experiments regarding the two-dimensional material graphene”. Two symmetric carbon graphene sublattices form a Dirac cone with the linear energy dispersion of carriers (Dirac fermions). A specific feature of these particles is that they are massless and behave as photons. As a consequence, graphene exhibits unique properties. Graphene is transparent (97.7%) and stable to extremely high current densities (its stability is a million times higher than that of copper). It has the highest electron mobility among known materials (~106 cm2 V–1 s–1, which is higher than that of silicon by 3 orders of magnitude) and a very high thermal conductivity (K > 5 × 103 W/(m K), which is higher than that of diamond). Graphene is a well stretchable (25%) material with unique mechanical strength E > 1012 Pa (higher than that of an atomic steel layer by a factor of 6). In addition, graphene is biocompatible. Neurons of the retina can survive and grow neurites on graphene [38]. Moreover, arrays of graphene field transistors can record the action potentials from cells [39], which makes it possible to develop an artificial retina display. Operation of this display is based on processing data, arriving from a video camera, integrated into glasses, and its further transmission to a graphene retina implant, which stimulates nerve cells for transferring the signal to the brain [40]. In 2017, at the Mobile World Congress in Barcelona, Catalan Institute of Nanoscience and Nanotechnologies (ICN2) a retina prosthesis was demonstrated (Fig. 5) [41], which was based on graphene due to its advantageous electrical properties, flexibility, and biocompatibility. These devices are promising for eyesight recovery for persons with lost functionality of photosensitive retina cells. The sensor responses of a built-in graphene transistor (GFET) to visual-cortex excitation are shown in Fig. 6 [42]. These recordings were obtained from the primary visual cortex and primary auditory cortex of rats during sensor stimulation. ON/OFF switching of the visually induced reaction occurred 40 ms after stimulus. The signal contains the main component with a frequency of ~20 Hz with a maximum amplitude of 250 µV and width of 70 ms.

Fig. 4.
figure 4

Hexagonal graphene lattice. A unit cell, determined by vectors a1 and a2, contains two carbon atoms from equivalent sublattices.

Full size image
Fig. 5.
figure 5

Graphene-based retina prosthesis [41].

Full size image
Fig. 6.
figure 6

Sensor evoked responses of the built-in GFET: (a) visual evoked potential, (b) spectrogram of the visual elicited potential, (c) auditory evoked potential, and (d) spectrogram of the auditory evoked potentials [42].

Full size image

Photosensitive structures, based on 2D photomemristor materials (molybdenum disulfide and graphene/graphene oxide (G/GO)) and quantum dots (QDs) on their basis [43–45], which are biocompatible and make it possible to use a neuromorphic architecture for analog calculations and self-assembly technologies, are of particular interest for the development of new photosensitive devices. The photocatalytic oxidation of graphene is an efficient method for preparing 2D memristors with photoresistive switching for ultrahigh-capacity nonvolatile memory. A new type of multifunctional memristors with photogates, controlled by electrical and optical signals in a wide wavelength range, can be used for image processing and the recognition of patterns, sounds, motions, and speech, which are necessary for the formation of artificial intelligence and for the interaction of electronic networks and neurons.

1.3 Memristors Based on 2D Crystals

Graphene/Graphene Oxide Memristor Heterostructure. In 2010, researchers from the Institute of Microelectronics Technology and High-Purity Materials of the Russian Academy of Sciences, Moscow State University, and Dongguk University demonstrated a memristor based on a “graphene/graphene oxide (G/GO)” structure, which switched at 0.7 V and 1 nA with an ON/OFF current ratio of ~103 [46, 47]. The formation of randomly distributed current channels-filaments (Fig. 7) was observed with high spatial resolution using the method of electron beam-induced current and the switching mechanism in this synapse-like device was studied. The resistance of this device in an electric field changed nonlinearly and the values of high and low resistance were non-volatile. Studying the resistive-switching mechanisms in G/GO showed that the migration of oxygen-containing groups plays an important role [48–51]. One carbon bond in sp3 hybridization (C–O bond) per 106 carbon bonds in sp2 hybridization decreases the conductivity of carbon nanomaterials by 50% [52]. Graphene oxide with dominant carbon in sp3 hybridization with low electrical conductivity exhibited switching in an electric field due to the local carbon transition to sp2 hybridization, which yielded a high electrical conductivity. This process could be controlled by varying the migration of oxygen groups.

Fig. 7.
figure 7

Memristor structure based on graphene/graphene oxide. (On the left) SEM images of the Al/GO/Al structure, obtained in the remote electron-beam induced current (REBIC) mode with modulation of the built-in potential barrier at the negatively biased Al electrode at different bias voltages (Vb) and forming voltages (Vf). (top left) Vb = 1 V, Vf = 0 V (REBIC); (center left) Vb = 0.4 V, Vf = 5 V (REBIC); and (bottom left) Vb = 0.7 V, Vf = 5 V (REBIC). (On the right) schematics of the Al/GO/Al structure at different Vb and Vf voltages [47].

Full size image

In 2017, IBM announced multilevel high-speed nonvolatile memory, based on graphene oxide [53]. A memory cell with a thickness of 8 nm and a size of 50 nm had four-level data storage (2 bits per cell) with writing and erasing times of less than 5 ns, and also exhibited high reproducibility and the possibility for long-term data storage. Memory with the possibility of storing more than one bit per cell (i.e., with multilevel memory states) is very attractive, because it provides a simple and economical way to increase memory capacity (e.g., modern CMOS NAND-Flash usually stores 2 or 3 bits per cell). Combining the possibility for multilevel storage with extremely high scalability is especially efficient for implementing memory with ultrahigh storage capacities. In this structure, based on graphene oxide, access to four well-separated and stable memory states in nanoscale cells was demonstrated by tracking the writing-pulse width and amplitude [53]. Excitation pulses with amplitudes from 2 to 6 V and widths from 20 to 80 ns were used to determine the conditions for the successful writing and erasing of multilevel memory states in a Pt/GO/Ti/Pt structure and to monitor the obtained cell resistance. The cells were completely switched from the RESET state, which can be considered state 00, to memory states 01, 10, and 11 using –2.5/60, –3.5/60, and –4.5 V/60 ns pulses, respectively. The erasure of cells from states 01, 10, and 11 back to state 00 was successfully achieved using +3/60, +4/60, and +5 V/60 ns pulses, respectively. Good separation of the intermediate-resistance levels made it possible to ensure a reliable reading process. Intermediate levels exhibited high reliability and time stability on both rigid and flexible substrates. Reversible resistive switching, which was observed in those devices, was related to oxygen migration and, as a consequence, lead to a change in the conductivity.

The photocatalytic oxidation of two–three graphene layers, covered with a layer of ZnO nanoparticles, spaced by a distance of ~10–15 nm, under UV irradiation in the presence of water molecules led to the formation of self-assembled G/GO memristors with a very high density (1012 cm–2) [49, 51]. Graphene was synthesized on a 25-µm thick-copper foil at 1020°C using chemical vapor deposition from a mixture of methane and hydrogen [49]. The grown graphene was transferred to a SiO2/Si substrate using polymethylmethacrylate. ZnO nanoparticles (NPs) were obtained via the thermal decomposition of a zinc salt [49]. A 0.005-M solution of acetate zinc in ethanol was deposited using centrifugation on a graphene/SiO2/Si substrate at 1000 rpm and then annealed at 350°C for 40 min with the formation of a NP layer. Graphene, consisting of two-three layers and covered with NPs, was irradiated in a flux of wet air at room temperature or above (80°C) using a quartz UV lamp with a luminous flux of 0.03 J min–1 cm2 for its photocatalytic oxidation. Light with a wavelength of more than 365 nm was filtered. The duration of UV irradiation was from 5 to 90 min. After UV treatment, ZnO NPs were dissolved in diluted 0.1-M HCl and the G/GO substrate was washed with deionized water and dried in nitrogen. Figure 8 shows the schematic diagram of the photocatalytic oxidation of graphene with ZnO NPs and the electronic band diagram of the graphene/ZnO interface under UV irradiation. ZnO NPs play a key role in the graphene photooxidation process. Band bending in ZnO NPs is caused by the lower work function in ZnO (3.6 eV) in comparison with graphene (4.5 eV). The electron–hole pairs, generated in ZnO (3.3 eV) under UV irradiation (reaction 1), are separated in an intrinsic electric field at the graphene/ZnO boundary, which provides a hole flux (3.3 eV) into graphene. As a result, graphene was oxidized by highly active hydroxyl radicals (OH) (reactions 3–5), formed during the photodecomposition of water molecules from wet air. Controling the distribution of ZnO NPs on graphene at the nanoscale level (~10 nm) makes it possible to form highly scalable G/GO nanoheterojunctions for ultrahigh-density memory (up to 1012 cm–2 or 1 TB per chip with a vertical geometry of crossing electrodes, Fig. 8c). The states of memristors with floating photogates can be read electrically and under optical excitation. The current–voltage (I–V) characteristic of graphene before oxidation is linear and exhibits high conductivity (Fig. 9а). The photocatalytic process decreases the conductivity of graphene and causes a nonlinear I–V characteristic, which indicates the formation of a band gap in oxidized graphene. An increase in the wet-air temperature reduces the oxidation time. The G/GO heterostructures, obtained using photocatalytic oxidation in wet air at temperatures of 20 and 80°C for 30 and 5 min, respectively, exhibit nonlinear I–V characteristics, which indicates the occurrence of a band gap (~3 eV) and a decrease in the oxidized-graphene conductivity by two orders of magnitude. The formed G/GO nanostructures exhibited good photosensitivity to illumination and photoresistive switching. The photocurrent increased by a factor of about 6 at a bias voltage of >3 V, which indicates that the light-induced electron–hole pairs, are separated efficiently in electrically biased G/GO heterojunctions. Figure 9b shows the I–V characteristics of the previously formed G/GO nanostructure (+5 V, 15 min) in the voltage range from –4 to 4 V under illumination and in the dark. The well-reproducible bipolar hysteresis indicates resistive switching of the structure with an ON/OFF current ratio of ~10 for four different resistive states HRSD, LRSD, LRSL, and HRSL in the dark and in light with switching voltages of 3.8/3.3 V (RESET/SET) and –3.5/4 V (SET/RESET), respectively (Figs. 9c, 9d).

Fig. 8.
figure 8

(a) Scheme of the photocatalytic oxidation of graphene, covered with ZnO nanoparticles, under UV illumination in the presence of water molecules with the formation of G/GO heterostructures on the Si/SiO2 substrate. (b) Band diagram of the G/ZnO interface under UV irradiation. The electron–hole pairs, generated in ZnO (Eg = 3.3 eV) under UV irradiation (reaction 1), are separated in the built-in electric field at the G/ZnO interface, providing a hole flux into graphene. (c) Schematic representation of the array of G/GO photomemristors in the geometry of crossing graphene electrodes, obtained by the photocatalytic oxidation of graphene with ZnO nanoparticles.

Full size image
Fig. 9.
figure 9

(a) I–V characteristics of the G/ZnO structure before and after (black and red curves, respectively) photocatalytic oxidation in wet air for 30 min at 20°C (the structure with lateral gold electrodes is shown in the inset). (b) I–V characteristics of the preliminary polarized (+5 V, 15 min) G/GO nanostructure (black curve) under illumination and (red curve) in the dark. (c) Resistive states of the G/GO photomemristor, switched by bias voltages of –3.8/3.3 V (RESET/SET) in the dark and –3.5/4 V (SET/RESET) under pulsed illumination and (d) read at 2.5 V [48].

Full size image

The vertical memristive G/GO structures were fabricated with ZnO nanorods, grown on graphene (Figs. 10а, 10b). The ZnO rods were used for selective excitation  of  the G/GO/ZnO structure using ~380-nm UV radiation. Resistive switching in such a structure was observed at voltages of <1 V with the a high/low resistance ratio of 103 after forming at 1 V (Figs. 10c, 10d) [51]. The obtained data indicate good prospects for using graphene-based structures for high-density memristive devices with low power consumption [54].

Fig. 10.
figure 10

(a) Scheme of the arrays of G/GO/ZnO nanorod photomemristors in the vertical geometry and (c) SEM image of the structures with their I–V characteristics (b) before and (d) after forming.

Full size image

2 PHOTOMEMRISTOR SYSTEMS

2.1 Photomemristors Based on 2D MoS2 Crystals

The MoS2 photomemristor structure, polarized in an electric field under various illumination conditions, exhibits multilevel switching [55]. MoS2 nanocrystals were grown using the hydrothermal method. A MoS2 dispersion in isopropyl alcohol was deposited using centrifugation on a Si/SiO2 substrate and then annealed at 200°C for 15 min. The electrodes were fabricated from gold using photolithography and the “explosive” process. After lithography, the samples were annealed in a flux of 200 cm3/min (H2/N2 = 2 : 3) at 450°C for 15 min. Figure 11 shows the I–V characteristic of the Au/MoS2/Au structure (Fig. 11а, inset) after polarization at 3 and 6 V. The nonlinear characteristics of the device with hysteresis indicate the memristive behavior. In addition, the memristor exhibits a photoresponse to illumination. When the device is polarized at 3 V, smooth HRSL3–LRSL3 switching is observed under illumination and HRSD3–LRSD3 switching is observed in the dark with ON/OFF current ratios of about 2 and 4 at 1.2 and 0.7 V, respectively (Fig. 11а). At higher voltages (6 V), the device demonstrates sharp HRSL6–LRSL6 switching upon illumination at 2.9 V with an ON/OFF current ratio of about 10 and smooth HRSD6–LRSD6 switching in darkness with the ON/OFF current ratio of about 3 at 0.7 V (SET process of recording of the ON state, Fig. 10b). When the applied voltage is changed from 0 to positive (4.2 V), the device returns to HRSL6 (RESET operation). The memristive behavior of the device in the dark and under illumination is well reproduced for more than 1000 cycles (Figs. 10c, 10d) and demonstrates the possibility for the formation of multilevel resistive switching in the structure and for controlling it using an electric field in the dark and under illumination. We note that resistive switching, controlled using the polarization of MoS2 nanospheres, is a more rapid process than ion transport and the optical-access frequency is much higher than electrical addressing. A memristor, polarized at 3 V in the dark or under illumination, demonstrates four states, which are read at voltages of 0.7 (HRSD3 and LRSD3) and 1.2 V (HRSL3 and LRSL3) in the dark or under white-light illumination [55]. Memristor polarization at 6 V in the dark or under illumination leads to the formation of another four states, which are read at voltages of 0.7 (HRSD6 and LRSD6) and 4 V (HRSL6 and LRSL6) in the dark or under illumination. These states can be controlled electrically and optically, which is confirmed by iterative memristor operation under various writing and reading conditions [55]. The polarization of nanospheres in a photomemristor using an electric field and light pulses induces the formation of multilevel states. Analysis of the conductivity in these resistance states yields that the polarization of nanospheres under illumination leads to the formation of conduction channels. A decrease in the gap between electrodes decreases the device operating voltage significantly. Modulation of the barrier height at the boundaries of nanospheres in an external electric field under illumination due to depolarization is a high-efficiency process for high-speed signal processing. A memristor, polarized at 3 and 6 V, has different states, which can be electrically read under illumination in the form of four high-resistance states and four low-resistance states. Optical and electrical polarization of the structure along with fast phase transitions in MoS2 [45, 56] provides several nonlinear dynamic processes, which make it possible to form a system with neuromorphic architecture, similar to a neural network.

Fig. 11.
figure 11

Resistive switching of the Au/MoS2/Au nanosphere photomemristor. (a) I–V characteristics of the structure in the dark and under illumination (excitation spectral maxima at 2.7 and 1.8 eV; the structure of the device under illumination is shown in the inset). The voltage-sweep directions are indicated by arrows on the curves. The I–V characteristic of the structure after its polarization at 3 V. The device is switched gradually from HRSL3 to LRSL3 under illumination and from HRSD3 to LRSD3 in the dark with ON/OFF current ratios of about 2 and 4 at 1.2 and 0.7 V, respectively. (b) I–V characteristic of the structure after polarization at voltage of 6 V. The device exhibits a sharp change in the resistance under illumination from HRSL6 to LRSL6 at –9.2 V with an ON/OFF current ratio of about 10 and a smooth transition from HRSD6 to LRSD6 without illumination with an ON/OFF current ratio of about 3 at 0.7 V. The memristive characteristics of the structure after a series of cycles (c) in the dark and (d) under illumination [55].

Full size image

3 IMAGE SENSORS AND PHOTONIC STORAGE DEVICES

3.1 Arrays of Image Sensors Based on Graphene–MoS2

In 2017, an ultrathin hemispherical array of image sensors was demonstrated as a visualization element in a soft retina implant [57]. The CurvIS array is used as a soft implantable optoelectronic device, similar to the human eye, which can detect optical signals and subject visual nerves to programmed electrical stimulation with minimum mechanical side effects on the retina. An ultrathin MoS2-based 2D crystal [58–60] is a promising candidate for the role of the photoabsorption component in the omnidirectional CurvIS array due to its unique advantages (e.g., excellent photoabsorption coefficient (>5 × 107 m–1) [61], high photosensitivity (2200 A/W) [62], and high failure strain (~23%) [59]). The softness [59–63] and hyperfine thickness [64, 65] of MoS2 are additional factors, which makes it possible to fabricate a soft optoelectronic device. The high-density CurvIS array from MoS2–graphene successfully recognizes various projected images without the IR noise of a biological object. The structure of a single photodetector is shown in the inset in Fig. 12a. The transfer curve (Id–Vg) shows the typical behavior of a photosensitive transistor (Fig. 12b). Under illumination (515 nm), a MoS2 channel generates a photocurrent, the normalized value of which (Id/Idark) is proportional to the illumination intensity (Fig. 12c). The photosensitivity of the MoS2–graphene photodetector is compared to the theoretical photosensitivity of a silicon photodiode, the silicon thickness in which is the same as that of MoS2 (Fig. 12d). It can be seen that the photosensitivity of the MoS2–graphene structure exceeds that of a silicon photodiode by 2–3 orders of magnitude, which is related to the efficient photoabsorption of MoS2 [57]. Conventional silicon image sensors absorb IR radiation (850 nm) (Fig. 12e, inset), which causes IR noise. The MoS2 photodetector does not absorb IR radiation because it has a wider band gap (Fig. 12e). Therefore, a MoS2-based device does not require an IR filter, which makes it possible to decrease the thickness and to increase the softness of the CurvIS array. The CurvIS array visualizes sigma (Σ) (Fig. 12f) using one planar-convex lens. IR radiation does not affect the captured image (Fig. 12g) because of IR transparency of the MoS2-based structure (Fig. 12e), whereas the array of conventional silicon photodetectors without IR filters shows IR noise (Fig. 12g, inset on the right). The ultrathin CurvIS array, the shape and mechanical softness of which are the same as those of the retina of a human eye, has great potential for use as a soft photodetector component in retina prosthesis. The human eye consists of a light-collecting lens; retina, which transforms light to action potentials; and visual nerves, which transfer action potentials to the brain (Fig. 13a). A soft implantable optoelectronic device detects light on a large area (like a human retina) (Fig. 13b), imitates the structural features of the human eye, and barely causes its deformation. The stress for this model of an artificial eye is lower by an order of magnitude than those for other models (0.61 MPa, Fig. 14) [57].

Fig. 12.
figure 12

Device characteristics and imaging using a curved array of image sensors. (a) Image obtained using an optical camera of the array of phototransistors in the form of a truncated icosahedron on a flat substrate. The device structure is shown schematically in the inset. (b) The transfer curves of the phototransistor at different light intensities (515 nm). (c) Normalized photocurrent change under different light intensities. (d) Photosensitivity of the MoS2–graphene structure in comparison with a silicon photodetector of the same thickness. (e) Normalized change in the photocurrent under IR illumination (850 nm) of different intensities. Inset shows the light absorbance of MoS2 and silicon. (f) Sigma-shaped image captured by the CurvIS array. (g) The same image with Fig. 12f but captured under IR illumination. Inset images are acquired by a commercial silicon photodetector array with (left) and without (right) an IR filter under IR illumination [57].

Full size image
Fig. 13.
figure 13

Soft optoelectronic device, similar to the human eye [57].

Full size image
Fig. 14.
figure 14

Mechanical stress, caused by various implanted devices [57].

Full size image

Figure 15 shows the recording of nervous signals from the visual cortex upon optical and electrical stimulation of the retina [57]. The excitation of visual nerves was controlled using electrodes, penetrating to the nerve. Stimulation of the retina [66, 67] is confirmed by the elicited spikes [68] and changes in the local-field potential (LFP) [69], which were measured simultaneously in the primary visual cerebral cortex of a rat [68, 70]. When pulsed optical signals are supplied to the rat eye, the rat retina accepts ON/OFF light switching. This optical information is transferred to the visual cortex through the visual nerves, which induces the formation of spikes (Fig. 15b) and a change in the LFP in the frequency range of 4–15 Hz (Fig. 15c). Exactly in the same manner, a soft optoelectronic MoS2–graphene device detects ON/OFF light switching, and the corresponding electrical pulses are fed to the visual nerve through ultrathin neurointerface electrodes. Therefore, the visual cortex is similarly activated, which leads to the occurrence of spikes (Fig. 15d) and changes in the LFP in the same frequency range (Fig. 15e).

Fig. 15.
figure 15

Stimulation of the retina using a soft optoelectronic device. Recording of neural signals from the visual cortex. (a) Measurement of elicited spikes and (b) changes in the local-field potential (LFP) in the visual cortex using optical stimulation. Measurement of (c) elicited spikes and (d) changes in the LFP in the visual cortex using electrical stimulation [57].

Full size image

3.2 Photon Chip with Photon Synapses

A photon chip, containing 70 photon synapses, was demonstrated in 2017 by a team from the Universities of Oxford, Münster, and Exeter [71]. In this case, data writing, erasure, and reading are performed exclusively by optical methods.

The photon synapse consists of a tapered waveguide (Fig. 16) with discrete islands of material with the phase transition from above, which connect optically pre-synaptic (pre-neuron) and post-synaptic (post-neuron) signals. The use of exclusively optical tools ensures an ultrafast operating speed, almost unlimited transmission band, and the absence of power losses in interconnections. It is noteworthy that the synaptic weight can be set randomly, simply by changing the number of optical pulses, which form a system with constantly changing synaptic plasticity, reflecting the true analog nature of biological synapses. Synaptic tuning of the device upon switching between the crystalline and amorphous states of Ge2Sb2Te5 (GST) islands with a recorded change in the relative transmittance is shown in Fig. 17. Five weighting states of a photon synapse are obtained by switching the optical-pulse energy (404.5 pJ, 50 ns). The photon synapse exhibits good reproducibility of the weighting numbers upon cyclic measurements. In this case, the photon-synapse weight is determined by the number of optical pulses.

Fig. 16.
figure 16

On-chip photonic synapse. (a) Structure of neuron and (inset) synapse. (b) Schematic of an integrated photonic synapse, which models the function of a nerve synapse. The synapse is based on a tapered waveguide (horizontal path) with vertical discrete PCM islands on top, which optically connect pre-synaptic (pre-neuron) and post-synaptic (post-neuron) signals. The open circle is a circulator with ports 2 and 3, which connects the synapse and post-neuron; weighting pulses are supplied to the synapse through port 1. (c) Optical image of the device with the active region (rectangle on top), used as a photon synapse. The optical input and output of the device are implemented using anodized diffraction couplers (rectangles at the bottom). Typical photonic chip, which contains 70 photonic synapses and is smaller than a coin. (d) SEM image of the active region of the photonic synapse, which corresponds to the top rectangle on the optical image, with six vertical GST strips (1 × 3 µm2) at the end of the waveguide. An enlarged view of the conical waveguide structure, indicated by the dotted frame, is shown in the inset [71].

Full size image
Fig. 17.
figure 17

Synaptic weight and plasticity. (a) Demonstration of the differential synaptic weight of the device, shown in Fig. 16, upon switching between crystalline and amorphous states of the GST islands with the recorded relative change in the coefficient ΔT/T0. Each weight can be obtained using identical number of pulses (50 ns at 243 pJ, 1 MHz) from any earlier weight. (b) Weight reproducibility during several cycles. The statistical analysis of a change in the weight readings “0”, “1”, and “4” is shown in the inset. The supplied pulse was 50 ps at 320 pJ, which is somewhat larger than upon measurements, shown in panel (a). (c) Five photonic synapse weights are obtained upon switching of the optical pulse energy (404.5 pJ, 50 ns). Dotted rectangles on the left and right correspond to the first and last weighing cycle, respectively. The up and down arrows in the rectangles indicate the weighting directions. (d) Photonic synaptic weight (ΔT/T0) as a function of the number of optical pulses. The left and right panels correspond to the left and right rectangles in panel (c). Filled triangles (open squares) are the data for the ascending (descending) weighting direction. Dotted curves correspond to the exponential curves, which are the closest to the experimental data [71].

Full size image

CONCLUSIONS AND PERSPECTIVE RESEARCH DIRECTIONS

Memristive systems, based on 2D crystals, comprise a new class of non-volatile electronic components, which make it possible to perform high-efficiency calculations and the real-time reading of visual information with low power consumption. Self-assembled synapse-like memristors, based on graphene and controlled by transitions of carbon in sp3 and sp2 hybridizations under an electric field, can be used for the interaction of artificial neural networks and biological neurons. 2D photomemristors with floating photogates demonstrate several states, which are controlled in a wide range of electromagnetic radiation, and can be used in a wide circle of problems, bound with neuromorphic calculations and the processing and recognition of patterns, which are required for solving the artificial-sight problems. Development of technologies of the formation of self-organizing artificial neural networks and heterointerface interactions, based on biocompatible 2D crystals with neurons, will make it possible to develop stable neuromorphic systems.