Introduction

Diamond (sp3-carbon material) and graphene (sp2-carbon material) are typical carbon allotropes with important electronic applications [1,2,3,4,5,6,7,8]. Graphene shows promise as a component of high frequency devices because of its very high carrier mobility that originates from the Dirac cone-type band structure of this material [1,2,3,4]. In addition, diamond has been considered for roles in next-generation high-power devices, especially those operating under harsh conditions such as high temperatures or exposure to radiation. These applications are suggested because diamond has suitable physical properties, such as a large band gap, high thermal conductivity, and a high breakdown field in addition to good chemical stability [5,6,7,8].

Recently, diamond/graphene (carbon sp3-sp2) interfaces have attracted much attention because these regions can exhibit various interesting electronic phenomena [9,10,11,12,13,14]. Such interfaces are expected to involve bonding states intermediate between those normally associated with sp3 or sp2 hybridization in conjunction with various carbon structures depending on the number of layers and the orientations of the diamond (that is, the three-dimensional (3D) carbon structure) and graphene (the two-dimensional (2D) structure) in the diamond/graphene heterostructure (Fig. 1). As an example, in the case for vertically aligned graphene/diamond (100) interfaces, trigonal carbon structures are formed, whereas hexagonal carbon structures are generated by normal graphene/diamond (111) interfaces. Different electronic properties can therefore be obtained depending on the interfacial carbon structure. The authors considered that an intermediate carbon structure could be switched between sp3 like (insulative) and sp2 like (conductive) by an external stimulus, such as an electric field or light, leading to variable resistance states (see the right lower part of Fig. 1). Such switching was achieved in the present study to produce photoinduced conductivity mimicking the functioning of the brain using graphene/diamond (C sp3-sp2) junctions. There have been many studies of interfaces exhibiting various interesting physical properties such as highly efficient photoelectronic conversion and the generation of different topological interfaces [9,10,11,12,13]. Even so, there have only been a limited number of experimental reports concerning the electronic characteristics of carbon sp3-sp2 interfaces [14], likely because of the difficulty in synthesizing high quality specimens.

Figure 1
figure 1

Research concept for diamond/graphene (carbon sp3-sp2) interfaces toward novel electronic applications.

Full size image

The authors have previously explored the novel characteristics of carbon sp3-sp2 interfaces based on fabricating various graphene/diamond heterostructures. This prior work demonstrated that heterojunctions between vertically aligned graphene and semiconducting diamond (100) can be used to form photo-controllable memristors (photomemristors). Such devices may exhibit both photo-switching and memory functions [15, 16]. These junctions also showed unusual responses to optical pulses and were able to mimic memory functions of the human brain. Specifically, the duration over which photoconductivity was retained (corresponding to the memory duration) was found to vary depending on the number and frequency of optical pulses to which the device was exposed [17]. This phenomenon resembles the transition of brain synapses from the short to long-term memory state. The results of this work also showed that arrays fabricated from graphene/diamond junctions could be used as special types of image sensors capable of storing optical information in a selective manner dependent on the relative importance of the data, similar to human brain functioning. These data suggest that diamond/graphene junctions show promise with regard to the development of novel neuromorphic devices.

As a note, neuromorphic computing devices using novel materials are being developed vigorously by many researchers to overcome the limitation of the present von Neumann-type computers, where sensing, storage and computing units are physically separated and data are processed sequentially in each unit (so-called, Neumann bottle neck). One of most promising materials for the neuromorphic computing is a memristor since its analog resistance changes with nonvolatile memory functions are equivalent to memory characters of synapses in human brains. More effective computing is possible by using the memristors because they have both memory and processing functions simultaneously (Fig. 2) and the Neumann bottle neck problem can be avoided. Our graphene/diamond photomemristors (photo-driven) are superior to the electric-field driven normal memristors because they also have a sensing function in addition to memory and processing functions. They are very promising materials for more efficient and lower-cost computing.

Figure 2
figure 2

Comparison of present Neumann computing and neuromorphic computing using novel materials such as memristors and photomemristors.

Full size image

The present study confirms that the graphene/diamond junctions have very high potential for use as novel optoelectronic neuromorphic devices, especially in the construction of complex image recognition systems mimicking brain functions.

Experimental methods

Semiconducting boron-doped homoepitaxial diamond films were grown on commercial diamond (100) Ib substrates by microwave plasma chemical vapor deposition (CVD) [18]. The acceptor concentration and mobility at room temperature in each film were typically on the order of 1017 cm−3 and 1000 cm2/Vs, respectively. Vertically aligned graphene layers (that is, carbon nanowalls; CNWs) were formed on these films via an in situ process. By precise tuning of the growth parameters, including the growth temperature, CH4/H2 ratio, plasma power, and working pressure, diamond and graphene films could be individually formed in situ using this method. Note that this film fabrication process has been described in greater detail in previous papers [16, 17]. Raman spectra of the resulting CNW/diamond heterostructures were acquired using an excitation wavelength of 532 nm. Interfacial structures between CNW and diamond layers were characterized by high-resolution scanning transmission electron microscopy (STEM), using a JEM ARM200F STEM operated at 80 kV. The cross-sectioned (parallel to the diamond [110] plane) thin film for STEM was picked up first using a focused Ga ion beam, followed by low energy Ar ion milling for final thinning and removing the surface damage layers, using Nano-Mill™ (Fischone Inc.). CNW/diamond junctions were fabricated from these heterostructures using standard photolithography processes and reactive ion etching [17]. The current–voltage characteristics of these devices with and without photo-irradiation were assessed based on exposure to blue or white LEDs in air at room temperature.

Results and discussion

Figure 3 shows a typical Raman spectrum obtained from a CNW/diamond heterostructure fabricated by plasma CVD, in which five major peaks can be observed. These comprise G, 2D, and D peaks [19, 20] that are associated with graphene layers having edge structures, a D' peak characteristic of CNWs [21] and a peak (labeled “Dia.”) originating from the diamond layers [22]. These results confirm that the desired CNW/diamond heterostructure was obtained. Figure 4 shows the high-resolution STEM image viewed along the [110] direction of the diamond substrate for CNW/diamond interfaces (upper part) and the bonding relations between graphene (001) layers in CNW and diamond (111) surfaces (bottom part). From the STEM image, graphene’s basal planes are connected to diamond (111) microfacets, which were formed during CVD growth, parallel to the diamond (111) surfaces. As a consequence of the different interlayer distances in the graphene (001) and diamond (111), unsaturated sp3 bonds were present within the diamond layers. This occurred because two graphene layers were connected to both end layers in three diamond (111) layers. Various slightly different types of connections at the interface are also evident in the TEM image and these complicated interfacial carbon structures are thought to be closely related to the unique optoelectronic characteristics of CNW/diamond junctions, although the associated mechanisms must still be elucidated.

Figure 3
figure 3

A typical Raman spectrum acquired from a CNW/diamond heterostructure.

Full size image
Figure 4
figure 4

Upper: A high-resolution STEM image viewed along the [110] direction of the diamond substrate for the CNW/diamond interfaces. Lower: A diagram showing the bonding relationship between a graphene (001) [G (001)] layer in CNW and a diamond (111) surface [D (111)].

Full size image

Figure 5(a) plots the photocurrent generated at three typical junctions at a bias voltage of 10 V before and after irradiation with 10 optical pulses using the blue LED. Each current was found to increase in response to the optical pulses and then to decrease after cessation of this stimulation. The decay rate of the current also varied between the specimens, suggesting that this factor could be controlled by modifying the growth conditions [17]. This unusual photoconductive behavior (that is, photomemristive behavior) is rarely observed and is usually only seen in oxide or transition metal chalcogenides heterostructures [23,24,25]. The current decay can be described by the equation G(t)/G0 = exp[− (t/τ)β], which is often used to analyze the relaxation of human memory [26, 27]. Here, G(t)/G0 is the normalized current, where G(t) and G0 are the current at time t and maximum current immediately after applying the 10 optical pulses, respectively, τ is the relaxation time (corresponds to memory duration time) for the photo-induced current, and β is an index ranging from 0 to 1. The present current decay curves were well fitted using this equation and the corresponding τ and β values were estimated to be 13,740 s and 0.48 for the long-term memory sample, 214 s and 0.33 for the short to long-term memory sample, and 1.9 s and 0.49 for the short-term memory sample, respectively. It is apparent from Fig. 5(b) that τ decreased as the ID/IG ratio (the ratio of the intensity of the D peak to that of the G peak in the Raman spectra) increased. A larger ID/IG ratio indicates a higher concentration of sp3-related defects in the graphene layers [19], which are considered to originate from oxidation and/or edge formation in the graphene layers (sp2 carbons) in the CNW. Hence, those junctions having more highly oxidized graphene layers and/or more edges showed smaller τ values. Depending on their τ values, the CNW/diamond junctions also exhibited different optoelectronic properties. In particular, the higher τ junctions (that is, the long-term memory junctions) were able to function as novel optical memory devices with arithmetic functions [15], whereas the medium τ junctions with short to long-term memory transitions could be used as brain-mimicking image sensors with photomemory behavior [17].

Figure 5
figure 5

(a) The photoinduced currents obtained from three junctions at a bias voltage of 10 V over time before and after irradiation with 10 optical pulses. The dotted lines show the results of fitting using the equation G(t)/G0 = exp[− (t/τ)β]. (b) The relaxation time values (τ) of CNW/diamond junctions as a function of the ID/IG ratio (the ratio of the intensity of the D peak to that of the G peak in Raman spectra).

Full size image

Very recently, the authors demonstrated that image recognition is possible using shorter τ (that is, short-term memory) junctions [28]. This was accomplished using the so-called physical reservoir computing (RC), which is a machine learning method based on recurrent neural networks substituted by the complicated dynamics of physical phenomena [29, 30]. Physical RC provides more rapid learning, prediction, and recognition with lower calculation costs than is the case for numerical calculations using a normal neural network, and is expected to be used in next-generation edge computing devices. In the previous work by the authors, simple images of the numerals 0 through 9 composed of 20 binary bits could be recognized by physical RC using the special photoconducting properties of the CNW/diamond junctions. In the research reported herein, the recognition of more complicated numeral images taken from the MNIST (Modified National Institute of Standards and Technology database) hand-written digital database is demonstrated using the CNW/diamond junctions having shorter τ.

Image recognition of the hand-written numeral images was performed during trials with the junctions having a τ value of 1.9 s. In these experiments, numerals such as “7” were presented as images generated using 256 gradations of gray in conjunction with a 28 × 28 grid (upper left panel in Fig. 6) and were recognized. For simplification, four edges composed of 2 × 28 grids within the 28 × 28 images were cut out to reduce the image size to 24 × 24 (upper middle panel). Following this, these 24 × 24 images were contracted to generate 8 × 8 images by taking the median of the 3 × 3 units in each area (upper right panel). Following these procedures, the 8 × 8 images of the numerals were changed to 8 × 8 binary patterns. These patterns were subsequently decomposed into individual rows, each consisting of eight sequential pixels (such as 00000000 or 01110000) and optical pulses corresponding to each row were imparted to the CNW/diamond junctions. As an example, in the case of the second row for the number “7,” the junction was irradiated with a pulse sequence in which the radiation was off (corresponding to 0), on (1), on (1), on (1), off (0), off (0), off (0), and off (0), employing a optical pulse width of 500 ms and a frequency of 1.6 Hz, as shown in the bottom left part of Fig. 6. The photocurrent data generated in response to these sequential pulses were recorded and eight current values (I1 to I8) were obtained. It should be noted that all the photocurrent data were automatically collected using programs written in the LabVIEW software package. The eight current values generated by each junction varied in a complex manner because of the short-term memory characteristics of these devices. Specifically, I1 through I8 were different for each numeral and so the recognition of each numeral required the raw data to be transferred into a simple, two-layered neural network constructed using LabVIEW. One-third of the 185 images of numerals were used for learning procedures to determine the weights to apply in the neural network, while the others were employed to evaluate recognition accuracy.

Figure 6
figure 6

Schematic of the image recognition procedures for the numeral “7” based on the use of the CNW/diamond junction having the shorter τ of 1.9 s.

Full size image

Figure 7 shows typical I1 to I8 values for each numeral (0 to 9). The length of each horizontal bar corresponds to the magnitude of the current, meaning that a longer bar reflects a higher photoconducting current. Eight different current values were obtained for each numeral, from which the numeral could be identified. As a result, a good recognition accuracy on the order of 80% was obtained, as shown in Fig. 8. This value was lower than the accuracy of approximately 92% obtained in prior work for the recognition of simpler numeral images composed of 20 bit patterns [28]. The loss of accuracy in the present study is thought to have occurred because the hand-written images of numerals lost many of their features during contraction from 28 × 28 images to an 8 × 8 format. In addition, some of the images were found to be difficult to classify even for a human. The loss of the image features could be avoided by instead using direct recognition of the hand-written 28 × 28 images, and image recognition without the contraction procedure will be assessed by our group in future work. In any case, these results indicate that CNW/diamond junctions have potential applications as novel neuromorphic visual information processing devices.

Figure 7
figure 7

Typical results for eight photo-induced currents (I1 to I8) corresponding to the numerals 0 to 9 generated by CNW/diamond junctions. The length of each bar indicates the magnitude of the current.

Full size image
Figure 8
figure 8

Confusion map for the image recognition of the simple numerals 0 through 9 based on the use of the CNW/diamond junctions. Dark and light green color corresponds to higher and lower value, respectively. Correct answers appear as a diagonal line going down to the right.

Full size image

Conclusion

This study demonstrated that diamond/graphene (C sp3-sp2) junctions can be used as novel neuromorphic devices with properties, such as image recognition that mimic brain functions, based on physical RC. The junctions produced in this work exhibited different photomemory times depending on their interfacial states, which in turn were controlled by adjusting the CVD growth conditions. High-resolution TEM analyses indicated that the diamond/graphene interfaces were primarily composed of graphene (001) planes situated parallel to diamond (111) surfaces. The complex interface structures involving unsaturated sp3 bonds within the diamond layers are believed to be closely related to the special optoelectronic properties of these junctions. Using junctions having shorter τ, MNIST hand-written numeral images were recognized with a suitable accuracy of approximately 80% by employing the physical RC method. This work is likely to lead to the eventual realization of high-performance brain-mimicking image processing systems comprising diamond/graphene (C sp3-sp2) heterointerfaces.