Reservoir Computing Using Vertically Aligned Graphene/Diamond Heterojunctions

Recently, reservoir computing (RC) using various physical phenomena such as laser light, rippling water surface, memristors, etc., which have complicated dynamics, has been paid much attention. Higher speed learning and prediction with lower calculation cost is expected to be realized by implementing the physical phenomena to the reservoir system. However, the research on physical RC had only just started, and physical reservoir systems suitable for high speed and high efficiency computing has been seeking.<br/>In this study, we show that it is possible to perform physical RC by using vertically aligned graphene (VG)/diamond heterojunctions, which are found to show multiple photo-memory [1] and brain-mimic optoelectronic memory functions [2], recently, and recognize simple digits (0~9) by using the junctions for the first time.<br/>Heterostructures of VG and boron-doped semiconducting diamond were grown <i>in situ</i> by using microwave plasma CVD. VG/diamond junctions were fabricated using the heterostructures by a standard photolithographic process and reactive ion etching (RIE). Current-voltage characteristics of the VG/diamond junctions were measured at room temperature (R. T.) in air with or without photo-irradiation by using visible LEDs.<br/>Conducting behaviors of various VG/diamond junctions formed in different growth conditions were examined during and after irradiating 10 optical pulses with positive bias voltage. Conductivity values of the junctions were increased step-by-step in response to each optical pulse and decayed with different relaxation time (τ). From results of Raman spectroscopy and TEM measurements, the relaxation time of the junctions were decreased as the interlayer distances of interfacial graphene layers were increased. It is suggested that the distances were controlled by the volume of sp³ type defects in the layers, and photo-memory functions of VG/diamond junctions were changed depending on the sp³/sp² ratio of the VG layers, which could be controlled by changing the growth temperature and the oxygen ratio during CVD growth. VG/diamond junctions with shorter-term memory functions (τ= 1.9 s) could be fabricated based on these results.<br/>RC using the VG/diamond junctions was performed. First, simple digits (0~9) were shown by binary 5*4 images as shown below. As a note, you can see number “2” by tracing the positions of 1 in the images.<br/>----- (5*4 images (= 20 pixels) of the number “2” shown by 1 or 0 (binary image)) -----<br/>0 1 1 0<br/>1 0 0 1<br/>0 0 1 0<br/>0 1 0 0<br/>1 1 1 1<br/>---------<br/>These images were decomposed into each raw containing 4 sequential pixels, like 0110, 1001, …, and optical pulses correspond to each raw were irradiated to the junctions. For example, in the case for first line of number “2”, the junctions were irradiated by sequential pulses of light-off (0), on (1), on (1), and off (0) because the first line of the images of “2” was shown by “0110”. Each conductivity value of the junctions was memorized after irradiation by each sequential pulse, and five conductivity values were obtained. The conductivity of the junctions was changed in a complex way because of short-term memory functions of them. The five conductivity values are different for each digit, and it is possible to recognize each digit (0~9) after transferring them to the simple, two-layered neural network. In this study, 50 images of simple digits were used to evaluate recognition accuracy. As a result, we have obtained excellent recognition ratio of above 90%. These results indicate that the VG/diamond photomemristors work as physical reservoirs and can be used as novel brain-mimic devices with both photo-memory and computing functions.<br/><br/>Ref.: [1] K. Ueda et al., Appl. Phys. Lett. 117 (2020) 092103, [2] Y. Mizuno, Y. Ito, and K. Ueda, Carbon. 182 (2021) 669. [3] Du et al., Nature. Comm. 8 (2017) 2204.