Takuma Matsuo1,2,Hisashi Shima2,Masaharu Yonezawa1,2,Yasuhisa Naitoh2,Hiroyuki Akinaga2,Toshiyuki Itoh3,Toshiki Nokami4,Masakazu Kobayashi1,5,Kentaro Kinoshita1
Tokyo University of Science1,National Institute of Advanced Industrial Science and Technology (AIST)2,Toyota Physical and Chemical Research Institute3,Tottori University4,NAGASE & CO., LTD.5
Takuma Matsuo1,2,Hisashi Shima2,Masaharu Yonezawa1,2,Yasuhisa Naitoh2,Hiroyuki Akinaga2,Toshiyuki Itoh3,Toshiki Nokami4,Masakazu Kobayashi1,5,Kentaro Kinoshita1
Tokyo University of Science1,National Institute of Advanced Industrial Science and Technology (AIST)2,Toyota Physical and Chemical Research Institute3,Tottori University4,NAGASE & CO., LTD.5
Physical reservoir computing (PRC), which is a hardware implementation for a unique paradigm of recurrent neural network (RNN) called RC, attracts considerable attention because PRC is of great promise for advanced and low-power information processing (IP) [1]. PRC has a simple layer configuration: an input, reservoir, and output layers. Since the weight update in PRC is carried out only for the weights between the reservoir and output layers, the lower-power IP is strongly expected compared to the conventional RNN. Although various devices have previously been studied as a reservoir layer (i.e., physical reservoir device (PRD)), an appropriate choice of the physical dynamics in PRD is essential in view of the low-power device operation. Electrochemical reactions (ERs) at a liquid/electrode interface, which are known to exhibit the time-varying current response when voltage is applied, becomes a strong candidate because the electrochemical current reduction is expected by decreasing the electrode area [2]. Considering that liquid materials for PRD are required to withstand repeated voltage inputs, ionic liquids (ILs) are quite suitable because they are more resistant to the electrolysis than aqueous solutions. Besides, the ILs can dissolve various metal cations, which influence the electrochemical properties of PRD [3]. Herein, we report the device size dependence of operating current in IL-based PRD and demonstrate IP with the extremely small operating current of about 0.5 uA.<br/>PRDs used in the present study had a square-shaped openings in the insulating SiO<sub>2</sub> layer on a pair of Pt input and output electrode patterns. The side length <i>L</i> of the square was 10, 100, and 300 um. By placing the IL droplet on those openings, the Pt/IL interfaces, which were the site of the ERs, were formed. As the IL material, a solvated IL, Cu(Tf<sub>2</sub>N)<sub>2</sub>:G3=1:1 (Cu-G3), was selected. The IP capability of the IL-based PRD was assessed based on the short-term memory (STM) task accuracy. In the present STM task evaluation, a binary data (0 and 1) stream <i>u</i>(<i>T</i>) was input to the PRD as a triangular shaped voltage pulse (TVP) stream and <i>u</i>(<i>T</i>-<i>T</i><sub>delay</sub>), the input signal <i>T</i><sub>delay</sub> timestep before, was used as a training data at the timestep <i>T</i>. The sign of the TVP height for 1 and 0 was positive and negative, respectively. For updating the weights between the reservoir and output layers, the linear regression was used. The deposits on the Pt electrode formed during the device operation were investigated by X-ray photoelectron spectroscopy (XPS) after the washout of the IL droplet.<br/>The output current from IL-based PRD showed <i>L</i>-dependent properties. Although the Faradaic current peak was observed in all the devices, the peak width became narrower for the smaller value of <i>L</i>. This result indicates that the ERs in IL-based PRD have a response time lag at the interface. Also, the output current decreased almost in proportional to the value of <i>L</i><sup>2</sup>, achieving about 0.5 uA for <i>L</i> = 10 um. The STM task accuracy slightly decreased with <i>L</i> probably due to the following two factors. One is the cycle-to-cycle variation in the net Pt/Cu-G3 interface area during the operation caused by the deposits on the Pt electrode, which causes the fluctuation in the electrochemical current values especially for smaller L. Another is presumably the cycle-to-cycle variation in the deposit compositional ratio. From the XPS results, it was found that Cu and some Cu compounds were formed on the Pt electrode depending on the applied voltage. The variation in compositional ratio for those materials reasonably causes the different output current even for the same input. Further operating power reduction by the device size miniaturization and IP performance boost in IL-based PRD are expected by realizing more repeatable and reliable ERs.<br/>[1] G. Tanaka et al., Neural Netw. 115, 100 (2019).<br/>[2] S. Trasatti et al., Pure & Appl. Chem. 63, 711 (1991).<br/>[3] H. Sato et al., Front. Nanotechnol. 3, 660563 (2021).