Masaharu Yonezawa1,2,Hisashi Shima2,Takuma Matsuo1,2,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., LTD5
Masaharu Yonezawa1,2,Hisashi Shima2,Takuma Matsuo1,2,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., LTD5
Edge computing is attracting considerable attention because the real-time secure and low-power information processing (IP) is expected. Especially for the IoT devices operated under the limited power-supply condition, the IP power reduction becomes crucial in addition to realizing higher performance. Therefore, a machine learning algorithm called reservoir computing (RC), which is a special type of the recurrent neural networks, is intensively investigated as a breakthrough to satisfy those requirements. The network structure of RC is quite simple, consisting of input, reservoir, and output layers. Because only the weights between the reservoir and output layers are updated in RC, the calculation cost reduction is expected. To implement RC by actual devices, physical reservoir devices (PRDs), in which actual physical phenomena plays a role of reservoir, are essential for edge computing and PRDs.<br/>We have developed PRD using Faradaic currents (FCs) generated by the electrochemical reactions (ERs) in metal ion-doped ionic liquids (IL), which possess advantages such as a high design flexibility of material properties and an immunity to the electrolysis <sup>[1]</sup>. Because the FCs generated by ERs are much smaller than the operating currents of solid-state devices, IL-based PRD is suitable for low-power IP <sup>[2]</sup>. In this study, we focused on the water contained in IL. According to the previous report, about 10 mol% of water can be dissolved even in hydrophobic ILs such as [Bmim][Tf<sub>2</sub>N] <sup>[3]</sup>. Also, it has been reported that the FC observed during the cyclic voltammetry for the dehydrated, low water content IL decreased below one-fourth of that for the high water content IL <sup>[4]</sup>. Therefore, further lowering of the operating power in the IL-based PRD is reasonably expected by reducing water content in IL. Herein, we evaluated the influence of water in IL on the operating power and IP performance of IL-based PRDs by comparing the electrical properties measured in the synthesized dry air (SDA) after vacuum dehydration and those measured in moisture-containing air (MCA).<br/>The present PRD had a pair of input and output terminals made of a Pt thin film. To control the location of the ER, those terminals were covered by the insulating SiO<sub>2</sub> layer excepting the 100 μm x 100 μm sized square-shaped regions on each terminal. A droplet of 0.4 M Cu ion-doped [Bmim][Tf<sub>2</sub>N] was placed between the terminals to form the Pt/IL interface inside the above-mentioned square-shaped regions. Time series data (TSD) consisting of "0" and "1" was input into the device as a triangular-shaped voltage pulse (TVP) stream, and the output current were measured. The signs of the TVP for "0" and "1" were defined to be negative and positive, respectively. TSDs consisting of alternately and randomly aligned “0” and “1” were used to investigate the basic electrical properties and the short-term memory (STM) characteristics of the PRD, respectively. In SDA, as expected, the peak current values during the device operation decreased below about one-fifth of those in MCA. Moreover, the memory capacity, which is a quantitative measure for the STM characteristics, increased by more than 40 %. The input-signal-dependence of output current waveforms became more prominent in SDA thanks to decrease in cycle-to-cycle current dispersion due to dehydration and consequently the memory capacity increased.<br/>We successfully demonstrated both the IP power reduction and IP performance improvement of IL-based PRD in SDA. In addition, the difference in the output current waveforms between SDA and MCA suggests that the different ERs occur in the device depending on the water amount in IL. The ERs optimized for the IP task to be executed are crucial for improving the physical RC accuracy.<br/>[1] H. Sato,<i> et al</i>., <i>Front. Nanotechnol</i>. <b>3</b>, 660563 (2021).<br/>[2] Y. Zhong, <i>et al</i>., <i>Nat. Commun.</i> <b>12</b>, 1 (2021).<br/>[3] F. Di Francesco, <i>et al</i>., <i>Green Chem.</i> <b>13</b>, 1712 (2011).<br/>[4] S. Caporali, <i>et al</i>., <i>Surf. Coat. Technol</i>. <b>264</b>, 23 (2015).