Physical Reservoir Computing (PRC) has recently attracted significant attention as a computational method suitable for the edge AI computing, which requires both the high performance of information processing and energy conservative operation. There are two standard methods for evaluating PRC performance: the short-term memory (STM) task for the memory capacity and the parity check (PC) task for the nonlinear conversion capacity. We have developed PR device utilizing Faradaic currents generated by the redox reaction of metal ions in ionic liquids (ILs) and the impact of metal ions in ILs on the STM characteristics has been evaluated by comparing the RC device using metal-ion doped IL with that using non-doped (pure) IL [1]. In this study, we investigated the effect of Faradaic current on the STM and PC tasks by extracting the Faradaic current from the output signal when the triangular shaped input voltage pulse was applied to the PR device. It was found out that the peak shaped Faradaic current in the output signal improves not only the memory capacity but also the nonlinear conversion capability.A reservoir device with a transverse Pt/SiO2/Pt structure was prepared (Fig. 1) and solvated IL, Cu(Tf2N)2-Glyme(G3)=1:1 [2], was provided between the Pt electrodes as a reaction field where electrochemical reaction of Cu actively takes place. In order to prevent unnecessary copper deposition on the top surface of the Pt electrode, all areas other than the Pt electrode tip was covered with SiO2. As a result, a structure that allows deposition only between the terminals was realized. Furthermore, to prevent the migration of the IL by the application of an electric field, an IL pool surrounded by a resist wall was formed by patterning a spin-coated photoresist AZ5214E with a thickness of 2 µm. Au/Ti (100/10 nm) was deposited as the contact pad. The device characteristics were evaluated by cyclic voltammetry. In addition, a response of the device to the triangular pulses was investigated by using the B1530A WGFMU (waveform generator / fast measurement unit). STM and PC tasks were used to evaluate the time series data processing ability.In the present study, an artificial time-series data consisting of randomly connected binary data (0 and 1) was input to one of the Pt electrodes as the triangular shaped voltage pulse stream, while the other electrode was grounded. The positive and negative voltages were defined as 2bit data, ‘1’ and ‘0’, respectively. As shown in Fig. 2, triangular voltage pulse streams shown by the blue line were input to the reservoir device, and output current shown by red line was observed. Current peak appears when the polarity of the input signal is switched from positive to negative and vice versa (black arrow), but the peak intensity decreases when application of voltage pulses with the same polarity are continuously repeated. In addition, Cu deposition on the Pt electrode was observed. These results indicate that the origin of the peak is the Faradaic current generated by the electroactive species near the electrode. We divided output current data into two parts; the first half (yellow region), the latter half (green region). The first half corresponds to the rising part of the triangular pulse and involves the Faradaic current. On the other hand, the latter half corresponds to the descending part of the triangular pulse and is more featureless compared with the first half. We found that the accuracy of STM task was much higher when the first half was used. This tendency was also confirmed for the PC task. In PRC based on the concept of virtual nodes using a single physical device [3], the output signal complexity is considered to be related to the multidimensional transformation capability, which is one of indispensable property for PRC [4]. The present results suggest that the redox reaction of electroactive species in the IL increases the complexity of the output signal and improves the ability to extract time-series features.We thank Mr. Hiroshi Sato for supporting device fabrication processes. A part of this work was supported by "Nanotechnology Platform Program", Grant Number JPMXPF21NM0006.[1] D. Sato et al., MEMRISYS 2021, 4A-7 (2021).[2] H. Yamaoka et al., Chem. Lett., 46, 1832-1835 (2017).[3] L. Appeltant et al., Nat. Commun., 2, 468 (2011).[4] L. Appeltant et al., Sci. Rep., 4, 3629 (2014). Figure 1
Read full abstract