Temperature-Robust Short-Term Memory Stored at Electrode/Ionic Liquid Interface and Its Applications for Energy-Efficient Information Processing

Abstract

Listen

Translate article

Electrochemical reactions (ERs) accompanied by electrodeposition at electrode/ionic liquid (IL) interface have been adopted to various application such as batteries [1] and metal plating for interconnects [2]. Herein, we demonstrate that the electrode/IL interface can be used for the physical device implementation of the advanced machine learning (ML) model called reservoir computing (RC), which is an edge AI computing-friendly ML model. In this application, the electrode/IL interface extracts the feature from the input data by converting the signal waveform via ER at the interface [3-5]. Our device having the functional electrode/IL interface is called “ionic liquid-based physical reservoir device (IL-PRD)”.Fig. 1(a) is an optical microscope image of the IL-PRD having Pt electrodes, where ERs occur. The external area of the white dotted line in Fig. 1(a) is covered by SiO2. Fig. 1(b) explains the data processing flow. We performed a benchmark task for ML called a short-term-memory (STM) task. The input signal u(TS) at the time step TS is a random time-series of the binary data (0 or 1), where the signals “0” and “1” are input to the IL-PRD as the triangular voltage pulses (TVPs) with the pulse height/width of -3.0 V/500 ms and +3.0 V/500 ms, respectively. By using the output current signal from the IL-PRD, the efficient ML in a simple neural network (NN) can be conducted. In the STM task, the target data of the ML is u(TS-TS delay), where TS delay was a delay time. The ML performance is evaluated by a memory capacity (MC), which was calculated by adding the square of the correlation coefficient between the target data and model output. A linear regression (LR) model was used after increasing the input variable (x i) for LR by a virtual node method [6]. The weight value w i (fitting parameters of LR) was optimized to minimize the least-square error. In the present study, we investigated the temperature dependence of MC using two different ILs, which were 1-Butyl-3-methylimidazolium Bis(trifluoromethanesulfonyl)imide ([C4mim][Tf2N]) containing Cu cations with the concentration C Cu of 0.1 mol/L and 0.4 mol/L. The ML at the moderately elevated temperature is indispensable when considering the high-temperature application such as the life-time prediction of batteries operated at high temperature [1] as well as the signal processing in the highly integrated circuits [7], which is quite thermogenetic.Fig. 2 is an example of the current-voltage (I-V) curves plotted using the randomly input TVPs and corresponding current obtained at 26 ºC using the IL-PRD with C Cu = 0.1 mol/L. The color-coding of the I-V curves in Fig. 2 was conducted according to the input signal history over three time steps. For instance, the figure legend of “001” shown in Fig. 2 indicates that the corresponding current data was obtained when “1” was input after “0” was input two times. As shown in Fig. 2, the shapes of the I-V curve differ depending on the input signal history, indicating that the output signal from the IL-PRD successfully extracted the feature of the input signal history, in other words, the memory about the time series input signal. The values of MC for IL-PRD with C Cu = 0.1 mol/L and 0.4 mol/L are plotted in Fig. 3 as a function of temperature. Compared with the case of C Cu = 0.4 mol/L, the IL-PRD with C Cu = 0.1 mol/L exhibited the better robustness against the temperature change. In our previous study [4], Cu and some Cu compounds are formed by the electrodeposition during the device operation, and Cu is especially soluble in the present IL even if no external voltage is applied. Such Cu dissolution is reasonably considered to be accelerated by the temperature rise. In the IL-PRD with C Cu = 0.1 mol/L, less amount of Cu is intrinsically involved in ERs at the Pt/IL interface compared to that with C Cu = 0.4 mol/L. Consequently, the electrical properties of the IL-PRD with C Cu = 0.1 mol/L and the consequent ML performance became more stable against the temperature change, which is preferred for the RC under temperature change.