Expanding Dynamic Range of Ionic Liquid Based Physical Reservoirs by Utilizing High Molecular Design Flexibility

In recent years, there has been a growing focus on reservoir computing (RC) due to its adaptability to edge computing. RC is composed of input, reservoir, and output layers, and is characterized by learning only in the readout part [1]. Furthermore, RC in which the reservoir is replaced with actual physical system is referred to as physical RC (PRC). Accordingly, PRC allows for physical implementation of the reservoir layer and is suitable for fast and low-load learning for time-series data processing. However, the time scale of input signal that can process is limited depending on the relaxation time peculiar to the selected physical system. We focus on ionic liquids (IL) that offer high molecular design flexibility, enabling fine- and wide-tuning of dynamics time scales.<br/>In this study, we showed that the time constant of reservoir could be tuned by employing IL as a physical reservoir and controlling the dielectric relaxation time. In addition, we successfully expanded the dynamic range of PRC for short-term memory (STM) tasks [2] by mixing ILs or parallelizing IL reservoirs with different dielectric relaxation times.<br/>IL reservoirs with EL/IL/EL structure were formed by supplying IL onto a gap between a pair of gold electrodes (ELs) deposited on a SiO2 substrate. We employed four types of ILs with different cations of [C_nmim⁺] (n = 2, 4, 6, 8) each while fixing the anion to [Tf₂N^-], in which the alkyl side chain length of the cation systematically becomes long with increasing n. Positive and negative triangular voltage pulses, which were defined respectively as binary data of “1”and “0”, were injected into the structure. The width (<i>t</i>_w) of the pulse was set in the range of 1-10 µs. We conducted STM tasks using the current response to the random binary data input, and estimated the performance of IL reservoirs as a function of <i>t</i>_w.<br/>Impedance measurement revealed that the dielectric relaxation time of [C_nmim⁺][Tf₂N^-] (n = 2, 4, 6, 8) increased with increasing the length of the cation alkyl side chain. Furthermore, it was confirmed that <i>t</i>_w at which each IL reservoir exhibits the best STM performance closely matches the dielectric relaxation time of each IL. This result enables us to optimize the STM performance of IL reservoirs by selecting IL with the dielectric relaxation time that is similar to<i> t</i>_w. However, once the IL is determined, it is difficult to learn input signals with <i>t</i>_w that is far from the dielectric relaxation time of the IL.<br/>Then, we configured an IL reservoir by connecting all the IL reservoirs for n = 2-8 in parallel, which is called a parallel reservoir hereafter. The STM performance of the parallel reservoir was significantly improved in all the range of <i>t</i>_w compared to that of each single IL reservoir. This is thought to be the result that four IL reservoirs with different time constant provide feature vectors which are linearly independent to each other, creating a higher-dimensional feature vector space.<br/>The above results suggest that a single reservoir which includes multiple relaxation time constants can process input signal with high performance over a wide range of <i>t</i>_w. Therefore, we evaluated the STM performance of an IL reservoir in which ILs for n = 2 and 8 were mixed at a ratio of 4:1, which is called a mixed IL reservoir hereafter. Although not as good as the parallel reservoir, the STM performance of the mixed IL reservoir improved over all the rage of <i>t</i>_w compared to those of single IL reservoirs. It is interesting that by mixing two different types of IL, STM performance can be cooperatively improved without interfering with each other's learning.<br/>The present study indicates the feasibility of wide dynamic range PRC with a single physical reservoir device.<br/><br/>[1] H. Jaeger,<i> GMD Technical Report </i> 148, 13 (2001).<br/>[2] H. Jaeger, <i>GMD Report 152</i> (2001).