Physical Reservoir Computing (PRC) is a technology attracting attention for overcoming the limitations of conventional digital computing. This technology is particularly useful in areas where complex nonlinear problems and real-time processing are required. PRC is a technique using physical systems (e.g., optical, electronic, magnetic, and mechanical systems) as reservoirs and utilizes their dynamic properties for computation. This enables efficient processing of tasks such as speech recognition, image processing, and predictive modeling. We asked top researchers to contribute their recent research to this special issue about the PRC.

A solid electrolyte-based physical reservoir utilizes ions’ diffusion and their redox reactions to achieve nonlinear transformation of input and its short-term memorization. One of the major advantage in using a solid electrolyte is easy fabrication of a recurrent neural network with a large number of ‘nodes’, which enables various functions unique to the material-based physical reservoir such as in-sensor computing.

Artificial intelligence (AI) has rapidly advanced and is being utilized across a wide range of fields by developing artificial neural network (ANN). However, constructing large-scale ANNs requires significant power consumption. In the field of materials and device engineering, “physical reservoir computing (physical RC)” is gaining attention to perform ANN computations with low power consumption. Unlike traditional ANNs, physical RC leverages physical properties for computation, resulting in energy efficiency. Conductive polymers such as polyaniline are expected to exhibit excellent performance in in-materio physical reservoir computing (IMRC) due to their superior environmental stability and reversible doping behavior. This paper introduces about utilizing the mixed polaron-ion conductivity of sulfonated polyaniline (SPAN), evaluating changes in electrical properties with humidity adjustments, and exploring its potential application in IMRC by investigating the correlation between the electrical properties and computational ability.

Magnetic skyrmions are topological spin textures, with great potential for energy-saving next generation devices such as non-volatile memory devices. Recently, the application of skyrmions to neuromorphic devices, including physical reservoir computing devices, has attracted much attention. In this article, we introduce the recent progress in research on skyrmion-based physical reservoir computing, mainly focusing on our results. We recently demonstrated physical reservoir computing using the nonlinear response originating from the magnetic-field induced dynamics of skyrmions formed in ultrathin multilayer films. We succeeded in waveform recognition and handwritten digit recognition. Notably, there is a positive correlation between the recognition accuracy and the number of skyrmions in the devices. In addition to our results, skyrmion-based reservoir computing devices have also been reported by other groups, highlighting the high potential of magnetic skyrmions for physical reservoir computing.

Reservoir computing is a framework of recurrent neural networks that offers the significant advantage of not requiring learning in the intermediate reservoir layer. Leveraging this feature, the concept of “physical reservoir computing,” where computations in the reservoir layer are substituted by dynamical systems, has emerged. This novel approach has formed a broad research field that spans various domains, including not only information science but also mechanical engineering, condensed matter physics, and materials science. In this paper, I focus on physical reservoir computing using nanometer-sized magnets, describing my research achievements and discussing future prospects.

Ferroelectric field-effect transistors (FeFETs) with silicon channels and hafnia-zirconia ferroelectric films have been employed as silicon-friendly physical reservoir in reservoir computing application by leveraging the rich dynamics of polarization-polarization and polarization-charge interactions. FeFET reservoir computing exhibits promising computing performance with low error rate when predicting time-series generated from a nonlinear function and when classifying spoken digits for speech recognition. We investigate several approaches to improve the computing performance of FeFET reservoir computing including the optimization of input and readout schemes, the introduction of antiferroelectric-like materials with more complicated polarization dynamics, and the re-training to compensate the device degradation under electrical stress.

To learn signals with relatively slow time scales arising in living environments, slow dynamics-based physical reservoirs with tunable relaxation time constants are required. As a suitable physical reservoir for this purpose, we propose an ionic liquid-based reservoir. By utilizing the high molecular designability of ionic liquids, we can control the relaxation time constant of the reservoir according to the time scale of the signal and optimize learning accuracy. Furthermore, because of the high ionic conductivity, ionic liquids can provide excellent chemical reaction, allowing us to advantageously use electrochemical reactions. The present study showed that by doping metal ions to the ionic liquid and generating faradaic current due to redox reaction of them, it is possible to improve learning performance for nonlinear tasks.

Nonlinear dynamical systems useful for reservoir computing enhance the physical implementation of computing systems. In this paper we show that electrochemical reactions and their versatility show great potential as reservoirs. The essence of signal processing in such systems is the different levels of ionic current passing through the solution and the detected electrochemical current based on a multidirectional data acquisition system. A basic electrochemical system is also tested to verify the distinction between Faradic and non-Faradic current. Faradic current exhibits a nonlinear effect on the applied voltage and increase the high-dimensional memory capacity.

Ocean plastic collected from the sea around Japan was evaluated using ToF-SIMS with tandem mass spectrometer (MSMS) and machine learning. The expanded polystyrene-like ocean plastic samples were selected and then they were classified into smaller size (less than 7 nm) and larger size (more than 7 nm). The depth profiles of the smaller and larger size expanded polystyrene-like ocean plastics were measured with ToF-SIMS and then some of the PS-related peaks were selected as precursor ions for MSMS to confirm their chemical structures. The MSMS spectra showed the pathways of each PS-related fragment ion. The ToF-SIMS datasets were analyzed using unsupervised machine learning methods such as principal component analysis (PCA) and sparse autoencoder (SAE). These methods showed that there were differences between the smaller and larger plastic samples and between the analysis depths of the depth profiles. These analysis results are useful for the evaluation of ocean plastic samples.