Reservoir Computing with Neuro-memristive Nanowire Networks Kaiwei Fu * , Ruomin Zhu * , Alon Loeffler * , Joel Hochstetter * , Adrian Diaz-Alvarez † , Adam Stieg †‡ , James Gimzewski †‡ , Tomonobu Nakayama ‡* and Zdenka Kuncic *‡ , * School of Physics and Sydney Nano Institute, University of Sydney, Sydney, NSW 2006, Australia Email: zdenka.kuncic@sydney.edu.au † International Centre for Materials Nanoarchitectonics, National Institute for Materials Science, Tsukuba, Japan ‡ California NanoSystems Institute, University of California at Los Angeles, California USA Abstract—We present simulations based on a model of self– assembled nanowire networks with memristive junctions and neural network–like topology. We analyze the dynamical volt- age distribution in response to an applied bias and explain the network conductance fluctuations observed in our previous experimental studies. We then demonstrate the potential of neuromorphic nanowire networks as a physical reservoir by performing benchmark reservoir computing tasks. The tasks include sine wave nonlinear transformation, sine wave auto– generation and forecasting the Mackey–Glass chaotic time series. Index Terms—Neuromorphic systems, Memristive systems, Nanowire networks, Reservoir computing I. I NTRODUCTION Artificial neural networks are computational models using algorithms that are loosely based on biological neural networks [1]. Learning from data is the most time and power consuming part of training such models and memory bandwidth is also an important practical consideration. These challenges have motivated further development of alternate approaches based on Reservoir Computing (RC). Originating from echo state networks and liquid state machines, which were proposed as a means to reduce the training cost of recurrent neural network models [2], RC leverages nonlinear dynamical properties of a high-dimensional reservoir to process information [3]. Impor- tantly, the reservoir itself is not trained, but rather the readout is trained using relatively straightforward methods such as linear regression or classification. RC has been most successfully demonstrated for computa- tional tasks in the temporal domain, such as wave generation and time series prediction, including the well-known bench- marking task Mackey–Glass chaotic time series prediction [4]– [6]. Physical reservoir computing is based on the concept that any physical dynamical system has the potential to serve as a reservoir if it meets several requirements. Tanaka et al. [7] list a number of such requirements, including high dimensionality, non-linearity and fading memory. A number of different physical reservoir models fulfil these requirements, including in particular ones based on analog circuits [8] and memristors [9]. Memristor-based physical RC is attractive because of the synapse-like dynamical electrical switching properties of memristive devices [9]–[13]. This has led to the development of a class of neuromorphic systems and circuits in which memristive RC has been successfully implemented, as evidenced by the ability to perform benchmark learning tasks such as hand-written digit and speech pattern recognition, as well as the Mackey-Glass forecasting task [14]–[16]. Nanowire networks [17]–[24] represent another class of neuro-memristive systems with an additional neuromorphic property: their neural network-like topology, which arises from their self–assembly, analogous to biological neural net- works [25]. Inorganic nanowires comprised of silver readily self–assemble into a complex network, forming two-terminal memristive junctions where nanowires intersect. Memristive switching occurs at the junctions as a result of the formation of a conductive Ag filament above a voltage threshold [19]. Previous experimental and simulation studies based on Ag- Ag 2 S-Ag nanowire networks developed by Stieg, Gimzewski and colleagues demonstrated their potential for physical RC through properties including higher harmonic generation, re- current dynamics and waveform transformation [17], [19]– [21]. While the synthesis of those particular nanowire net- works was aided by a pre-patterned substrate, the resulting network topology was sufficiently complex to observe emer- gent nonlinear (i.e. power-law) dynamics at the network level. More recently, we showed that self-assembled polymer (PVP) coated Ag nanowire networks exhibit similar emer- gent dynamical properties as Ag-Ag 2 S-Ag networks, even though their memristive junctions differ [24]. Training in hardware demonstrated these Ag-PVP-Ag nanowire networks can associatively learn spatial patterns by recalling previously established current pathways [26]. In addition to meeting the requirements of high dimensionality, non-linearity and fading memory, this suggests that physical RC may also be imple- mented on our Ag-PVP-Ag nanowire networks. In this study, we present simulations based on a model of our experimental self-assembled Ag-PVP-Ag nanowire networks. Simulations of these networks allow us to modify various parameters at the junction level, which is impossible to do experimentally. We first explore the switch junction and network dynamics under an applied bias. Following this, we implement several reservoir computing tasks, including nonlinear wave transformation, sine wave generation and Mackey–Glass signal prediction. 978-1-7281-6926-2/20/$31.00 ©2020 IEEE