Reservoir computing based on electric-double-layer coupled InGaZnO artificial synapse

Physical reservoir computing (PRC) is thought to be a potential low training-cost temporal processing platform, which has been explored by the nonlinear and volatile dynamics of materials. An electric-double-layer (EDL) formed at the interface between a semiconductor and an electrolyte provided a great potential for building high energy-efficiency PRC. In this Letter, EDL coupled indium-gallium-zinc-oxide (IGZO) artificial synapses are used to implement reservoir computing (RC). Rich reservoir states can be obtained based the ionic relaxation-based time multiplexing mask process. Such an IGZO-based RC device exhibits nonlinearity, fade memory properties, and a low average power of similar to 9.3 nW, well matching the requirement of a high energy-efficiency RC system. Recognition of handwritten digit and spoken-digit signals is simulated with an energy consumption per reservoir state of similar to 1.9 nJ, and maximum accuracy of 90.86% and 100% can be achieved, respectively. Our results show a great potential of exploiting such EDL coupling for realizing a physical reservoir that would underlie a next-generation machine learning platform with a lightweight hardware structure.

Automated summary

Physical reservoir computing (PRC) is a potential low-cost temporal processing platform, utilizing the nonlinear and volatile dynamics of materials. An electric-double-layer (EDL) formed at the interface between a semiconductor and an electrolyte is used to build a high energy-efficiency PRC. In this study, EDL coupled indium-gallium-zinc-oxide (IGZO) artificial synapses are used to implement reservoir computing (RC), which exhibits nonlinearity, fade memory properties, and low power consumption, making it suitable for high energy-efficient RC systems. The results show the potential of using EDL coupling for a lightweight hardware physical reservoir that can underlie a next-generation machine learning platform.