Development of a physical reservoir that operates by the diffusion of Cu cations

Abstract

We developed a physical reservoir using Cu2S and Cu-doped Ta2O5 as a material of a reservoir layer, in both of which Cu cations contribute to the reservoir operation. The reservoirs showed nonlinearity and short-term memory required as reservoirs. The memory capacity becomes maximum with the input frequency at around 104 Hz. The t-distributed stochastic neighbor embedding analysis revealed that a Cu2S reservoir can classify input of five bit pulse trains, and a Cu-doped Ta2O5 reservoir can classify input of six bit pulse trains. These are longer than four bit pulse trains that a Ag2S island network reservoir achieved in our previous study. Using the superior performance, NARMA task was also carried out.