Information processing capabilities of two parallel gold-nanoparticle reservoirs dependent on operation temperatures

Abstract We demonstrate the fabrication and performance evaluation of physical reservoir computing (PRC) using a random network of gold nanoparticles (GNPs). We fabricated two random arrays of GNPs with six electrodes on one sample. Electrical measurements for each array and two parallelized arrays were conducted at temperatures of 4.2 K, 77 K, and 294 K. The random connections of GNPs brought varied tunneling resistances, resulting in various output voltages. The computational performance was assessed using the second-order nonlinear autoregression moving average task. The parallelized reservoir configuration achieved the normalized mean square error as small as 0.03 at 294 K. This higher performance was evaluated through information processing capability and was attributed to the increased second-order capacity at 294 K. Although PRC with GNPs was traditionally regarded to rely on the Coulomb-blockade-induced nonlinearity, nonlinear dynamics possibly due to thermal noise and non-uniform tunnel barriers were effective in reservoir calculations even at room temperature.