Energy efficient and robust reservoir computing system using ultrathin (3.5 nm) ferroelectric tunneling junctions for temporal data learning

Reservoir computing (RC) can compute temporal data with low training cost. To enhance data processing capability, high dimensionality of reservoir is required, which poses a significant challenge on RC hardware implementation using Si-friendly devices. In this work, for the first time, we use ultra-thin (3.5 nm) ferroelectric tunneling junctions (FTJs) with transient depolarization property as physical nonlinear virtual nodes to address this challenge. By constructing an FTJ-based dynamic reservoir and combining it with RRAM-based binarized readout layer, the high energy efficiency (35 pJ), processing speed (500 ns), and recognition accuracy (92.3%) have been demonstrated for digital sequence classification.