Echo-state networks based on voltage-controlled magnetic tunnel junctions
Phys. Rev. Applied 25, 064015 – Published 3 June, 2026
DOI: https://doi.org/10.1103/qm64-x5g8
Abstract
Spintronic devices, such as magnetic tunnel junctions (MTJs), offer nonvolatility and ultralow power operation, making them promising candidates for building physical reservoirs in neuromorphic computing. However, realizing low-power and hardware-integrable reservoir computing (RC) systems remains a major challenge. In this work, we construct an echo-state network based on voltage-controlled stochastic dynamics of ten distinct MTJs. These devices exhibit an S-shaped switching probability as a function of the driving-voltage pulse width, thereby offering the essential nonlinearity, stochasticity, and tunability required for RC. Moreover, by adopting a time-division-multiplexing strategy that integrates outputs from all ten distinct devices, we implement a reservoir with enhanced dynamical diversity. The resulting system achieves an accuracy of over 96% on the Modified National Institute of Standards and Technology (MNIST) task and reduces the normalized root-mean-square error (NRMSE) of the nonlinear autoregressive moving average (NARMA-10) task by about 35% compared with the single-device case. In addition, the MTJs exhibit a 50% switching time of about 660 ps and the energy consumption per switching event is estimated to be about 17.6 fJ at 2 V and 150 kΩ. This study demonstrates the feasibility of using voltage-pulse-controlled MTJs for RC and provides a practical pathway toward hardware-based neural network implementations.
