First Demonstration of Ge Ferroelectric Nanowire FET as Synaptic Device for Online Learning in Neural Network with High Number of Conductance State and Gmax/Gmin

In this paper, optimum weight update scheme for improved linearity and asymmetry of channel conductance potentiation and depression in a Germanium ferroelectric (FE) nanowire FET (NWFET) was experimentally demonstrated and simulated for the first time. It was found that −5 V, 320 pulses and +5V, 256 pulses both with 50 ns pulse width were the optimum pulsing conditions for potentiation and depression process, respectively. With the optimized scheme, non-linearity for potentiation and depression were extracted to be $\alpha_{p}=1.22$ and $\alpha_{d}=-1.75$, respectively resulting in asymmetry $(\vert \alpha_{p}-\alpha_{d}\vert)$ of 2.97 based on models embedded in MLP simulator and NeuroSim [1]. $\mathrm{G}_{\max}/\mathrm{G}_{\min}$ ratio (few hundreds) and number of conductance states (>256) are both very large. 9 alternating consecutive conductance updates (potentiation followed by depression) were executed to observe variability in conductance profiles. Multilayer perceptron neural network was simulated over 1 million MNIST images with extracted experimental parameters which yielded in online learning accuracy of ∼ 88%.