High Conductance Margin for Efficient Neuromorphic Computing Enabled by Stacking Nonvolatile van der Waals Transistors

High-performance artificial synaptic devices are key building blocks for developing efficient neuromorphic computing systems. However, the nonlinear and asymmetric weight update of existing devices has restricted their practical applications. Herein, floating gate nonvolatile memory (FG NVM) devices based on two-dimensional (2D) ${\mathrm{Hf}\mathrm{S}}_2$/$h$-BN/FG-graphene heterostructures have been designed and investigated as multifunctional NVM and artificial optoelectronic synapses. Benefiting from the FG architecture, the ${\mathrm{Hf}\mathrm{S}}_2$-based NVM device exhibits competitive performances, such as a high on:off ratio ($>{10}^5$), large memory window (approximately 100 V), excellent charge retention ability ($>{10}^4\mathrm{s}$), and robust durability ($>{10}^3$ cycles). Notably, the artificial optoelectronic synapses based on ${\mathrm{Hf}\mathrm{S}}_2$ FG NVM show an impressive large conductance margin and good linearity, owing to the ultrahigh photoresponsivity and photogain of ${\mathrm{Hf}\mathrm{S}}_2$. The energy consumption of per spike in our artificial synapse is as low as 0.2 pJ. Therefore, a high recognition accuracy up to 91.5% of the artificial neural network on the basis of our ${\mathrm{Hf}\mathrm{S}}_2$-based optoelectronic synapse at the system level has been achieved, which is superior to other reported 2D artificial optoelectronic synapses. This work paves the way forward for all 2D material-based memory for developing efficient optogenetics-inspired neuromorphic computing in brain-inspired intelligent systems.

Figure 1

(a) Three-dimensional (3D) schematic diagram of the device composing of ${\mathrm{Hf}\mathrm{S}}_2$/$h$-BN/graphene layers. (b) Optical image of the stacked ${\mathrm{Hf}\mathrm{S}}_2$, $h$-BN, and graphene flakes in FET with electrical contacts on $\mathrm{Si}$/$\mathrm{Si}\mathrm{O}$${}_2$ substrate. (c) Surface topography AFM image of the ${\mathrm{Hf}\mathrm{S}}_2$ flake. Note that the thickness of the blue line is obtained from the line profile. (d) Raman spectra of the ${\mathrm{Hf}\mathrm{S}}_2$, $h$-BN, graphene, and ${\mathrm{Hf}\mathrm{S}}_2$/$h$-BN/graphene stacked heterostructure. (e) Output characteristic curves showing good Ohmic behavior for different control gate bias. (f) Transfer characteristic curves with a large memory window at different $V_{\mathrm{DS}}$ from 0.02 to 1 V.

Figure 2

(a) Transfer characteristics of the ${\mathrm{Hf}\mathrm{S}}_2$ device on silicon. The hysteresis windows between the forward and reverse transfer characteristic curves are pretty small. (b) Transfer characteristic curves of the ${\mathrm{Hf}\mathrm{S}}_2$-based NVM device with increasing hysteresis for different control gate-voltage sweeping ranges. Note that the directions of the hysteresis loops are indicated by the solid arrows. (c) Variation of the threshold voltage shift under various $V_{\mathrm{CG\; max}}$. (d) Memory window extracted from transfer characteristic curves depending on $V_{\mathrm{CG\; max}}$. The inset presents the calculated stored carrier density $n$ as a function of $V_{\mathrm{CG}}$. (e) Retention performance of the ${\mathrm{Hf}\mathrm{S}}_2$-based NVM device after being programmed ($+80$ V, 1 s) and erased ($-80$ V, 1 s), both program and erase states are read with $V_{\mathrm{DS}}=0.1\mathrm{V}$, $V_{\mathrm{CG}}=0\mathrm{V}$. (f) Retention endurance property over 1000 cycles of the memory device. Schematic energy band diagram of nonvolatile characteristic under (g) zero, (h) positive, and (i) negative control gate bias conditions, $\mathrm{\Phi }$ and $\chi$ are the work function and the electron affinity, respectively.

Figure 3

(a) Schematic diagram of a biological synapse and its emulation with a three-terminal ${\mathrm{Hf}\mathrm{S}}_2$-based NVM device. (b) Excitatory postsynaptic current (EPSC) generated by single-input spike with different amplitudes ($-20$, $-30$, and $-40$ V). (c) Inhibitory postsynaptic current (IPSC) triggered by positive spike with different pulse amplitudes (20, 30, and 40 V). Note that the input spike has a fixed width of 50 ms in (b) and (c). (d) EPSC behavior generated by two consecutive input spikes ($-30$ V, 50 ms) with an interval time of 0.5 s. (e) STP shown by the PPF and PPD index as a function of interval time between two successive input spikes ($\pm 30$ V, 50 ms). (f) The dynamic long-term potentiation of PSC stimulated by a serious of input spikes ($-20$ V, 50 ms) with interval time of 1 s. (g) The dynamic long-term depression behavior in PSC under a series of voltage pulses (15 V, 50 ms) with an interval time of 1 s. (h) LTP and LTD synaptic behaviors of the artificial synapse in response to the application of 50 negative voltage pulses ($-20$ V, 100 ms) and 50 positive voltage pulses (15 V, 100 ms) with a period time of 5 s. (i) The extracted $G_{\max }:G_{\min }$ ratio as a function of pulse duration.

Figure 4

(a) Transient photoresponsive characteristics of the NVM device under different light intensities from 2.6 to $245.3{\mathrm{mW/cm}}^2$ for 200 ms with fixed wavelength of 405 nm. (b) Transient photoresponsive characteristics of the NVM induced by light pulses ($\lambda =405\mathrm{nm}$, $P=245.3{\mathrm{mW/cm}}^2$) with different exposure time (1 ms–2 s). (c) Typical photoresponsive characteristics of the NVM generated by a pair of light pulses (405 nm, $245.3{\mathrm{mW/cm}}^2$, $\mathrm{\Delta }\mathrm{t}$=0.8 s). $A_1$ and $A_2$ represent the change of values at the end of the first and the second light pulse, respectively. (d) PPF index plotted as a function of interval time between light-pulse pairs, fitted by exponential curves. (e) The STM-to-LTM transition triggered by increasing the number of input light pulses. (f) Forgetting curve and fitted line using exponential relaxation model. The PSC achieved after 50 consecutive light pulses. (g) The measured “learning-experience” behavior emulated by pulsed light stimuli. Light pulse in (e)–(g): $\lambda =405\mathrm{nm}$, $P=245.3{\mathrm{mW/cm}}^2$; pulse width, 0.2 s; pulse frequency, 0.5 Hz.

Figure 5

(a) LTP characteristic curve measured under light pulse ($\lambda =405\mathrm{nm},P=245.3{\mathrm{mW/cm}}^2$, pulse width 0.2 s, pulse frequency 0.5 Hz.) and LTD characteristic curve obtained under different voltage spike amplitudes from 0.5 to 25 V with a step of 0.5 V. (b) Schematic of a three-layer neural network for recognition tasks. (c) Recognition accuracy evolution with training epoches based on the LTP and LTD characteristic curves measured under light and voltage spikes, respectively. (d) The simulated recognition accuracy as a function of training epoches based on our artificial synapse measured under voltage pulse.