High‑Performance ReS₂ Charge‑Trapping Synaptic Device Achieves 100% Face‑Recognition Accuracy

Abstract

Synaptic devices underpin the next generation of energy‑efficient, high‑speed neuromorphic systems. We have fabricated a charge‑trapping synapse based on anisotropic rhenium disulfide (ReS₂) that faithfully reproduces long‑term potentiation and depression. To validate its suitability for large‑scale neural networks, we trained a three‑layer artificial neural network (ANN) on 120 images from the Yale Face database and tested it on 45 unseen images. By mapping 120 finely tunable conductance states of the ReS₂ device onto the ANN’s >10⁵ weights, the network achieved a perfect 100 % recognition rate, demonstrating the device’s promise for practical neuromorphic applications.

Background

Traditional von Neumann architectures separate computation from memory, creating a bandwidth bottleneck and high power consumption. In contrast, the human brain operates with ≈ 20 W by leveraging a highly interconnected network of ∼10¹¹ neurons and ∼10¹⁵ synapses. Synaptic plasticity—dynamic adjustment of synaptic weight—is the core mechanism that enables learning and memory.

Two‑dimensional (2D) materials, such as graphene, transition‑metal dichalcogenides (TMDCs), and black phosphorus, have emerged as promising channel media for artificial synapses. While symmetric‑lattice TMDCs (MoS₂, WSe₂) dominate the literature, rhenium disulfide (ReS₂) offers a distorted 1T structure with a direct bandgap up to ten layers, simplifying fabrication and enabling robust charge trapping.

Previous synaptic demonstrations have focused on short‑term plasticity, with only a few reports capturing long‑term potentiation/depression and integrating conductance states into ANN calculations. Our work bridges this gap by providing a ReS₂ device that not only emulates LTP/LTD but also delivers 120 stable conductance levels usable as ANN weights.

Methods

The device stack (Figure 1d) comprises a 70‑nm ITO back gate on 200‑nm SiO₂/Si, a 12/4/4‑nm Al₂O₃/ZrO₂/Al₂O₃ charge‑trapping layer grown by ALD, and a mechanically exfoliated ∼3.6‑nm (≈ 5‑layer) ReS₂ channel with Ti/Au source/drain contacts (10/70 nm). The channel length is 1.5 µm, and the ReS₂ crystal orientation along the b‑axis is chosen for its superior electron mobility. Pulsed back‑gate voltages modulate the device, while a constant 0.1‑V drain‑to‑source bias reads the postsynaptic current.

Results and Discussion

Figure 2a shows an on/off ratio >10⁶ at V_ds = 100–700 mV, confirming strong channel control by the 20‑nm ITO gate. The transfer characteristics exhibit a superlinear low‑V_ds region, indicating good Schottky contacts. A 3.5‑V memory window is observed in the I_bg hysteresis (Fig. 2b), reflecting efficient charge trapping in the ZrO₂ layer (Fig. 2c,d).

Excitatory postsynaptic currents (EPSC) and inhibitory postsynaptic currents (IPSC) are triggered by −1 V/10 ms and +1 V/10 ms back‑gate pulses, respectively (Fig. 3a,b). Negative pulses induce LTP by injecting electrons into the ReS₂ channel; positive pulses produce LTD by trapping electrons in ZrO₂. Repeated −2 V pulses with 1‑s intervals produce a monotonic increase in current (Fig. 3c,d), enabling the generation of multiple stable conductance states.

After 120 successive −2 V, 10 ms pulses, the device exhibits 120 linearly spaced conductance levels (Fig. 4a). These levels map directly onto the >10⁵ ANN weight values, as illustrated in the 3‑layer ANN architecture (Fig. 4b): 1024 input neurons, 256 hidden neurons, and 15 output classes. Training on 120 Yale Face images and testing on 45 unseen images, the network reaches 100 % accuracy after 600 training iterations (Fig. 5b). Weight distributions converge to a scattered set of values, which are then matched to the 120 conductance states via a linear transformation C_j = A I_j + B (A = 1.3769 × 10¹⁰, B = −65.784). Substituting these conductance values into the ANN preserves the 100 % recognition rate, confirming the viability of ReS₂ synapses for large‑scale neural networks.

Conclusions

We have demonstrated a high‑k dielectric‑stacked ReS₂ synaptic device that reliably exhibits long‑term potentiation and depression. By applying 120 periodic gate pulses, the device yields 120 distinct, stable conductance states that replace >10⁵ ANN weights, achieving perfect face‑recognition accuracy. These results establish ReS₂ as a compelling platform for neuromorphic hardware.

Availability of Data and Materials

The authors confirm that all data, materials, and protocols are available to readers and fully contained within this article.

Abbreviations

2D:

Two‑dimensional

ALD:

Atomic layer deposition

ANN:

Artificial neural network

LTD:

Long‑term depression

LTP:

Long‑term potentiation