Ultra‐Low‐Power and Reconfigurable Optoelectronic Memtransistor Based on Vertical Nb‐WSe2/Te Van Der Waals Heterostructure

Why It Matters

Such sub‑attojoule operation could dramatically shrink the power budget of brain‑inspired processors, accelerating the shift away from energy‑hungry von Neumann architectures.

Key Takeaways

• Energy per spike below 1 attojoule.
• Vertical Nb‑WSe2/Te heterostructure enables reconfigurable synapses.
• Device performs optical‑electrical AND/OR logic operations.
• CNN built from memtransistors reaches 92.3% CIFAR‑10 accuracy.
• Maintains 72.8% accuracy under multilevel noise.

Pulse Analysis

Neuromorphic computing seeks to mimic the brain’s parallelism and energy efficiency, but conventional silicon circuits remain constrained by the von Neumann bottleneck. Two‑dimensional materials and van der Waals heterostructures have emerged as promising platforms because they allow atomically sharp interfaces and tunable electronic properties without lattice‑matching constraints. By stacking Nb‑doped WSe₂ with tellurium layers, researchers created a vertical channel that can be modulated both electrically and optically, opening a pathway to truly hybrid synaptic devices.

The standout feature of this memtransistor is its sub‑attojoule energy consumption per optical spike—approximately 1 × 10⁻¹⁸ J—far surpassing the ~10⁻¹⁴ J typical of biological synapses. This efficiency stems from the ultra‑thin active layer and the strong light‑matter interaction in the heterostructure, which together enable precise charge trapping and release with minimal power. Beyond basic plasticity, the device supports programmable AND/OR logic through combined photo‑electronic gating and can emulate associative learning such as Pavlovian conditioning by leveraging wavelength‑selective illumination.

When integrated into a convolutional neural network, the memtransistor array delivers 92.32 % accuracy on the CIFAR‑10 benchmark and retains 72.75 % accuracy under multilevel noise, showcasing robustness comparable to conventional digital accelerators. This performance, combined with the dramatically lower energy envelope, positions the technology as a candidate for next‑generation AI hardware that can operate at the edge or in power‑restricted environments. As the industry pushes toward scalable neuromorphic chips, the Nb‑WSe₂/Te platform may accelerate commercialization by offering a manufacturable, reconfigurable, and ultra‑efficient synaptic element.