Path-Decoupled Cation-Eutaxy III-V van der Waals Memristive Semiconductors for Mitigating the Neuromorphic Accuracy-Energy Trade-off.

Abstract

Transistor-based computing faces a fundamental energy-resolution trade-off: lowering the conductance reduces the programming energy (Eprog) but simultaneously narrows the dynamic range (G_max/G_min) required for multilevel state discrimination. A memristor offers analog programmability, but it suffers from the same limitation because reducing the conductance decreases G_max/G_min, degrading the learning accuracy. Here, a path-decoupled III-V van der Waals (vdW) memtransistor overcomes this constraint via the spatial separation of the ionic and electronic transport pathways. Using H_xK_1-xGaSb₂, K⁺ vacancies confined to the vdW gap serve as mobile ionic species, while holes conduct within the covalently bonded [GaSb₂] layers. This decoupling yields a high K⁺ diffusivity and enables memristive switching at markedly reduced voltages. The memristive window G_max/G_min-which is set by ionic motion-remains invariant under gate modulation, whereas E_prog decreases via electrostatic control of the channel conductance. Consequently, the synaptic plasticity and neuromorphic inference maintain a high accuracy (>80%), while E_prog is reduced by more than an order of magnitude. The results establish ionic-electronic path decoupling as a general strategy for breaking the accuracy-energy trade-off in emerging neuromorphic hardware and position III-V vdW materials, which are promising candidates for application in low-energy, artificial intelligence accelerators.