Damage-free van der Waals metal/NbOx-NbSe2 integration for reliable and flexible memristor in neuromorphic computing.

Ultrathin native oxides formed on two-dimensional (2D) metallic transition metal dichalcogenides (MTMDs), such as NbSe2, have recently emerged as promising active switching layers for memristive devices. However, these oxides are highly susceptible to surface damage during conventional top-electrode deposition processes, which leads to interfacial disorder and degraded switching reliability. Here, we demonstrate a highly reliable and flexible NbOx-NbSe2 memristor utilizing a damage-free van der Waals (vdW) metal transfer technique. By gently laminating a prefabricated gold (Au) electrode onto the NbOx-NbSe2, a pristine and chemically undisturbed metal/oxide interface is achieved. While metallization-induced barrier damage causes a large OFF-state direct tunneling (DT) current in the evaporated-metal memristor (Ev-MEM), preserving the pristine oxide interface in the transferred-metal memristor (T-MEM) effectively suppresses this OFF-state DT current. Consequently, the T-MEM exhibits an ON/OFF ratio of 105-a 1,000 times improvement over the Ev-MEM (102)-along with a stable retention time of > 104 s and an endurance exceeding 103 cycles. Furthermore, the T-MEM exhibits exceptional synaptic linearity (β = 0.6) and a full dynamic range (0-1), in contrast to the Ev-MEM, which suffers from poor linearity (β = 3.1) and a limited dynamic range (0.2-1). As a result, a hardware-aware convolutional neural network (CNN) simulation shows a near-ideal MNIST image recognition accuracy of 98.1% for the T-MEM based system, significantly outperforming the Ev-MEM counterpart (75.8%). In addition, capitalizing on the inherent flexibility of the 2D NbSe2 and the vdW interface, devices fabricated on a polymer substrate demonstrate remarkable mechanical robustness, maintaining stable switching characteristics over 200 bending cycles at a 30° angle. Our findings establish that damage-free vdW metal integration is a crucial prerequisite for developing reliable, highly accurate, and flexible neuromorphic hardware.