A Single Quasi-Vertical van der Waals Heterostructure Memristor as Opto-Reconfigurable Artificial Intelligence Hardware.

Abstract

The explosive advancement of artificial intelligence (AI) has driven vast amounts of unstructured data that directly emerges at sensory interaction interfaces. Memristors with superior computing ability are among the most promising computing blocks. However, conventional vertical Metal/Resistive layer/Metal (V-MRM) memristors tend to generate intense electric fields, suffer from limited sensory response, inflexible, and rigid functional modes, as well as long-standing variability and overshoot issues. Here, we present an unconventional two-terminal quasi-vertical (QV) van der Waals heterostructure (vdWH) memristor as opto-reconfigurable hardware. Unlike the V-MRM memristor with a strong, rigid vertical electric field, this QV-vdWH memristor realizes a significantly softened, tunable electric field, enabling parametric configurable and wavelength-dependent reconfigurable modes across nonvolatile/volatile memory, neurons, synapses, and optical perception. Moreover, the light-memristive switching-vdWH coupling triggers a super-additive synergistic enhancement of memristive performance (> 100-fold) and photoresponse (> 3000-fold). Meanwhile, the long-standing variability and overshoot issues are overcome, yielding ultra-uniform memristive behavior with even a single-valued distribution. Leveraging the opto-reconfigurability, high-accuracy object tracking, and human motion recognition (∼96.8%) are successfully performed. Breaking through the inherent functional limitation of the conventional V-MRM memristive structure, the QV-vdWH memristor provides a flexible, opto-reconfigurable hardware platform for emerging computing paradigms, and AI applications.