Charge-Selective 2D Heterointerface-Driven Multifunctional\nFloating Gate Memory for In Situ Sensing-Memory-Computing

Flash\nmemory, dominating data storage due to its substantial storage\ndensity and cost efficiency, faces limitations such as slow response,\nhigh operating voltages, absence of optoelectronic response, etc.,\nhindering the development of sensing-memory-computing capability.\nHere, we present an ultrathin platinum disulfide (PtS<sub>2</sub>)/hexagonal\nboron nitride (hBN)/multilayer graphene (MLG) van der Waals heterojunction\nwith atomically sharp interfaces, achieving selective charge tunneling\nbehavior and demonstrating ultrafast operations, a high on/off ratio\n(10<sup>8</sup>), extremely low operating voltage, robust endurance\n(10<sup>5</sup> cycles), and retention exceeding 10 years. Additionally,\nwe achieve highly linear synaptic potentiation and depression, and\nobserve the reversibly gate-tunable transitions between positive and\nnegative photoconductivity. Furthermore, we employed the VGG11 neural\nnetwork for in situ trained in-sensor-memory-computing to classify\nthe CIFAR-10 data set, pushing accuracy levels comparable to pure\ndigital systems. This work could pave the way for seamlessly integrated\nsensing, memory, and computing capabilities for diverse edge computing.