Reconfigurable Artificial Synapse Based on Ambipolar Floating Gate Memory.

Artificial synapse networks capable of massively parallel computing and mimicking biological neural networks can potentially improve the processing efficiency of existing information technologies. Semiconductor devices functioning as excitatory and inhibitory synapses are crucial for developing intelligence systems, such as traffic control systems. However, achieving reconfigurability between two working modes (inhibitory and excitatory) and bilingual synaptic behavior in a single transistor remains challenging. This study successfully mimics a bilingual synaptic response using an artificial synapse based on an ambipolar floating gate memory comprising tungsten selenide (WSe 2 )/hexagonal boron nitride (h-BN)/ molybdenum telluride (MoTe 2 ). In this WSe 2 /h-BN/MoTe 2 structure, ambipolar semiconductors WSe 2 and MoTe 2 are inserted as channel and floating gates, respectively, and h-BN serves as the tunneling barrier layer. Using either positive or negative pulse amplitude modulations at the control gate, this device with bipolar channel conduction produced eight distinct resistance states. Based on this, we experimentally projected that we could achieve 490 memory states (210 hole-resistance states + 280 electron-resistance states). Using the bipolar charge transport and multistorage states of WSe 2 /h-BN/MoTe 2 floating gate memory, we mimicked reconfigurable excitatory and inhibitory synaptic plasticity in a single device. Furthermore, the convolution neural network formed by these synaptic devices can recognize handwritten digits with an accuracy of >92%. This study identifies the unique properties of heterostructure devices based on two-dimensional materials as well as predicts their applicability in advanced recognition of neuromorphic computing.