Perception, Synaptic Plasticity, and Spiking Neuron Function Enabled by a 2D Ferroelectric NbOBr₂ for Neuromorphic Vision.

Abstract

2D ferroelectrics integrate electrically tunable polarization dynamics with anisotropic light-matter interaction, providing a unified platform for co-localized sensing and in-memory computing. However, most neuromorphic visual systems still rely on heterogeneous components that exploit only a single material functionality at a time, thereby increasing system complexity and fabrication costs. Here, we demonstrate a device-algorithm co-design for neuromorphic visual recognition based on multifunctional NbOBr₂. By leveraging its anisotropic broadband photoresponse, robust in-plane ferroelectricity, and integration with multilayer graphene, we realize three bio-inspired core functions: anisotropic photoelectric synaptic preprocessing for retina-like encoding, ferroelectric synaptic weight modulation for in-memory weighted operations, and leaky integrate-and-fire (LIF) neuronal emulation for spike generation. With BDD100K driving scenes converted to drivable maps and encoded into two orthogonal polarization channels (c- and b-axis; 0°/90°), a device-aware spiking neural network achieves a recognition accuracy of up to 94.2%, outperforming a standard SNN baseline (91.8%). These results illustrate how the intrinsic multifunctionality of a 2D ferroelectric can be harnessed to create compact, biologically plausible perception-computation hardware for high-performance neuromorphic vision.