Reconfigurable High‐Performance Memristors Based on Few‐Layer High‐κ Dielectric Bi ₂ SeO ₅ for Neuromorphic Computing

Abstract Memristors are pivotal for energy‐efficient artificial intelligence (AI) hardware, potentially eliminating the von Neumann bottleneck by in‐memory realizations of synaptic operations. However, the dynamic requirements of neuromorphic computing on specific electronic devices pose reliability and universality challenges, limiting progress toward more widely applicable computing platforms. Here, a 2D high‐κ dielectric‐based memristor with the desired reconfigurable resistive switching behavior is successfully demonstrated. Utilizing a few layered Bi 2 SeO 5 possessing excellent electrical insulation properties as the switching medium, the device features a low operating voltage (≈0.5 V), low operation current (10 pA), long memory retention (>10 3 s), large switching window (≈10 8 ), steep slope (<1 mV dec −1 ), fast switching speed (40 ns), and low energy dissipation (≈1 pJ). The switching characteristics between volatile and non‐volatile memory can be achieved on demand by regulating compliance currents, offering the possibility of implementing multiple neural computational primitives. A simulated convolutional neural network (CNN) based on long‐term potentiation/depression (LTP/D) achieves 85% accuracy in complex image recognition. Furthermore, MNIST and fashion‐MNIST recognition with built reservoir computing (RC) utilizing volatile behaviors reach 97% and 85% accuracy, respectively. This work opens new opportunities for 2D high‐κ dielectrics in next‐generation AI hardware with enhanced energy efficiency and computational versatility.