Linear and Symmetric Artificial Synapses Driven by Hydrogen Bonding for Accurate and Reliable Neuromorphic Computing

Abstract

Neuromorphic computing addresses the von Neumann bottleneck by integrating memory and processing to emulate synaptic behavior. Artificial synapses enable this functionality through analog conductance modulation, low-power operation, and nanoscale integration. Halide perovskites with high ionic mobilities and solution processabilities have emerged as promising materials for such devices; however, inherent stochastic ion migration and thermal instability lead to asymmetric and nonlinear characteristics, ultimately impairing their learning and inference capabilities. To overcome these limitations, this study introduces a polyvinyl alcohol (PVA)-based hydrogen-bonding interface engineering strategy to stabilize CsPbI3 artificial synapses. Density functional theory calculations and experimental analyses indicate that the hydroxyl groups in PVA form robust O─H···I− bonds with surface iodides, promoting vertical lattice ordering. This suppresses grain boundary defects and enables directional ion migration, resulting in extremely linear and symmetric optoelectronic conductance modulation (αp = 0.004, αd = 0.020), over eight fold reduction in interfacial trap density, and high-temperature retention (>104 s). When integrated into a neural network, artificial synapses show large-scale image classification accuracy within 1.62% of the theoretical limit. The proposed strategy provides a scalable pathway toward overcoming the existing limitations of artificial synapses, exhibiting high potential for application in edge AI, autonomous systems, and material-based cognitive modeling.