Highly Reliable Bi2O2Se Dendritic Neuron Enabling Spatial-Temporal Signal Processing for Real-World Image Classification

• Published in: ACS nano 2024-12
• Main Authors: Oh, Jungyeop, Ahn, Wonbae, Ham, Ayoung, Lee, Hyeonji, Lee, Sejin, Cha, Jun-Hwe, Seo, Seunghwan, Kang, Kibum, Choi, Sung-Yool
• Format: Article
• Language: English
• ISSN: 1936-086X; 1936-086X
• DOI: 10.1021/acsnano.4c11133

Artificial intelligence (AI) has made significant strides by imitating biological neurons and synapses through simplified models, yet incomplete neuron functionalities can limit performance and energy efficiency in handling complex tasks. Biological neurons process input signals nonlinearly, utilizing dendrites to process spatial-temporal information. This study demonstrates the compact artificial dendrite device employing memristors based on bismuth oxyselenide (Bi2O2Se). Transfer-free Bi2O2Se switching medium is directly grown on the metal-patterned substrates via 350 °C selenization process. The layered Bi2O2Se structure, limiting metal injection, results in reliable dynamic resistive switching with excellent cycle uniformity and exceptional endurance over 2 million cycles. The highly reliable current response of dynamic resistive switching is modeled with respect to the spatial-temporal voltage input. With the Bi2O2Se dendrite device, dendritic neuron model is implemented, and the proposed neural network achieved high recognition rates of 78.3% with the street view house numbers (SVHN) data set.Artificial intelligence (AI) has made significant strides by imitating biological neurons and synapses through simplified models, yet incomplete neuron functionalities can limit performance and energy efficiency in handling complex tasks. Biological neurons process input signals nonlinearly, utilizing dendrites to process spatial-temporal information. This study demonstrates the compact artificial dendrite device employing memristors based on bismuth oxyselenide (Bi2O2Se). Transfer-free Bi2O2Se switching medium is directly grown on the metal-patterned substrates via 350 °C selenization process. The layered Bi2O2Se structure, limiting metal injection, results in reliable dynamic resistive switching with excellent cycle uniformity and exceptional endurance over 2 million cycles. The highly reliable current response of dynamic resistive switching is modeled with respect to the spatial-temporal voltage input. With the Bi2O2Se dendrite device, dendritic neuron model is implemented, and the proposed neural network achieved high recognition rates of 78.3% with the street view house numbers (SVHN) data set.