Open AccessSpin‐Torque Memristors Based on Perpendicular Magnetic Tunnel Junctions for Neuromorphic Computing
Author(s) -
Zhang Xueying,
Cai Wenlong,
Wang Mengxing,
Pan Biao,
Cao Kaihua,
Guo Maosen,
Zhang Tianrui,
Cheng Houyi,
Li Shaoxin,
Zhu Daoqian,
Wang Lin,
Shi Fazhan,
Du Jiangfeng,
Zhao Weisheng
Publication year - 2021
Publication title -
advanced science
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 5.388
H-Index - 100
ISSN - 2198-3844
DOI - 10.1002/advs.202004645
Subject(s) - neuromorphic engineering , memristor , condensed matter physics , tunnel magnetoresistance , torque , perpendicular , spin (aerodynamics) , materials science , physics , electrical engineering , computer science , engineering , ferromagnetism , mechanical engineering , artificial neural network , quantum mechanics , mathematics , artificial intelligence , geometry
Spin‐torque memristors are proposed in 2009, and can provide fast, low‐power, and infinite memristive behavior for neuromorphic computing and large‐density non‐volatile memory. However, the strict requirements of combining high magnetoresistance, stable domain wall pinning and current‐induced switching in a single device pose difficulties in physical implementation. Here, a nanoscale spin‐torque memristor based on a perpendicular‐anisotropy magnetic tunnel junction with a CoFeB/W/CoFeB composite free layer structure is experimentally demonstrated. Its tunneling magnetoresistance is higher than 200%, and memristive behavior can be realized by spin‐transfer torque switching. Memristive states are retained by strong domain wall pinning effects in the free layer. Experiments and simulations suggest that nanoscale vertical chiral spin textures can form around clusters of W atoms under the combined effect of opposite Dzyaloshinskii–Moriya interactions and the Ruderman–Kittel–Kasuya–Yosida interaction between the two CoFeB free layers. Energy fluctuation caused by these textures may be the main reason for the strong pinning effect. With the experimentally demonstrated memristive behavior and spike‐timing‐dependent plasticity, a spiking neural network to perform handwritten pattern recognition in an unsupervised manner is simulated. Due to advantages such as long endurance and high speed, the spin‐torque memristors are competitive in the future applications for neuromorphic computing.