High-Accuracy Deep Neural Networks Using a Contralateral-Gated Analog Synapse Composed of Ultrathin MoS₂ nFET and Nonvolatile Charge-Trap Memory

• Published in: IEEE electron device letters Vol. 41; no. 11; pp. 1649 - 1652
• Main Authors: Chung, Yun-Yan, Cheng, Chao-Ching, Chou, Yu-Che, Chueh, Wei-Chen, Chung, Wan-Hsuan, Yu, Zhihao, Hung, Terry Yi-Tse, Huang, Lin-Yun, Wang, Shin-Yuan, Teng, Li-Cheng, Chang, Wen-Ho, Li, Lain-Jong, Chien, Chao-Hsin
• Format: Journal Article
• Language: English
• Published: New York IEEE 01-11-2020; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Artificial neural networks; Charge-trap memory; contralateral-gated; Dielectrics; Hafnium oxide; Logic gates; Molybdenum; MoS; Neural networks; nonvolatile; Nonvolatile memory; Synapses; transition metal dichalcogenide (TMD)
• ISSN: 0741-3106; 1558-0563
• DOI: 10.1109/LED.2020.3026931

The development of high-accuracy analog synapse deep neural networks entails devising novel materials and innovative memory structures. We demonstrated an analog synapse with contralateral gates based on a two-dimensional (2D) field-effect transistor and nonvolatile charge-trap memory. Vertical integration of a MoS 2 -channel FET with a charge-trapping layer provided excellent charge controllability and gate-tunable nonvolatile storage. In the proposed contralateral-gate design, the read and write operations were separated to mitigate read disturb degradation. Reducing the MoS 2 channel thickness to the ultrathin scale allowed large threshold voltage shifts and on-resistance (<inline-formula> <tex-math notation="LaTeX">\text{R}_{\text {ON}} </tex-math></inline-formula>) modulations. This vertically integrated MoS 2 synapse device exhibited 55 conductance states, high conductance max-min ratio (<inline-formula> <tex-math notation="LaTeX">{G}_{\text {MAX}}/ ~{G}_{\text {MIN}} </tex-math></inline-formula>; ~50), low nonlinearity of <inline-formula> <tex-math notation="LaTeX">\alpha _{\text {p}} </tex-math></inline-formula> = −0.81 and <inline-formula> <tex-math notation="LaTeX">\alpha _{\text {d}} </tex-math></inline-formula> = −0.31, near ideal asymmetry of 0.5, and free of read disturb degradation. High neural network accuracy (>87%) is also obtained.