High Linearity Synaptic Devices Using Ar Plasma Treatment on HfO<sub>2</sub> Thin Film with Non-Identical Pulse Waveforms
Ke-Jing Lee,
Yu-Chuan Weng,
Li-Wen Wang,
Hsin-Ni Lin,
Parthasarathi Pal,
Sheng-Yuan Chu,
Darsen Lu,
Yeong-Her Wang
Affiliations
Ke-Jing Lee
Program on Semiconductor Process Technology, Academy of Innovative Semiconductor and Sustainable Manufacturing, National Cheng-Kung University, Tainan 701, Taiwan
Yu-Chuan Weng
Department of Electrical Engineering, Institute of Microelectronics, National Cheng-Kung University, Tainan 701, Taiwan
Li-Wen Wang
Department of Electrical Engineering, Institute of Microelectronics, National Cheng-Kung University, Tainan 701, Taiwan
Hsin-Ni Lin
Department of Physics, National Sun Yat-sen University, Kaohsiung 804, Taiwan
Parthasarathi Pal
Department of Electrical Engineering, Institute of Microelectronics, National Cheng-Kung University, Tainan 701, Taiwan
Sheng-Yuan Chu
Department of Electrical Engineering, Institute of Microelectronics, National Cheng-Kung University, Tainan 701, Taiwan
Darsen Lu
Department of Electrical Engineering, Institute of Microelectronics, National Cheng-Kung University, Tainan 701, Taiwan
Yeong-Her Wang
Program on Semiconductor Process Technology, Academy of Innovative Semiconductor and Sustainable Manufacturing, National Cheng-Kung University, Tainan 701, Taiwan
We enhanced the device uniformity for reliable memory performances by increasing the device surface roughness by exposing the HfO2 thin film surface to argon (Ar) plasma. The results showed significant improvements in electrical and synaptic properties, including memory window, linearity, pattern recognition accuracy, and synaptic weight modulations. Furthermore, we proposed a non-identical pulse waveform for further improvement in linearity accuracy. From the simulation results, the Ar plasma processing device using the designed waveform as the input signals significantly improved the off-chip training and inference accuracy, achieving 96.3% training accuracy and 97.1% inference accuracy in only 10 training cycles.
