Advanced artificial synaptic thin-film transistor based on doped potassium ions for neuromorphic computing via third-generation neural network

• Published in: Journal of materials chemistry. C, Materials for optical and electronic devices Materials for optical and electronic devices, 2022-02, Vol.10 (8), p.3196-3206
• Main Authors: Cao, Yixin, Zhao, Tianshi, Zhao, Chun, Liu, Yina, Song, Pengfei, Gao, Hao, Zhao, Ce Zhou
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial intelligence; Artificial neural networks; Cognitive tasks; Computer memory; Computer simulation; Machine learning; Neural networks; Neuromorphic computing; Object recognition; Parallel processing; Potassium; Semiconductor devices; Synapses; Thin film transistors; Time dependence; Transistors
• ISSN: 2050-7526; 2050-7534
• DOI: 10.1039/D1TC04827A

As the basic and essential unit of neuromorphic computing systems, artificial synaptic devices have great potential to accelerate high-performance parallel computation, artificial intelligence, and adaptive learning. Among the proposed artificial synaptic devices, the synaptic transistors are well considered to be one of the most suitable devices for simulating artificial intelligence. So far, synaptic transistors based on iontronic have been proposed and proved to demonstrate great potential in artificial intelligence applications. However, little research specifically focused on improving the device's ability to mimic synaptic behaviour. Here, we proposed the enhancement of synaptic properties of the solution-based thin-film transistors based on potassium ion conduction in the dielectric layer for the first time. Due to the formation of a gated electrical double-layer, the transistor exhibited an enlarged memory window. Based on this, the excitatory postsynaptic current in the synaptic thin-film transistor was modified accordingly, which further enhanced the suitability of the proposed synaptic thin-film transistor for simulating biological synapses. In addition, considerable synaptic properties were evaluated elaborately, including paired-pulse facilitation, short-term memory, long-term memory, and spike-time-dependent-plasticity. Most importantly, according to the impressive results of the Artificial Neural Network algorithm's image recognition simulation, the simulation image recognition rate based on the mentioned artificial synaptic devices reached as high as 92%. Last but not least, in order to simulate biological neurobehavior more closely, the Spiking Neural Network algorithm was also successfully implemented to complete the specified machine learning task, which further proved the great potential of the synaptic devices in advanced low-power neural network systems.