Highly Linear Weight Update of IGZO Photonic-Synaptic Transistors by Controlling Oxygen Vacancy for Neuromorphic Computing

Optoelectronic synaptic devices attract considerable attention for neuromorphic computing due to high bandwidth and ultrafast signal transmission. IGZO-based photonic transistors have been studied for synaptic devices because of their CMOS compatibility, ultralow-off current, and transparency. They exhibit excellent synaptic functions, such as paired-pulse facilitation (PPF), photonic potentiation/electrical depression, and transition from short-term plasticity to long-term plasticity. However, IGZO-based synaptic transistors have a problem with the weight update nonlinearity, the most critical factor for the accuracy of artificial neural network (ANN).<br/>Now, we suggest a method to improve the linearity of potentiation/depression plasticity of IGZO TFTs by controlling oxygen vacancies in IGZO thin film. First, a substrate bias was applied during IGZO deposition to remove unstable oxygen bonds. Second, nitrogen plasma was followed to fill the generated oxygen vacancies. X-ray photoelectron spectroscopy (XPS) data showed oxygen vacancies effectively reduced by 35 % using our approach. As oxygen vacancies in IGZO thin film decreased, the recovery behavior of excitatory post-synaptic current (EPSC) was almost improved by two times. PPF measurement, indicating an increase in the post-synaptic response by the second stimulation compared to the first stimulation, exhibited that oxygen vacancy-controlled IGZO-based synaptic transistors perfectly mimic biological synapses. Especially, the nonlinearity factor of weight update was dramatically improved from 1.55 to 0.41, which is a sufficient level for highly accurate ANN. By solving the problem of the weight update nonlinearity in IGZO-based synaptic transistor, which has a benefit in CMOS compatibility and ultralow off current, our approach enables the implementation of ultra-low-power and high-performance artificial optoelectronic circuits for neuromorphic computing.