A 2T Neuromorphic Device based on Metal Oxide Transistor with Self-Assembled Monolayer for High-Precision Artificial Intelligence

State-of-the-art artificial intelligence computation consumes tremendous power and requires a long time during training as the separated memory and processor of the current computing architecture hinder data transfer. To deal with this problem, researchers have developed a neuromorphic device that can operate weight storage, update, and matrix-vector multiplication near memory. However, conventional neuromorphic devices based on new non-volatile memories suffer from abrupt conductance shifts and poor uniformity, which degrade the accuracy of the neural network.<br/>This work presents a neuromorphic device with excellent synaptic characteristics by utilizing two InGaZnO (IGZO) transistors. The proposed neuromorphic device consists of a read transistor and a write transistor, where the write transistor's source/drain electrode is connected to the gate electrode of the read transistor so that the write transistor precisely adjusts the gate potential of the read transistor. The ideal synaptic properties were achieved by balancing the gate electrode charging/discharging speed and the transconductance change rate to have an identical programming slope within the operation range. The ultra-low leakage current of IGZO prevents the charge leakage from the read transistor gate through the IGZO channel, preventing the programmed weights change during training. Furthermore, with the large-area uniformity of the IGZO transistor, the programming slopes of the proposed neuromorphic device exhibited a very low standard deviation to the average ratio of less than 3%. For large array compatibility, we regulate the output current of the neuromorphic device by introducing an alkyl-phosphonic acid self-assembled monolayer (SAM) between IGZO and source/drain to form an ultra-thin energy barrier. A phosphonic acid SAM with 12 alkyl chains reduced the current of the IGZO transistor by 30%. In addition, the energy barrier formed by the SAM suppressed leakage current during retention, maintaining the stored weights after 300 seconds from programming. With these strategies, the 2T neuromorphic device exhibited outstanding synaptic characteristics within a low conductance range, which is suitable for a high-precision artificial intelligence system.