Dual Synaptic Weights Emulation for Artificial Nociceptors and Energy Efficient Neuromorphic Computing

With the increased diverse sets of information in the era of big data, neuromorphic computation is the crucial technology for advancing the data-processing speed. Electronic devices, which mimic the neural activity with good energy efficiency, will be the key for successful hardware implementation of neuromorphic computing. Inspired by the charge transfer mechanism, the capacitance-based synaptic weight is highly advantageous for the energy efficient neuromorphic computing because the capacitor consumes lower power dissipation and is also free of sneak-path issue due to the voltage-expressed transmitted signal. Furthermore, three-terminal artificial synapse is also realized with the same stacking layer to emulate the synaptic functions and features the concurrent signal transmitting-and-learning.<br/>To be compatible with CMOS process technology, here we fabricated a metal oxide-based synaptic capacitor on p⁺-Si substrate. The indium tin oxide (ITO) and tantalum oxide (TaO_x) are stacked as the capacitive device. In parallel, the same stacking layers are employed as the channel layer (ITO) and high-k electrolyte dielectric layer (TaO_x), respectively, to form the synaptic transistor. The relocation of oxygen vacancies in TaO_x under the application and removal of the electric field contributes to the analog transistor conductance and capacitance modulation. Consequently, multiple synaptic functions, including pair-pulse facilitation, high-pass filter, long-term plasticity as well as the transition from short-term memory to long-term memory, can be successfully demonstrated in both transistor and capacitive device. On top of that, it is unparalleled to mimic the nociceptive behavior through capacitance as the response. To confirm the feasibility of both electronics, the dataset classification simulation is carried out by incorporating the results of weight updating process to the CrossSim software. A robust stability of cycle-to-cycle variation is found in capacitive device (0.9%) and transistor (1.8%), causing a high recognition accuracy of 95% and 84%, respectively. It indicates that the ITO/TaO_x-based capacitor shows superior performance and has promising potential in the hardware capacitive neural network.