Synaptic Characteristic of Li Doped HfO₂ with TiN Electrode

The human brain is capable of simultaneously processing and memory with low power consumption. A computing method that imitates these characteristics is called neuromorphic computing. Neuromorphic computing has high energy efficiency by mimicking brain works in hardware. To mimic the brain, we need to imitate synaptic behavior: electrical stimulus causes a continuous resistance change in the synaptic device. Resistive random access memory(RRAM) is one of the strong candidates attended for the synaptic device. The two-terminal structure of RRAM enables high-density integration. The simple structure makes crossbar array: Structure used in neuromorphic semiconductors. But RRAM needs the forming process, higher than the operating voltage. The forming process degrades the endurance of RRAM.<br/>In this study, we confirmed the synaptic behavior, forming voltage controllability of lithium-doped hafnium oxide with a titanium nitride electrode. The structure is Pt/Li: HfO₂/TiN. The bottom electrode, 85nm-thick TiN as Li reservoir was deposited by magnetron sputter. The switching layer, Li:HfO₂ was co-sputtered in a lithium-rich to have a forming-free characteristic by Magnetron sputter. The top electrode, 200nm-thick Pt was deposited by Magnetron sputter. The richer lithium concentration (Li:Hf = 1:1) modulates the forming voltage from 2.5V to 2V. The device shows synaptic characteristics. The synaptic characteristics are Habituation, Sensitization, and Paired Pulse Facilitation (PPF). Habituation is that the conductivity increases as the pulses are applied, and the element becomes insensitive to the electrical pulse. Sensitization is when the conductivity decreases as the pulse is applied, making the device sensitive to the electrical pulse. PPF is a phenomenon in which short intervals of electrical pulses are input like a signal with strong conductivity, and wide intervals are input as a weak signal. We also checked the 8x8 crossbar array learning speed simulation of the learning speed based on the nonlinearity : (G : synaptic weight, G₀ : initial weight A : Constant to fit the graph, B : nonlinearity, P : number of pulses, P₀ : initial number of pulses) Our device’s nonlinearity was 0.1/3.6 (Potentiation/Depression). When simulated with this nonlinearity, the accuracy was confirmed to be 96.3% at 500epoch. —through simulation, the accuracy increased to 97.9% at 500epoch when the nonlinearity of depression was only 1— our research showed the possibility of a forming-free device with modulation of lithium concentration. We confirmed even if the synaptic device has insufficient linearity and symmetry, it can be used as a neuromorphic device.