Resistive random-access memory (RRAM) devices with hydrogen plasma treated HfO2 have shown low power switching (1) and good conductance quantization with programing pulsed operation (2) that qualify them to be used for in-memory computing. Engineering the distribution of defects or oxygen vacancies near the top and bottom electrodes has a significant impact on reducing the switching power and improving the multi-level cell (MLC) characteristics of the device. A graded distribution with higher concentration of oxygen vacancies closer to the top electrode (TE) due to hydrogen plasma treatment and lower concentration near the bottom electrode (BE) reduces switching energy (power) (1). However, the impact of the bottom electrode is not well understood. In this work we investigated three RRAM stacks as shown in Figure 1 that have the same top electrode metal (50nm PVD TiN/5nm ALD TiN) and dielectric (HfO2 with hydrogen plasma treatment at the mid-point) but with different BE for each stack: TiN (device A (control)), TiN plus Mo with thermal anneal (device B), and TiN plus Mo with plasma nitridation (device C). Choosing a more inert metal as the BE provides a lower oxygen vacancy concentration at the interface between the BE and the dielectric (3) which in turn reduces the switching power as noted with device C (TiN plus Mo with plasma nitridation) which switches at a minimum compliance current (Icc) of 500pA. It is possible that the presence of nitrogen prevents the increase in oxygen vacancy concentration near the BE. On the other hand, a higher concentration of oxygen vacancies at the BE and the dielectric interface leads to a larger magnitude of band bending and increases the height of Schottky barrier. This possibly leads to an increase in the switching power in device A and device B as compared to device C (3-4). Note that device A (TiN) switched at Icc=30µA and device B (TiN plus Mo with thermal anneal) switched at Icc=8µA, which is superior compared to device A in switching power. This suggests that the alloy of TiN with Mo is reducing the oxygen vacancies near the BE. To study the MLC characteristics of the devices, we performed a train of 36 RESET pulses with fixed amplitude Vp= -1V and pulse width starting at 4µs and increasing by 20µs every 3 pulses, while we applied Vr=0.1V and tr=10µs after each pulse to measure the conductance. Figure 2 shows that device B (TiN plus Mo with thermal anneal) has good MLC with no overlap between the states while device A shows unstable MLC behavior due to time constant variation. Figure 3a shows the results of device C for the same RESET pulse train that was applied to devices A and B. Since device C switches at very low current and voltage, the conductance quantization was unstable after 15 pulses due to overstress of the conductive filament. Figure 3b shows the results after selecting the proper pulse set Vp= -0.5V and pulse width starting at tp=100ns and increasing by 100ns every 3 pulses, while we applied Vr=0.05V and tr=100ns after each pulse, resulting in a better conductance quantization compared to the previous setup. Appropriate engineering of the BE and managing of the oxygen vacancy distribution by hydrogen plasma treatment can significantly reduce the switching power.
Read full abstract