Impact of ALD Growth Mode on the Stochasticity and Reliability of Pt/Ag/HfO2/W/Ti Memristors
Please login to view abstract download link
In this work, we investigate the influence of the Atomic Layer Deposition (ALD) growth method (Thermal vs. Plasma-enhanced) on the resistive switching (RS) performance of Pt/Ag/HfO2/W/Ti memristors (Fig. 1) with a fixed 10 nm oxide thickness. Electrical characterization was performed using a Keysight B2912b SMU, applying the voltage to the Pt/Ag top electrode while keeping the W/Ti bottom electrode grounded. Both samples exhibited high endurance, sustaining a large number of switching operations (Fig. 2). Nevertheless, the Plasma-Enhanced ALD (PEALD) process significantly improved switching stability compared to its Thermal ALD counterpart. Switching voltages (extracted as described in [1]) show lower stochasticity in the PEALD devices, exhibiting a tighter distribution, whereas the Thermal ALD samples present broader variability (Fig. 3). This trend is further confirmed by the cumulative probabilities of the resistance states, where the PEALD films show superior cycle-to-cycle stability, especially for the LRS (Fig. 4a). Furthermore, the correlation plot of the LRS vs. HRS reveals that the PEALD devices show a less variable HRS/LRS ratio between the 5x and 10x lines, while the Thermal ALD data results in a highly dispersed data cloud with frequent excursions toward the x20 ratio (Fig. 4b). These findings suggest that the PEALD process favours more confined conductive filaments that form in well-defined regions, effectively reducing the randomness of the switching transitions. This aligns with the findings of Zhang et al. [2] and Martínez-Puente et al. [3], who reported that the higher density of intrinsic defects and charged oxygen vacancies associated with the thermal ALD process can lead to the formation of multiple, unstable conductive paths, thereby increasing the stochasticity of the switching mechanism. Consequently, while both methods provide functional RRAM devices, PEALD is demonstrated as a superior growth mode for applications requiring high uniformity and robust memory operation.