Ultra‑Low‑Power HfO₂/TiOₓ Bi‑Layer RRAM Achieves 1.12 µW Switching via Controlled Oxygen Vacancy Engineering

Abstract

We have successfully fabricated resistive random‑access memory (RRAM) devices comprising an atomic‑layer‑deposited HfO₂ layer and an RF‑sputtered TiOₓ layer. By fine‑tuning the oxygen partial pressure during TiOₓ deposition, we achieved an unprecedented low‑power operation: 1.52 µW for the set transition (1 µA@1.52 V) and 1.12 µW for reset (1 µA@1.12 V). Detailed analysis reveals that a moderate oxygen vacancy concentration in the TiOₓ layer suppresses “soft‑breakdown” during forming/set, limiting the current and dramatically reducing power consumption.

Introduction

Resistive random‑access memory (RRAM) is a leading candidate for beyond‑CMOS scaling because of its simple two‑terminal structure, non‑volatile behavior, fast switching, and high on/off ratios. The 1‑transistor‑1‑resistor (1T1R) configuration is now standard for large arrays, mitigating sneak‑path currents. However, the high switching currents of conventional oxide‑based RRAMs hinder scaling, especially when CMOS drive currents drop below 40 µA at 27 nm nodes. Likewise, neuromorphic circuits that emulate biological synapses require energy per event in the femtojoule range. Thus, engineering RRAMs to operate reliably below 10 µA is critical for both high‑density storage and energy‑efficient artificial neural networks.

In this work, we systematically investigate Pt/HfO₂/TiOₓ/Pt devices with varying oxygen content in the TiOₓ layer. We demonstrate that lowering the oxygen partial pressure during sputtering creates an optimal vacancy profile that confines the conductive filament, yielding sub‑micro‑watt switching power. The underlying conduction mechanism is analyzed, providing actionable guidance for designing low‑power oxide RRAMs.

Methods

The device stack was grown on a Si/SiO₂/Ti substrate. A 100‑nm Pt bottom electrode (BE) was sputtered at room temperature. A 3‑nm HfO₂ layer was deposited by ALD (Picosun R200) at 300 °C using TEMAH and H₂O precursors. Subsequently, 30‑nm TiOₓ films were RF‑sputtered under a total gas flow of 20 sccm (Ar + O₂). By adjusting the O₂ partial pressure to 9 %, 11 %, and 13 %, we fabricated three device series (S1, S2, S3). A 70‑nm Pt top electrode (TE) was then sputtered and patterned by lithography. Reactive ion etching defined 100 µm × 100 µm square cells. Cross‑sectional high‑resolution TEM (HRTEM) verified the layer integrity, and X‑ray photoelectron spectroscopy (XPS) quantified the oxygen vacancy levels.

Electrical measurements employed an Agilent B1500A semiconductor parameter analyzer with a 1 µA compliance current for set/reset operations. Temperature‑dependent I–V curves were recorded between 25 °C and 125 °C to probe conduction mechanisms.

Results and Discussion

XPS O 1s spectra of TiOₓ films show two components: lattice O²⁻ and non‑lattice O²⁻ associated with oxygen vacancies. The vacancy fraction is 28.23 % (9 % O₂), 24.06 % (11 % O₂), and 23.63 % (13 % O₂), confirming that higher O₂ partial pressure reduces vacancies.

During current‑forming (CF) with 1 µA compliance, only S1 (9 % O₂) successfully forms a stable conductive filament. S2 and S3 require up to 20 mA to reach the low‑resistance state (LRS), indicating excessive vacancy density that promotes soft‑breakdown across the dielectric stack.

100‑cycle bipolar I–V sweeps reveal that S1 achieves stable switching with <10 µA compliance, whereas S2/S3 demand 10 mA. The low‑power performance of S1 originates from a high vacancy concentration that limits the current during CF/set, preventing excessive filament growth.

Statistical analysis (cycle‑to‑cycle and device‑to‑device) shows modest variation for S1 and acceptable on/off ratios (>100) for all samples. Retention tests demonstrate that S1 retains its states for >10⁴ s at 85 °C, confirming its suitability for non‑volatile storage.

Power calculations yield P_set = 1.52 µW and P_reset = 1.12 µW for S1 at 1 µA compliance. In contrast, S2/S3 exhibit tens of milliwatts due to their high switching currents.

Temperature‑dependent resistance measurements reveal that S1’s resistance decreases with temperature, consistent with Mott’s variable‑range hopping (VRH) model. This indicates electron hopping via localized oxygen vacancy states in TiOₓ dominates conduction. For S3, resistance is temperature‑insensitive, suggesting metallic‑like filamentary transport through a dense vacancy network. The latter behavior leads to higher currents and power consumption.

These results underscore the importance of a balanced vacancy concentration: sufficient to enable hopping conduction and limit filament growth, yet not so high that metallic filaments form. This insight is applicable to a wide range of oxide‑based RRAMs, guiding the design of energy‑efficient memory for both data storage and neuromorphic computing.

Conclusions

We demonstrated 1‑µA switching currents in Pt/HfO₂/TiOₓ/Pt RRAMs by controlling the oxygen vacancy profile in the TiOₓ layer. Electron hopping dominates in low‑oxygen films, confining current and yielding sub‑micro‑watt power. In contrast, high‑oxygen films form metallic filaments, causing soft‑breakdown and high power consumption. Optimizing the oxygen vacancy density is thus a universal strategy to achieve low‑power, high‑density RRAM suitable for IoT and embedded applications.