Conductive filament formation in the failure of Hf_0.5Zr_0.5O₂ ferroelectric capacitors

Ferroelectric materials show promising properties for building the next-generation logic and memory devices. In particular, the switchable non-volatile states allow information to be stored via the remanent polarization direction. This offers much-improved bit stability, scalability, and significantly reduced power dissipation compared to complementary metal oxide semiconductor (CMOS) devices^1–3 and dynamic random access memory (DRAM).

Since ferroelectricity was discovered in doped hafnium oxide,⁴ it has been considered a promising material for devices.⁵ This is partly due to the conformal coverage and fine thickness control achieved using atomic layer deposition (ALD). In addition, an alloy of hafnium oxide and zirconium oxide, Hf_0.5Zr_0.5O₂ (HZO), facilitates a low crystallization temperature, making it back end of line (BEOL) compatible.^6–10 This cements HZO as a promising ferroelectric material in terms of its potential for integration into emerging logic and memory devices, in addition to the benefits of scalability and low power consumption.¹¹ However, the large fields required to switch HZO and difficulties with scaling film thickness¹² present further challenges to its integration.

An important area of study is the reliability of these ferroelectric films. This includes factors such as ferroelectric wake-up, retention loss, and fatigue failure, which are all limiting factors for the implementation of HZO in non-volatile memory.¹³ The wake-up effect can be a major issue in HZO-based memory devices, with films initially demonstrating a low remanent polarization for the first ∼100–1000 cycles.¹⁴ While the ferroelectric fatigue performance has been studied extensively at room temperature, focusing on improving the fatigue behavior through electrode selection and annealing conditions, the underlying mechanisms and size dependence remain a puzzle.^15–19 

Some studies suggest that a phase transformation from the polar orthorhombic phase to the non-polar monoclinic phase could be responsible for the decrease in polarization during cycling and the eventual failure in HZO devices.^20,21 It has also been suggested that this decrease in polarization is due to charge trapping at oxygen vacancies, which typically become pinned at interfaces, such as the interface with the electrode and at grain boundaries within the ferroelectric.^22–25 Another symptom of charge trapping at the electrode interface is an increase in the bias of the ferroelectric hysteresis loop (the positive coercive field is increased at the expense of the negative coercive field, or vice versa, commonly referred to as imprint).

Despite many reports describing several different failure mechanisms that depend on factors such as film growth and processing, there is limited evidence of the filamentary failure mechanism, even with filament formation being highly studied in HfO₂-resistive switching materials.^29,30

In this work, a systematic study of the ferroelectric fatigue behavior of W/HZO/W capacitors is performed as a function of temperature and device size. Temperatures up to 110 °C and ferroelectric device sizes as small as 0.125 μm² are studied. The lack of polarization degradation during cycling suggests that an orthorhombic to monoclinic phase transformation is not responsible for failure in these HZO thin films. Furthermore, the number of cycles to breakdown follows an Arrhenius behavior with activation energy between 150 and 400 meV, suggesting electric field modulated motion of charged oxygen vacancies. An examination of the leakage current post-failure shows a leakage current that lacks a dependence on capacitor area. The results suggest the formation of conductive filaments via the ordering of charged oxygen vacancies are likely the cause of failure. Consistent with this picture, conductive atomic force microscopy (c-AFM) and simulations of the electric field distribution indicate the failure is local, likely residing on the boundary of the capacitor where the local field is amplified.

Our samples consist of ∼10-nm-thick thin films of HZO grown on a 50-nm-thick W back electrode on a Si substrate and topped with a 20-nm-thick W top electrode. W/HZO/W devices were prepared following previously published procedures.¹⁸ First, DC magnetron sputtering was used to deposit 50 nm tungsten onto a (001)-oriented silicon substrate using a custom sputter system equipped with a Meivac MAK 50.4 mm diameter sputter gun. A 3.3 W cm⁻² sputter power density and 5 mTorr argon processing gas pressure controlled by an Alicat Scientific mass flow controller and a Meivac conductance flow valve were used. An Oxford FlexAL II plasma-enhanced atomic layer deposition instrument was then used to prepare Hf_0.5Zr_0.5O₂ with TEMA-Hf and TEMA-Zr precursors for Hf and Zr, respectively. A 3:2 cycle ratio was used in 16 supercycles to produce an HZO layer of 9.6 nm thickness. Plasma powers of 250 and 300 W were used with oxygen gas as the oxygen precursor for the HfO₂ and ZrO₂ layers, respectively. The growth temperature was 260 °C and a 90 s purge was used following each cation precursor dose. A 20 nm thick W layer was then deposited as the top electrode using the same parameters as the bottom electrode. Devices were annealed in an Allwin AccuThermo AW 610 rapid thermal annealer at 600 °C in a N₂ atmosphere for 30 s with a heating ramp rate of 50 °C s⁻¹.

To measure the phase constitution in the HZO films, grazing-incidence x-ray diffraction was performed using a Rigaku Smartlab with a Cu Kα radiation source and a 0.7° incidence angle [Fig. 1(a)]. The LIPRAS software package was then used to fit the pattern using Pearson VII peak shapes to determine phase fractions.³¹ X-ray reflectivity was utilized to measure the thickness of the HZO films using an Empyrean x-ray diffractometer with a Cu Kα radiation source. Patterns were fit using the GSAS-II software package,³² which confirmed the absence of the monoclinic non-polar HZO phase.³³ 

HZO film microstructure was assessed using atomic force microscopy (AFM) with an Asylum Research Cypher-S instrument in AC mode off tip resonance. A BudgetSensors Tap300Al-G cantilever was used (40 N/m force constant, 300 kHz resonance frequency). The Heyn linear intercept method was used to measure average feature size.³⁴ AFM reveals an average lateral feature size (grain size) of ∼25 nm diameter following a rapid thermal anneal [Fig. 1(b)].

Measuring the pseudo-DC polarization (100 kHz) vs the electric field response of the film in a vertical metal–insulator–metal configuration reveals a typical ferroelectric hysteresis loop for polarization vs electric field, with a remanent polarization of ∼20 μC cm⁻² [Fig. 1(c)]. A Radiant Precision Multiferroic II was used for the pseudo-DC ferroelectric loop measurement. A series of photolithographic and dry etch steps were used to isolate the ferroelectric pillars. Figure 1(d) shows a through-microscope photograph of the electrode configuration used in this work. The samples were prepared using e-beam and optical lithography with an Ar ion mill etch to obtain a ferroelectric capacitor pillar (diameter spanning 0.4–10 μm), which was then encapsulated in an insulator and capped with a Ti/Au metal for probe contact.

Endurance testing was conducted using a read/write pulse sequence intended to better mimic operational memory device speeds. A Keysight 33600A waveform generator and a 13 GHz capable Keysight Infiniium UXR0134A real-time oscilloscope were used to source the pulse sequence and measure the current response of the ferroelectric device, respectively. Initial measurements prior to cycling at each temperature and device size were performed using a Berkeley Nucleonics Model 765 Fast Rise Time Pulse Generator. The same setup was used for post-failure characterization. A heating stage with a Eurotherm nanodec controller was used.

These sequences and the circuit configuration are illustrated in Fig. 2(a). A 10 ns write pulse sequence illustrates the high switching speed potential in HZO-based ferroelectric memory devices. A longer read pulse is used after 10, 100, 1000, and so on, cumulative cycles to ensure an accurate polarization measurement. The polarization cycling results with respect to device size are shown in Fig. 2(b), and the fatigue results are shown in Fig. 2(c).

Using the mean electric potential, estimated from a constant electric field and mean film thickness used in these measurements, this equates to a vacancy charge of ∼0.4–0.6e ± 0.16e. Different from what would be expected for oxygen vacancies with a charge of +2, this could be due to other mobile vacancies also being moved, including singly-charged oxygen vacancies. Using the diffusion pre-exponential factor reported by Shvilberg et al.³⁷ of D₀ $≈$ 10⁻⁹ cm² s⁻¹ and the average field-modified barrier, we calculate a RMS diffusion length of ∼10 nm at 30 °C.

$EA′$ is shown in Fig. 3(a) as a function of ferroelectric device area. This area dependence reinforces that failure is dependent on area size, possibly due to the increased number of defects in a larger device or larger perimeter length in larger devices, as will be discussed below. Following ferroelectric device failure, the leakage current of the device was measured to elucidate the failure mechanism. The measured leakage current for each device size is shown in Fig. 3(b). The measured current displays independence of the ferroelectric pillar area. This is indicative of the formation of conductive filaments, formed from the motion of oxygen vacancies within the ferroelectric film.³⁸ We note that we have not observed a reversibility of these filaments under cyclic voltage application, as one would find in a memristor. It has previously been hypothesized that oxygen vacancies may become pinned at grain boundaries, resulting in the formation of these irreversible conductive filaments.¹⁸ The pinning of vacancies at a grain boundary would support the irreversible filament formation. Thus, a strategy to enhance endurance may be to increase the mean grain size of the film. The conductive filament-based failure mechanism hypothesis is further supported using c-AFM where it is possible to write a conductive filament using an applied voltage following the removal of the top electrode, as seen in Figs. 3(c) and 3(d). For conductive AFM, to expose the bare ferroelectric layer, the top electrode (300 nm Au/10 nm Ti/20 nm W) needed to be stripped away. This was done by first dipping the sample into an Au etch (GE-8111) for 3 min, followed by ion milling to etch away the remaining Ti and W. The etch was confirmed with electrical measurements showing a complete removal of the conductive material. Conductive-AFM was performed using atomic force microscopy (NT-MDT NTEGRA) with a conductive platinum-coated tip (20 kHz, 0.3 N/m MikroMasch HQ:CSC37/Pt). The image was acquired by sourcing a voltage of 1 V through the tip and tracking the current during the scan, operated in contact mode. Conductive filaments were written using a voltage of up to 4 V. Prior work has verified the potential to write conductive regions in the HZO film through the use of a large DC bias without cycling.³⁹ However in this work, we hypothesize that during the writing process, filaments that partially formed during cycling [that occurred near the edge due to field amplification (Fig. S1)] finish forming, resulting in multiple conductive pathways near the edge of the device, as observed. This can help explain the dependencies observed in Figs. 2(b) and 2(c). The temperature dependence observed can be explained by the increased vacancy mobility. This means the vacancies can move further with each applied field cycle, resulting in failure in fewer cycles.

The area dependence of cycling endurance and initial polarization (before any cycling) is shown more clearly in Figs. 4(a) and 4(b), respectively. From this, failure is shown to occur more quickly in devices with larger ferroelectric pillars. We theorize that the device area dependence is due to larger ferroelectric pillar areas containing a larger number of defects at which a filament could nucleate, resulting in a higher statistical probability of filament formation. While there does appear to be an area dependence, it is lower than expected when compared to the literature.^42, Figure 4(b) highlights that the initial polarization is relatively invariant with respect to area and decreases significantly with a change in temperature. The remanent polarization dependence observed contradicts some literature studies, where an increase in polarization is observed upon heating.^43–45 This can be correlated with a wake-up behavior, which was not present in a significant way in these samples, as shown in Fig. 2(b). In other works, it has been hypothesized that a decrease in remanent polarization could be due to a reduction in the ferroelectric phase due to a partial transition to the antiferroelectric orthorhombic phase and/or an increase in defect formation,^44,46,47 although this was generally only noticed at temperatures above 100 °C. This is possibly recovered with cycling through the annealing and wake-up effect, as the remanent polarization at different temperatures seems to converge to a similar value with cycling.

In this work, W/Hf_0.5Zr_0.5O₂/W devices were fabricated into isolated ferroelectric pillars with diameters ranging from 0.4 to 10 μm. Using 9 MHz endurance tests with 10 ns, 3 MV cm⁻¹ pulses, Arrhenius behavior of breakdown was revealed, indicative of activated ion transport. After failure, a high leakage current is observed, invariant of ferroelectric device area, indicating the formation of conductive filaments. Conductive AFM measurements reveal the ability to create local regions of high conductivity after failure, consistent with filament formation. Simulations indicate an amplified field at the boundary of the capacitor, suggesting the filament at the boundary. In light of the irreversibility of the conductive filaments, potentially due to the pinning at grain boundaries, we recommend further work relating the endurance of Hf_0.5Zr_0.5O₂ to the grain size and/or the grain boundary density. While we report that the vacancies have lower charge than anticipated, it is ultimately their charge that facilitates their motion in an electric field and endurance failure. The potential to produce only neutral oxygen vacancies would mitigate an influence by an applied electric field. One such way of producing such chargeless oxygen vacancies could be appropriate doping that compensates for the charge of the missing oxygen atom.