Impact of vacancies and impurities on ferroelectricity in PVD- and ALD-grown HfO₂ films

HfO₂ is a key multifunctional material for current device technology. Recently, it is also been successfully used as a ferroelectric or memristive material in nonvolatile memory devices and for novel neuromorphic computing concepts.^1–3 Atomic layer deposition (ALD) has become the most widely employed growth process since it allows for a precise thickness control on the level of single atomic layers at CMOS-compatible low deposition temperatures.

However, ALD-grown HfO₂ thin films can contain a significant amount of impurities like carbon or nitrogen if the decomposition of the precursor gas in ozone was incomplete or removed ligands readsorb on the surface. The impact of N or C impurities on the HfO₂ thin film ferroelectric properties is not fully clarified to date.

As an alternative technique, physical vapor deposition (PVD) by sputtering from an HfO₂ target was explored to gain even better control on the stoichiometry and defect concentration in HfO₂ thin films.

Oxygen vacancies are recognized as those defects, which have a positive impact on the stabilization of the ferroelectric orthorhombic phase in undoped HfO₂.^4–6 Here, we present a comparative study of both ALD- and PVD-grown HfO₂ thin films interfaced with TiN bottom/top electrodes (BE/TEs) integrated into TiN/HfO₂/TiN capacitors. We apply hard x-ray photoelectron spectroscopy (HAXPES) as an equally depth-sensitive and element-specific chemical characterization technique,^7,8 which allows us to gain a comprehensive picture of the chemical state of the HfO₂ layers. In our analysis, we focus, in particular, on identifying spectral signatures of oxygen vacancies and nitrogen/carbon impurities. We correlate these spectroscopic findings with electrical measurements of the remanent polarization of the ALD- and PVD-grown TiN/HfO₂/TiN capacitors. In this way, possible mechanisms of stabilization of the ferroelectric HfO₂ phase, resulting from the specific PVD or ALD conditions, are identified.

Sample preparation for PVD and ALD samples is sketched in Fig. 1. For further details, see the supplementary material.

In order to access chemical information from the TiN/HfO₂ bottom interface, we performed HAXPES measurements on interstitial areas between the etched TiN top electrodes. The observation of interfacial TiO₂ in the case of PVD (see Fig. 2) arises from the oxygen supply $ṁ$ provided before the in situ PVD process (see Fig. 1, left).⁹ The effective thickness of the TiO₂ intralayer saturates in a self-limiting process at about 3.5 (±0.5) nm. As a consequence, the chemical exchange between the TiN electrode and the HfO₂ is strongly reduced or even suppressed by the TiO₂ intralayer, which serves as a chemical buffer.⁹ For the ALD-grown samples, we also observed spectral contributions of interfacial TiO₂ (see Fig. 2). Here, a sizeable surface oxidation is a result of an exposure to air for at least 24 h before ALD growth of HfO₂. Indeed, the TiO₂-related spectral weights of both ALD_10 s and ALD_60 s samples are comparable to those of the PVD_2 sccm sample at $ṁ=2$ sccm. Independent of its chemical origin, however, the TiO₂ layer serves as a chemical buffer also in ALD-grown samples.

Figure 3 shows the Hf 4f core levels for both sets of PVD- and ALD-grown samples. The intensity is normalized to the Hf 4f_5/2 peak maximum. We find the PVD_0 sccm and PVD_2 sccm Hf 4f peak maxima located at higher binding energies compared to the spectra of both ALD-grown samples. The largest binding energy shift $ΔEB=270$ meV is observed between the PVD_0 sccm and ALD_10 s samples. It should be noted that we observe these binding energy shifts between PVD- and ALD-grown samples in all spectral features, which are related to the HfO₂ layer (also see the supplementary material). On the other hand, core level features of the TiN electrodes or the TiO_2−x interface layer do not show any relative binding energy shift $ΔEB$⁠.

Next, we determined the Hf 4f difference spectra relative to the ALD_60 s sample (see the supplementary material). They reveal additional Hf 4f contributions at lower binding energies originating from Hf ³⁺, which are attributed to the presence of oxygen vacancies $V0$⁠.^9–13 For a quantitative analysis, the Hf 4f spectra were fitted consistently by two doublets. The fits for the Hf³⁺ contributions are shown as shaded peaks in the lower part of Fig. 3, and the particular Hf³⁺/Hf ⁴⁺ areal intensity ratios are given in Table I.

As expected, we find a sizeable Hf³⁺ component for the PVD_0 sccm sample and only a small one for PVD_2 sccm, while both ALD samples are nearly free of Hf³⁺. For ALD_10 s, we still find a tiny Hf³⁺ component that disappears completely after a longer ozone dose in ALD_60 s, in agreement with recent findings.¹⁴ 

Figure 4 shows the Hf 4s and O 1s core levels, which have been normalized to the Hf 4s peak. The O 1s peak intensity shows a gradual increase from PVD_0 sccm, PVD_2 sccm, ALD_60 s to ALD_10 s. The O 1s/Hf 4s intensity ratios obtained from an integration over the normalized spectra are given in Table I. For PVD samples, a larger additional oxygen supply leads to a larger oxygen intensity. Moreover, a larger ozone dose (ALD_60 s) results in a lower oxygen spectral intensity compared to the ALD_10 s sample.

The O/Hf intensity ratio may decrease for two reasons: First, undesired processes such as the readsorption of organic precursor ligands can likely occur for longer ozone times and may modify the O/Hf ratio in a poorly controllable manner. Second, the relative O/Hf ratio may also decrease by increasing the absolute Hf content, which is most likely the case for the ALD_60 s sample. We note that such large O₃ doses (60 s) are typically not applied in ALD, and we use it here for the purpose of studying the impact of impurities.

Furthermore, the O 1s spectra show a slight increase in intensity at a binding energy of about 533 eV (valley region) for both ALD spectra as compared to PVD. This intensity increase indicates a satellite peak, which is split off from the main O 1s peak by about 1.8 eV (see the supplementary material, Sec. II B). This feature has been discussed in the literature as a signature for an oxygen vacancy V₀ in different oxide materials,^15–20 with either a 1+ or 2+ valency.^12,13,21

A major difference between ALD- and PVD-grown HfO₂ thin films results from using a metal-organic precursor molecule (TEMA-Hf) in the ALD case. We find clear signatures of C and N impurities incorporated in the ALD-grown HfO₂ films, as discussed in the following.

In Fig. 5, the Carbon C 1s core levels recorded at different photoelectron emission angles are shown for (a) ALD- and (b) PVD-grown samples. All spectra were normalized to the neighboring Hf 4d peaks (not shown). The amount of carbon detected in the PVD-grown HfO₂ layers is, as expected, very low. Moreover, the C 1s peaks do not show any dependence either on the oxygen flow during PVD growth or on the emission angle of the measurement. This finding indicates a homogeneous distribution of carbon impurities within the sample stack, which most likely originates from residuals in the vacuum chamber present throughout the deposition.

Clear variations are, however, observable in the C 1s core level of the ALD-grown samples: The C 1s peak intensity is reduced to $∼40%$ while increasing the O₃ dosage from 10 to 60 s. Moreover, the C 1s peak intensity shows a strong decrease at a larger surface sensitivity of $41°$⁠. Hence, we conclude on an accumulation of carbon impurities toward the bottom TiN electrode [see the inset of Fig. 5(a)]. This finding suggests that the cycled ALD reaction of the TEMA-Hf precursor molecule with O₃ is less effective on a TiO₂ surface, i.e., at the beginning of the HfO₂ growth process, thereby enhancing the amount of organic C residuals in the first HfO₂ layers.

The HAXPES spectra of the nitrogen N 1s core levels, shown in Fig. 6 for (a) ALD and (b) PVD samples, are both dominated by the Ti–N contribution from the bottom electrode and O–Ti–N contribution from the TiN/HfO₂ interface.^9,10 The peak shoulder at E_B = 396 eV is related to Hf⁴⁺-N coordination,²² whereas the broad emission around E_B = 399 eV to C–N bonds and interstitial Nitrogen contributions.^23,24

In the PVD samples shown in Fig. 6(b), the Hf-N component only appears for PVD_0sccm and is related to an Hf-N formation at the TiN/HfO₂ interface, which is suppressed by the TiO₂ intralayer formed in PVD_2sccm samples: Since an interface reaction in ALD samples is also protected by a TiO₂ layer, the Hf-N component in ALD samples has to be associated with a residual impurity of the ALD process. The same applies to the N component around E_B = 399 eV originating from C-N or rather interstitial N. This component shows hardly any angular dependence but is strongly reduced by increasing the O₃ dose from 10 s to 60 s. The intensity of the Hf-N component, however, remains basically constant for an increased O₃ dose, which indicates a more stable bonding. From its angular dependence (see the supplementary material, Sect. IIC), we conclude on its accumulation at the bottom electrode.

Next, we discuss the experimental observations in terms of the interactions between oxygen vacancies (V₀) and impurities (C or N).

Generally, the ferroelectric phase in HfO₂ can be stabilized by defects.⁶ For PVD samples, the oxygen vacancies likely take this function.⁴ In HAXPES spectra, oxygen vacancies V₀ manifest themselves as a small Hf³⁺ component at the low binding energy side of Hf 4f spectra (see Fig. 3). The Hf³⁺ signature appears, if the two electrons left behind at the vacancy site from a removed neutral oxygen atom are distributed among two neighboring Hf⁴⁺ atoms.

The reduced Hf³⁺ component in the PVD_2sccm compared to PVD_0sccm originates from the larger amount of oxygen provided during growth, which fills up the vacancy sites. The emerging ferroelectric properties are fully consistent with this chemical state: Whereas the PVD_0sccm sample is ferroelectric with 2P$r,105=18.2$ uC/cm², see also Fig. 7, the PVD_2sccm sample does not show a significant ferroelectric remanence, as discussed in our previous work.⁹ 

One might conclude that the absence of the Hf³⁺ component in both ALD Hf 4f spectra (Fig. 3) should result in a reduced ferroelectric remanence. Yet, quite the contrary is the case: Both ALD_{10 s,60 s} samples reveal a sizeable remanent polarization with 2P$r,10510s=13.8$ μC/cm² and 2P$r,10560s=11.1$ μC/cm², respectively. This is remarkable because the O/Hf ratio increases for the ALD samples compared to the PVD samples. Thus, we suggest that the role of the other impurities, i.e., carbon and nitrogen, needs to be taken into closer consideration.

Kim et al.²⁴ argued that C impurities can also stabilize the ferroelectric HfO₂ phase, and Rodenbücher et al.²⁵ found that a high concentration of oxygen vacancies considerably lowers the formation energy of Hf-C bonds.

However, the C 1s spectra do not show signatures of such an Hf-C contribution, and thus, we rule out a stabilization of the ferroelectric phase in HfO₂ by such a process in the ALD samples.

Rather, we shall discuss the stabilization of the ferroelectric phase by the incorporation of nitrogen impurities into HfO₂ thin films. Indeed, nitrogen has been reported both theoretically and experimentally to be the only anion dopant that can stabilize the ferroelectric HfO₂ phase.^26,27 In their ab initio models, Umezawa et al.^28,29 and Xiong et al.³⁰ proposed an incorporation of nitrogen at the two oxygen nearest neighbor sites of a vacancy. This configuration was found to be energetically most favorable compared to other possible occupation sites for nitrogen, including the vacancy site itself, since two nitrogen will capture both electrons from the oxygen vacancy and form a closed shell VN₂ complex.^28,31,32 Locally, the valencies then amount to V$O2+$ for the oxygen vacancy site, N³⁻ for nitrogen, and hafnium remains in a Hf⁴⁺ state.

On the experimental side, Lomenzo et al. concluded on the formation of this nitrogen-vacancy complex as an interface reaction at the TaN/HfO₂ interface.³³ In our case, such an interface reaction can be excluded due to the passivation by the TiO₂ interlayer.

This modeling is fully consistent with the HAXPES data of the ALD samples, which reveal vanishing of oxygen vacancy signatures V₀ in the Hf 4f spectra and appearance of stable Hf⁴⁺-N bonds in the N 1s spectra and vacancy signatures in the O 1s spectra. Moreover, the incorporated N impurities may even compensate the reduced oxygen vacancy density by stabilizing the ferroelectricity in ALD samples.^26,27

Next, we consider the peak shifts of Hf 4f and O 1s toward lower binding energies (Fig. 3). They can be attributed to the decreasing density of oxygen vacancy states within the bandgap. For the PVD samples, this is reflected by the decreasing Hf³⁺ component, caused by the replenishing of vacancy sites by oxygen atoms and the increasing O 1s intensity (Fig. 4). For the ALD samples, the larger shift of Hf 4f toward lower binding energies as compared to the PVD samples can also be attributed to a further decrease in neutral oxygen vacancy states within the bandgap. However, here the N impurities play an additional role in the electronic structure formation. The vacancies are not simply replenished but also electronically passivated by nitrogen impurities, thus forming a VN₂ complex. The former V₀ gap states are now shifted into the valence band.³⁰ 

As highlighted before, oxygen vacancies are essential for stabilizing the ferroelectric phase in HfO₂ films and also determine the onset of the wake-up effect during electric field cycling of the TiN/HfO₂/TiN capacitors.^4,5,34,35 A key parameter is the mobility of the oxygen vacancies under the application of an electric field. It determines their spacial redistribution and increases the ferroelectric polarization during the wake-up cycling. Hence, a change in the mobility of oxygen vacancies should result in a sizable alteration of the electric wake-up behavior.

In the case of ALD-grown samples, the formation of a vacancy-impurity VN₂ complex should result in a decreased mobility compared to vacancies V₀ in PVD-grown samples, due to the chemical bonding of the oxygen vacancy to the nitrogen impurity.³⁰ Fully consistent with this assumption, indeed, the electric field cycling behavior of the PVD-grown samples compared to the ALD ones shows an essentially different shape (see Fig. 7). However, the PVD_0sccm sample reveals a pronounced increase in the remanent ferroelectric polarization 2P_r during wake-up, which may be related to a high mobility of the oxygen vacancies within the HfO₂ layer, like de-pinning of domains pinned to the defect states. Both ALD samples show only a minor increase and a smaller slope of the 2P_r curves. Hence, this observation might be related to a decreased mobility of the VN₂ complex formed in the ALD-grown HfO₂ thin films compared to the oxygen vacancies V₀ present in PVD-grown HfO₂.

In summary, we presented an HAXPES study of TiN/HfO₂/TiN capacitors, for which the ferroelectric HfO₂ layers were either prepared by PVD sputtering or by ALD deposition. A comparative picture of the emerging chemical states dependent on the specific PVD or ALD growth conditions was obtained, in particular of the HfO₂ layers. The spectral signatures observed in the Hf 4f, O 1s, and N 1s core levels consistently suggest different mechanisms of oxygen vacancy formation in PVD- and ALD-grown HfO₂ thin films and, in the case of ALD, their electronic interaction with nitrogen impurities, which may directly correlate with the HfO₂ ferroelectric properties. By linking the HAXPES results to electric field cycling experiments on the TiN/HfO₂/TiN capacitors, a novel view on how the stabilization of the ferroelectric HfO₂ phase directly relates to specific PVD or ALD growth conditions is provided. Hence, our proposed model may inspire further studies on the complex electronic interactions between oxygen vacancies (V₀) and impurities (C or N) in ferroelectric HfO₂ and their impact on electric characteristics of TiN/HfO₂/TiN capacitors.

See the supplementary material for details about materials growth and crystallization as well as additional ferroelectric and HAXPES results.

This project received funding from the European Union's Horizon 2020 research and innovation programme under Grant Agreement No. 780302. We acknowledge DESY (Hamburg, Germany), a member of the Helmholtz Association HGF, for the provision of experimental facilities. Funding for the HAXPES instrument at beamline P22 by the Federal Ministry of Education and Research (BMBF) under Contract Nos. 05KS7UM1 and 05K10UMA with Universität Mainz; Nos. 05KS7WW3, 05K10WW1, and 05K13WW1 with Universität Würzburg is gratefully acknowledged.