Ferroelectric Hafnium Oxide Based Materials and Devices: Assessment of Current Status and Future Prospects

In recent years extensive research has been conducted on HfO₂ and ZrO₂ based thin films and related deposition methodologies. The renewed interest in thin films of these thoroughly investigated ceramic systems was mainly driven by scaling considerations in microelectronics engineering and the industry's striving to find a suitable "high-k" replacement for integrated gate and capacitor dielectrics. As part of this development and in accordance with earlier investigations of bulk ceramics and powders several methodologies were identified to influence the polymorphism of crystalline phases in HfO₂ and ZrO₂ based systems. The stabilization of the desired higher-k tetragonal or cubic phase against the monoclinic phase was found to depend on doping (Table I and references therein), thermal treatments,^3–5 deposition methodology/chemistry and growth temperature,^6,7 surface energy effects and strain given by grain size and film thickness,^4,5 as well as on the mechanical encapsulation provided by the electrodes.⁸

However, detailed structural as well as electrical investigations of Si-doped HfO₂ thin films conducted by Böscke et al.⁹ in 2011 revealed deviations from the commonly assumed polymorphism (monoclinic-tetragonal-cubic). This led to the rediscovery of an intermediate orthorhombic phase that was first and until then solemnly identified in Mg-doped, partially stabilized ZrO₂ by Kisi et al.¹⁰ in 1989. The ferroelectric nature of this non-centrosymmetric phase in HfO₂ thin films is confirmed by characteristic features in its hysteretic P-V, C-V and macroscopic (double beam laser interferometry, DBLI) as well as microscopic (piezoresponse force microscopy, PFM) electromechanical response.^9,2

The intrinsic nature of ferroelectricity in HfO₂ and therewith the correlation to its relative phase stability is supported by the systematic shown in Table I. Besides Si other known phase stabilizers in HfO₂, such as Y or Al, were later found to provoke ferroelectricity in a similar manner. During those investigations the precise stoichiometry control provided by atomic layer deposition proved highly valuable to detect the electrical and structural impact of small dopant quantities in HfO₂. This sequential and self-limiting growth technique allows for the insertion of sub-monolayer quantities of the dopant material during thin film growth by simply adjusting the cycling ratio of the host and the dopant precursor pulsing scheme. The linear incorporation of the various dopant species with variable cycling ration was detected by X-ray photoelectron spectroscopy and quantified by the standard free Rutherford backscattering spectroscopy.

Suitable, but with respect to ferroelectric properties until now untested dopants are additionally suggested in Table I. Based on the corresponding literature given in the table, these dopants are also known to influence phase stability in HfO₂ and therewith provide the basic requirement of destabilizing the monoclinic phase at room temperature. Due to this commonality with the already investigated systems, the appearance of an intermediate orthorhombic phase appears possible.

However, following this line of argumentation and considering the multiple parameters influencing phase stability in HfO₂ stated above, it is further suggested that ferroelectricity in this system is not restricted to doping alone. The recent demonstration of ferroelectricity in the HfO₂-ZrO₂ solid solution can be viewed as an example where surface energy differences between crystalline phases of HfO₂ and ZrO₂ can be utilized to achieve a similar kind of transition.²⁰ Consequently, using some of the additional parameters provided above, details of which will be published elsewhere, the stabilization of a ferroelectric phase in pure HfO₂ has proven possible (Fig. 1).

Zoom In Zoom Out Reset image size

Supportive of this argumentation are earlier²¹ as well as more recent^2,22 ab initio simulations that found that the relative stability of this orthorhombic phase in HfO₂ and ZrO₂ suggested by Kisi et al. lies within a narrow energy range of the commonly assumed polymorphism. Moreover, due to this small energy difference between polar and non-polar phases in HfO₂ and ZrO₂ it is assumed that the experimentally observed double hysteresis loop in samples close to the ferroelectric stability region can be attributed to a field driven, ferroelectric phase transition.^20,23

The coercive field strength E_c and remanent polarization P_r are the most important parameters to describe the hysteretic behavior and functionality of a ferroelectric material. In practice, however, additional parameters such as CMOS-compatibility, scalability, available thin film technologies and thermal budget requirement/tolerance are equally important and decisive upon the applicability of a ferroelectric to microelectronic engineering. In this context Table II summarizes key parameters of FE-HfO₂ and compares them to the common perovskite based ferroelectrics. Due to its stable ferroelectric properties at film thicknesses in the single digit nanometer range and straightforward implementation into a standard CMOS-environment, FE-HfO₂ emerges as an ideal material choice for high density integration of ferroelectric devices.

Special attention needs to be paid to the extraordinary high coercive field strength of FE-HfO₂. This rather unique property of FE-HfO₂ is crucial to the major scaling advantages as well as reliability disadvantages of this new ferroelectric material.

The scaling advantage is twofold. The high E_c of FE-HfO₂ allows for a gate length scaling of the ferroelectric field effect transistor (FeFET, see next section) while at the same time providing a high resilience toward internal depolarization of the ferroelectric. This internal depolarization is intrinsic to the metal-ferroelectric-insulator field effect transistor (MFIS-FET), but can also occur in metal-ferroelectric-metal (MFM) capacitors due to the so called "dead layer effect" at the electrode interfaces.³⁴ Considering such a paraelectric interfacial or depletion layer of a non-ideal electrode with a thickness d_PE and a permittivity ɛ_PE the resulting depolarization field E_dep is given by:

Assuming an invariable dead layer it can be deduced that with decreasing ferroelectric thickness d_FE the depolarization field increases. For simplicity and in favor of an underestimation of the depolarization field in the high permittivity perovskites, we assume . Based on Eq. 1 and as illustrated in Fig. 2 it can be readily understood that a high coercive field strength prevents depolarization even at decreased layer thickness and therewith provides stable ferroelectric properties. Consequently, Fig. 3 provides experimental proof that FE-HfO₂ preserves its ferroelectric properties down to a film thickness of 5 nm. It should be noted that this thickness does not yet represent a fundamental limit of ferroelectricity in hafnium oxide. At this point in time the issues that arise when decreasing thickness further are rather technological in nature. These are mainly leakage current and reliability related problems due to a change in the conduction mechanism, an increasing crystallization temperature and an increasing impact of electrode roughness in the MFM configuration. Alternative methods such as PFM might be required to access the fundamental thickness scaling limit of ferroelectricity in HfO₂.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

However, in terms of reliability and low voltage operation the high coercive field strength of FE-HfO₂ needs to be viewed rather critical. Table II introduces the ratio between the coercive field strength and the field necessary for a hard dielectric breakdown (HBD) as a new figure of merit. It can be seen that compared to its perovskite contenders the dielectric breakdown strength and the operating window for ferroelectric polarization hysteresis in FE-HfO₂ lies within the same order of magnitude. Consequently, it has been shown that an abrupt HBD rather than slower, defect mediated mechanism observed in classical ferroelectric fatigue are the major reliability concern in FE-HfO₂, whose MFM endurance is currently limited to about 10⁹ switching cycles.^30,31 As a consequence, strength and weakness provided by the high coercive field strength in FE-HfO₂ needs to be considered when designing appropriate devices and applications.

Classical FRAM is usually realized as a 1T-1C or 2T-2C memory cell. Like in the volatile DRAM the memory state of this cell is defined by the (polarization-) charge stored in a capacitor element, which can be accessed via a cell transistor (Fig. 4a–4b). However, when using a ferroelectric material instead of a linear capacitor dielectric this memory state becomes nonvolatile. To realize this bipolar ferroelectric switching, an additional plate line is added to the common cross point setup of word and bit lines known from DRAM.

Zoom In Zoom Out Reset image size

Besides this minor increase in complexity, FRAM and DRAM are quite alike, using comparable sensing schemes and consequently facing similar scaling challenges in terms of decreasing lateral area and therewith decreasing values of absolute (polarization-) charge per bit. In the case of high density standalone and most recent in the case of embedded DRAM (e.g. BEOL trench capacitors³⁵), this issue has been counteracted by extending the capacitor area into the third dimension. For FRAM to follow a similar pathway has proven rather challenging.³⁶ This is mainly due to the high complexity and limited thickness scalability of perovskite ferroelectrics (compare Table II). Considering FE-HfO₂ as a material choice for FRAM, these limitations can be overcome. Using a similar material combination like in the metal-insulator-metal (MIM) capacitor of advanced DRAM, an MFM capacitor can be implemented into high aspect ratio structures with minimal planar P_r penalty^30,2 (Fig. 4c). Additionally, the stability of the polarization state according to FRAM retention requirements for 10-year data storage has been proven at elevated temperature.³¹

Besides these promising results, it should be noted that due to the destructive read out in the 1T-1C FRAM concept not only the write, but also the read operation is limited by the cycling endurance of the ferroelectric material. Taking into account the comparably low endurance of FE-HfO₂ addressed in the previous section, the potential field of application appears severely limited. Due to the fundamental issue of the coercive field being close to HBD, considerable improvement of switching endurance and therewith read endurance beyond 10¹⁰ remains questionable.

However, considering the high endurance and low retention requirements of a NVM during power on and vice-versa for power off, a solution to this problem may lie in a hybrid mode, including volatile and nonvolatile memory operation. Due to the many similarities of HfO₂- and ZrO₂-based DRAM and FE-HfO₂ based FRAM, a manufacturable NV-DRAM solution can be realized. Unlike in conventional NV-DRAM the need to include a separate NVM-element for power off data storage into the DRAM cell becomes obsolete. As depicted in the simplified schematic in Fig. 5, the switch between volatile DRAM and nonvolatile FRAM operation mode is defined by the operating voltage alone. As summarized in Table III the volatile mode, defined by heavy memory usage, provides unlimited read endurance, whereas the nonvolatile mode provides sufficient retention for scheduled/unscheduled power off/loss.

Zoom In Zoom Out Reset image size

Compared to the capacitor-based FRAM the ferroelectric field effect transistor provides a smaller cell size and non-destructive read operation. Despite these clear advantages the commercialization of this technology is held back by scaling and CMOS-compatibility issues of the most common MFIS-FET approach. Due to the instability and contamination risk of perovskites in direct contact with silicon, large paraelectric buffer layers have to implemented, which, according to Eq. 1, significantly enhance the internal depolarization and reduce retention time.

Moreover, to achieve a sufficiently large memory window (ΔV_t) of the FeFET a large d_FE and/or E_c is required (MW ≈ 2 · E_C · d_FE, Fig. 4d). This leads to the dilemma that when using perovskite ferroelectrics, a relatively large physical height of the gate stack has to be implemented to compensate for small values of E_c (Table II). This increases processing complexity and hinders lateral scaling.

The large E_c and CMOS-compatibility of FE-HfO₂ solves both issues addressed above. The recent demonstration of FE-HfO₂ based FeFET at the 28 nm node (Fig. 4f) shows that this approach enables scaling alongside high-k metal gate transistor technology.³² This compatibility of an NVM concept to logic is unique and proves its potential as a high density embedded memory technology.

However, also in the case of the MFIS-FET, the ambivalence of the high E_c has to be considered. According to the band diagram depicted in Fig. 6c a large electric field in the insulator evolves when the FE-HfO₂ based FeFET is programmed to a saturated, high retention memory window (mode 1 in the MFIS-FET simulation depicted in Fig. 6a–6b). As a consequence severe charge injection into the gate stack leads to a device degradation after only 10⁴–10⁵ switching cycles.^32,2 It has been shown that, at the cost of retention, unsaturated programming conditions in combination with a thicker SiO₂ interface can be used to significantly enhance the endurance of the FE-HfO₂ based FeFET, reaching up to 10¹⁴ switching cycles.³⁷ Denoted as mode 2 in Fig. 6a–6b, this retention loss can be understood as a result of the higher E_dep and unsaturated polarization hysteresis loops, generally believed to be more prone to depolarization. Similar like in the case of the capacitor based 1T-1C FRAM this tradeoff between endurance and retention (Table III) requires an application-driven design of the memory concept. Both extremes, reaching from NOR-FLASH replacement to a capacitor-less, low refresh rate DRAM are within reach.

Zoom In Zoom Out Reset image size

Finally, considering the most primitive cell array architectures, i.e. AND, NOR and NAND, the most feasible option for FeFET cells is seen in the AND architecture (see Fig 4e). Assuming that each column in a memory array possesses a dedicated sourceline, the so called V_DD/3 operational scheme can provide disturb free operation as indicated in previous studies.³³ For a common source or NOR architecture, there are no sufficient transistor terminals available that could provide the necessary inhibit signals which are required for avoidance of polarization disturb. Moreover, for a NAND architecture the V_pass and V_program disturbs are considered quite critical, however, promising results for SBT-based FeFETs were e.g. presented in Ref. 24.