Ferroelectricity in hafnium oxide

In the late 1990s, hafnium oxide became one of the most prominent high-k (k represents the dielectric constant or permittivity) materials to replace the well-established silicon dioxide in metal oxide semiconductor field-effect transistors. Hafnium oxide has reasonably high permittivity and bandgap, and is stable on silicon.^{Reference Robertson14} In terms of minimum leakage, the amorphous phase was preferred in transistors that became available in 2007.^{Reference Bohr, Chau, Ghani and Mistry15} However, crystalline phases of HfO₂ (see Figure 1e) have higher permittivity than the amorphous phase. Among these, the monoclinic (space group: P2₁/c), the tetragonal (space group: P4₂/nmc), and the cubic (space group: Fm$\bar{3}$m) phases can be formed under reasonable conditions. They have permittivities of 22, 46, and 36, respectively (Figure 1e).^{Reference Materlik, Künneth and Kersch16} In bulk material, the monoclinic phase is the stable phase at room temperature and 1 atm, and it transforms to the tetragonal and cubic phase at 1973 and 2773 K, respectively.

The tetragonal or cubic phase induced by doping in hafnium oxide has been studied intensively. Boescke first observed a clear fingerprint of ferroelectricity in silicon doped hafnium oxide samples with doping concentrations in the few percent range.^{Reference Böscke, Müller, Bräuhaus, Schröder and Böttger17,Reference Böscke, Teichert, Bräuhaus, Müller, Schröder, Böttger and Mikolajick18} Since this discovery was unexpected, researchers initially used different evaluation techniques to experimentally prove that the ferroelectric hysteresis does not originate from artifacts such as leakage^{Reference Scott19} or charge trapping. Piezoelectricity^{Reference Martin, Müller, Schenk, Arruda, Kumar, Strelcov, Yurchuk, Müller, Pohl, Schröder, Kalinin and Mikolajick20} and pyroelectricity^{Reference Hoffmann, Schroeder, Künneth, Kersch, Starschich, Böttger and Mikolajick21,Reference Smith, Kitahara, Rodriguez, Henry, Brumbach and Ihlefeld22} have been demonstrated as well. Additionally, it was shown that different dopants such as Si,^{Reference Böscke, Müller, Bräuhaus, Schröder and Böttger17} Y,^{Reference Mueller, Schröder, Böscke, Müller, Böttger, Wilde, Sundqvist, Lemberger, Kücher, Mikolajick and Frey23} Al,^{Reference Mueller, Mueller, Singh, Riedel, Sundqvist, Schroeder and Mikolajick24} Zr,^{Reference Müller, Böscke, Bräuhaus, Schröder, Böttger, Sundqvist, Kücher, Mikolajick and Frey25} and others^{Reference Schroeder, Yurchuk, Müller, Martin, Schenk, Polakowski, Adelmann, Popovici, Kalinin and Mikolajick26,Reference Xu, Shibayama, Izukashi, Nishimura, Yajima, Migita and Toriumi27} can induce ferroelectricity. Even in undoped HfO₂,^{Reference Polakowski and Müller28} it is possible to achieve ferroelectricity by further reducing the film thickness and choosing proper processing conditions. From the beginning, it was speculated that the noncentrosymmetric orthorhombic Pca2₁ phase could be responsible for this behavior.^{Reference Böscke, Müller, Bräuhaus, Schröder and Böttger17,Reference Böscke, Teichert, Bräuhaus, Müller, Schröder, Böttger and Mikolajick18}

However, with standard methods such as x-ray diffraction, a bulletproof verification of the orthorhombic phase in a mixture of monoclinic and tetragonal phases could not be achieved. Sang et al.^{Reference Sang, Grimley, Schenk, Schroeder and LeBeau29} used position averaged convergent beam electron diffraction analysis to prove that the orthorhombic Pca2₁ phase is present in ferroelectric hafnium oxide. Using this knowledge in modeling makes identification of phase fractions more convenient.^{Reference Park, Schenk, Fancher, Grimley, Zhou, Richter, LeBeau, Jones, Mikolajick and Schroeder30} Nonetheless, the factors leading to the formation of the orthorhombic phase are still not clear. Many factors such as stress, grain size, and doping may play a role. Materlik et al.^{Reference Materlik, Künneth and Kersch16} proposed that surface energy can be the most important factor, based on the fact that the gain in free energy of these factors is not sufficient to explain the stabilization of the orthorhombic phase and why the stabilization works much better in thinner films.^{Reference Yurchuk, Müller, Knebel, Sundqvist, Graham, Melde, Schröder and Mikolajick31} Park et al.^{Reference Park, Lee, Kim, Schenk, Lee, Kim, Fengler, Mikolajick, Schroeder and Hwang32} recently compared this model to experimental data in hafnium zirconium oxide. It was found that the basic experimental trends with respect to film composition confirm the model, but the best composition to achieve ferroelectric behavior is shifted to a different Hf:Zr ratio. Moreover, the observed increasing monoclinic phase fraction with increasing annealing temperature could not be explained.^{Reference Park, Kim, Moon, Kim, Lee, Hyun, Mikolajick, Schroeder and Hwang33} This points to the conclusion that not only thermodynamics, but also kinetic effects, need to be considered to explain the physical origin of the formation of the orthorhombic phase.^{Reference Park, Kim, Moon, Kim, Lee, Hyun, Mikolajick, Schroeder and Hwang33}

Although not all controlling parameters are understood in detail, it is now established that ferroelectricity can be reproducibly achieved in thin hafnium oxide-based films in the thickness range of 5–30 nm, thus it is suitable for further device development. As shown in Table I, ferroelectric hafnium oxide doped with various dopants has one order of magnitude higher E _c and similar polarization compared to ferroelectric perovskites. E _c is even higher than that of the polymer ferroelectric poly(vinylidene fluoride). Additionally, the permittivity is much lower than in perovskites, which is an important aspect in reducing the depolarization field in the FeFET.

Table I.

Comparison of coercive field, E _c, and switched polarization charge, 2P _R, for strontium bismuth tantalate (SBT), lead zirconium titanate (PZT), poly(vinylidene fluoride):tetrafluoroethylene (PVDF-TRFE), and doped hafnium oxide.

The values given for the doped hafnium oxide span the range that have been observed using different dopants such as Si, Al, Y, La, as well as mixed crystals from Hf and Zr.

FeRAMs

The most straightforward way to realize a ferroelectric memory is to use the DRAM architecture, replacing the dielectric by a ferroelectric and pulsing the plate line instead of keeping it grounded (see Figure 2a). This modification is necessary since in the ferroelectric capacitor, compared to DRAM where the electrical field is supplied by the charged capacitor, the storage is maintained at 0 V and the capacitor needs to be actively driven via the plate line to read the stored polarization (see Figure 1d). To increase the signal, in the early days, two such cells, where one is written into the direction that should be stored (e.g., 0) and the other is written into the opposite direction (e.g., 1), were combined to realize one bit. Later, the second cell was replaced by one reference cell for a large number of cells in order to save space. The first products appeared on the market as early as 1993,^{Reference Bondurant5} and today, advanced products use 130-nm technology.^{Reference McAdams, Acklin, Blake, Du, Eliason, Fong, Kraus, Liu, Madan, Moise, Natarajan, Qian, Qiu, Remack, Rodriguez, Roscher, Seshadri and Summerfelt10} However, further success beyond niche markets is hindered by the fact that a 3D capacitor has not been practically realized using perovskite materials.^{Reference Koo, Seo, Kim, Shin, Lee, Baik, Lee, Lee, Bae, Lim, Yoo, Park, Kim, Han, Baik, Choi, Park and Park8,Reference Yeh, Lisker, Kalkofen and Burte9}

With ferroelectric hafnium oxide, the full advantages of 3D integration can be attained^{Reference Polakowski, Riedel, Weinreich, Rudolf, Sundqvist, Seidel and Müller34} (see Figure 2b). However, now, the high coercive field of ferroelectric hafnium oxide is the limiting factor. On the one hand, device operation at low operating voltages is challenging. On the other hand, field cycling uses fields close to the breakdown value, limiting the device endurance with respect to the number of possible switching cycles.^{Reference Yurchuk, Mueller, Martin, Slesazeck, Schroeder, Mikolajick, Müller, Paul, Hoffmann, Sundqvist, Schlosser, Boschke, van Bentum and Trentzsch35} Note that due to the destructive read in the FeRAM concept, read cycles also contribute to the total number of endurance cycles. Moreover, effects like wakeup (increase in remanent polarization during the early stages of cycling) and fatigue (reduction in remanent polarization for high cycle numbers) still need to be engineered to levels as those found in perovskite ferroelectrics.^{Reference Schenk, Hoffmann, Ocker, Pešić, Mikolajick and Schroeder36,Reference Pešić, Fengler, Larcher, Padovani, Schenk, Grimley, Sang, LeBeau, Slesazeck, Schroeder and Mikolajick37} The issue of operating at low supply voltage can be tackled by engineering the coercive field. To date, the fabrication parameters have been observed to have only weak influence on the coercive field. An oxide thickness as low as 5 nm was fabricated with good electrical parameters.^{Reference Müller, Polakowski, Mueller and Mikolajick38} Attempts at further lowering the thickness were also made,^{Reference Chernikova, Kozodaev, Markeev, Negrov, Spiridonov, Zarubin, Bak, Buragohain, Lu, Suvorova, Gruverman and Zenkevich39} although it is challenging to verify the ferroelectricity in the case that leakage becomes dominant.^{Reference Schenk, Schroeder and Mikolajick40}

Our understanding of endurance has also significantly progressed in the last three to four years,^{Reference Schroeder, Pešic, Schenk, Mulaosmanovic, Slesazeck, Ocker, Richter, Yurchuk, Khullar, Müller, Polakowski, Grimley, LeBeau, Flachowsky, Jansen, Kolodinski, van Bentum, Kersch, Künneth and Mikolajick41} though much can still be learned from the optimization of lead zirconium titanate.^{Reference Fengler, Pešić, Starschich, Schneller, Künneth, Böttger, Mulaosmanovic, Schenk, Park, Nigon, Muralt, Mikolajick and Schroeder42} Recently, another approach has been proposed to increase the endurance in the 1-transistor-1-capacitor (1T-1C) architecture. When stabilizing ferroelectricity in hafnium oxide by dopants with radius smaller than that of Hf and at doping concentrations higher than the optimum to achieve ferroelectricity, a pinched or antiferroelectric hysteresis is observed.^{Reference Böscke, Müller, Bräuhaus, Schröder and Böttger17,Reference Mueller, Mueller, Singh, Riedel, Sundqvist, Schroeder and Mikolajick24,Reference Schroeder, Yurchuk, Müller, Martin, Schenk, Polakowski, Adelmann, Popovici, Kalinin and Mikolajick26,Reference Yurchuk, Müller, Knebel, Sundqvist, Graham, Melde, Schröder and Mikolajick31} (see Figure 2c; dotted line). Specifically, in the hafnium-zirconium-oxide system, this behavior is even observed for pure zirconium oxide.^{Reference Müller, Böscke, Schröder, Mueller, Bräuhaus, Böttger, Frey and Mikolajick43} The cycling endurance for such a structure is much higher compared to that of the ferroelectric case.^{Reference Pešić, Knebel, Hoffmann, Richter, Mikolajick and Schroeder44} However, since there is no remanent polarization, P _R, when the electrical field is removed, the system is not nonvolatile. Pesic et al. proposed the use of a bias field that can be generated by either electrodes with different work-function values^{Reference Pešić, Knebel, Hoffmann, Richter, Mikolajick and Schroeder44,Reference Pešić, Hoffmann, Richter, Mikolajick and Schroeder45} or additional fixed charges or dipoles placed inside the dielectric stack.^{Reference Pešić, Hoffmann, Richter, Slesazeck, Kämpfe, Eng, Mikolajick and Schroeder46} They verified this possibility using capacitor test structures.^{Reference Pešić, Knebel, Hoffmann, Richter, Mikolajick and Schroeder44–Reference Pešić, Hoffmann, Richter, Slesazeck, Kämpfe, Eng, Mikolajick and Schroeder46} The advantage of this approach is that now the cycling is performed between a polarized and an unpolarized state. As a result, the apparent coercive field is cut in half, and the endurance is drastically improved. Antiferroelectric hysteresis was also observed in dielectric stacks similar to those used in state-of-the-art DRAM capacitors,^{Reference Pešić, Hoffmann, Richter, Mikolajick and Schroeder45} making this a possible extension of today’s DRAM technology toward a nonvolatile RAM memory.^{Reference Pešić, Hoffmann, Richter, Slesazeck, Schroeder and Mikolajick47} Thus, ferroelectric hafnium oxide can overcome the scaling limitation of FeRAMs since mature atomic layer deposition processes exist that allow deposition of material on high-aspect-ratio capacitors. However, the higher E _c imposes additional challenges. Material engineering or an alternative approach using the antiferroelectric hysteresis are possible options to overcome these limitations. To date, no fully integrated 1T-1C FeRAM array structures have been reported and research has been mainly carried out on capacitor test structures.

FeFETs

The main issues of a 1T-1C memory cell can be overcome if the ferroelectric is integrated into the gate stack of a FeFET (Figure 3a), as explained in the Introduction section.^{Reference Ross11} Assuming an n-channel transistor and high enough polarization that if oriented in a way that the negative charge is close to the channel and the positive charge is close to the gate electrode, the channel will be in accumulation mode. We can also assume that if the polarization is the other way around, then the channel will be in the inversion condition. If we start in the OFF state and the gate voltage is swept to a point where the positive coercive field of the ferroelectric is reached and finally exceeded, the channel will turn on. When sweeping back beyond the point where the negative coercive field is reached, the channel will be switched OFF again. Therefore, if the positive coercive field E _c+ and negative coercive field E _c– (see Figure 1c) have the same absolute value E _c, the memory window (MW) (the difference between the two threshold voltages of the device for the two different polarization states) is given by the difference between the voltages where the respective coercive fields are reached:^{Reference Miller and McWhorter48}

(1)$$MW = 2 \cdot E_{\text{c}} \cdot d_{\text{Fe}} .$$

Here, d _Fe is the thickness of the ferroelectric film.

It should be noted that Equation (1) was deduced using a simplified picture. In practice, deviations from the previously discussed assumptions and the depolarization fields discussed later need to be considered. However, the consequence that the maximum attainable memory window is limited by the coercive field holds true and is important. Hence, the high coercive field of hafnium oxide ferroelectrics is actually a benefit rather than a drawback.

Using perovskites, much thicker ferroelectric films would be needed to achieve the same memory window, which limits scalability. Only for remanent polarization of the ferroelectric that is too low to cause sufficient electrical field can we expect a significant influence of the remanent polarization on the memory window.^{Reference Miller and McWhorter48} Eventually, the remanent polarization will influence the effective gate overdrive of the device, and therefore, will influence the ON current.^{Reference Miller and McWhorter48} Because of threshold voltage shifts during polarization switching, a lower subthreshold swing, compared to a nonferroelectric transistor that has the same effective gate dielectric thickness, can be achieved. The subthreshold swing is the voltage swing that is required to change the transistor current by one order of magnitude and describes the ability to efficiently switch the transistor between the ON and OFF states. However, if the effect is not a consequence of a stabilized negative capacitance effect,^{Reference Salahuddin and Datta49} the polarization reversal would lead to a switching hysteresis.

In the transistor, a serial capacitor consisting of not only the capacitance of the depletion layer in the semiconductor, but also the unavoidable interface oxide between the channel and the ferroelectric, is an integral part of the structure. This series connection means the voltage that has to be applied to the device for polarization switching is increased compared to the case of a pure ferroelectric capacitor, and in the case of no applied electrical field (e.g., the case of retention), an increased internal depolarization field will be the consequence.^{Reference Ma and Han12} Since perovskite-based ferroelectrics have permittivity in the range of a few hundreds, a large thickness is required to balance this inherent capacitive divider. The much lower permittivity of hafnium oxide is beneficial here, and additionally, the high coercive field helps to stabilize the polarization in the retention case.^{Reference Gong and Ma50} Therefore, hafnium-based ferroelectrics seem to be favorable to realize FeFETs.

The first demonstrations occurred in 2011.^{Reference Böscke, Müller, Bräuhaus, Schröder and Böttger51,Reference Müller, Boescke, Schroder, Hoffmann, Mikolajick and Frey52} In 2012, it was verified that use of ferroelectric hafnium oxide can close the scaling gap between FeFETs and conventional FETs.^{Reference Müller, Yurchuk, Schlösser, Paul, Hoffmann, Müller, Martin, Slesazeck, Polakowski, Sundqvist, Czernohorsky, Seidel, Kücher, Boschke, Trentzsch, Gebauer, Schröder and Mikolajick53} Encouraged by this work, in 2016, the first fully integrated technology, where FeFET memory arrays were embedded in a 28-nm CMOS process, was demonstrated^{Reference Trentzsch, Flachowsky, Richter, Paul, Reimer, Utess, Jansen, Mulaosmanovic, Müller, Slesazeck, Ocker, Noack, Müller, Polakowski, Schreiter, Beyer, Mikolajick and Rice54} (Figure 3b). One year later, scaling to 22-nm fully depleted silicon-on-insulator technology together with further advances in the performance of the memory arrays were shown.^{Reference Duenkel, Trentzsch, Richter, Moll, Fuchs, Gehring, Majer, Wittek, Müller, Melde, Mulaosmanovic, Slesazeck, Müller, Ocker, Noack, Löhr, Polakowski, Müller, Mikolajick, Höntschel, Rice, Pellerin and Beyer55}

Thus, the vision of realizing nonvolatile memories based on FeFETs has moved to realization much faster than in the previous 30 years of research. However, memories are not the only application for FeFETs. Salahuddin and Datta^{Reference Salahuddin and Datta49} proposed the concept of negative capacitance transistors in 2008 to overcome the so-called Boltzmann tyranny. This means that the subthreshold slope of FETs are limited to 60 mV/decade at room temperature, therefore, scaling of the supply voltage is limited. This concept makes use of stabilizing the ferroelectric in the negative differential capacitance region between the two stable polarization states by using a suitable dielectric capacitor in series. The negative differential capacitance region is caused by an energy barrier giving rise to a negative change in charge with changing voltage (see Figure 1b). Eventually, the stabilized negative capacitance will lead to a voltage boost on the internal node between the ferroelectric and the transistor dielectric. Consequently, a smaller subthreshold swing together with a higher ON current is possible.^{Reference Lin, Khan, Salahuddin and Hu56}

Again, the integration challenge of combining ferroelectrics with CMOS processes needs to be overcome and ferroelectric hafnium oxide seems to be the material of choice for such devices. Initial device demonstrations have been shown,^{Reference Li, Chen, Lai, Lin, Cheng, Chen, Wei, Hou, Liao, Lee, Chen, Sheih, Yeh, Yang, Salahuddin and Hu57,Reference Lee, Fan, Tang, Chou, Chen, Kuo, Xie, Liu, Liao, Jong, Li, Chen and Liu58} although the issue of stabilization of the negative capacitance state remains to be resolved. As a first step, the transient negative capacitance was observed.^{Reference Hoffmann, Pešić, Chatterjee, Khan, Salahuddin, Slesazeck, Schroeder and Mikolajick59} However, the exact boundary conditions for stabilizing the negative capacitance operation region are still under scientific debate and need clarification.^{Reference Hoffmann, Pešić, Slesazeck, Schroeder and Mikolajick60–Reference Kim, Yamada, Moon, Kwon, An, Kim, Kim, Lee, Hyun, Park and Hwang62}

Summary and future prospects

From a physical viewpoint, ferroelectrics are an almost ideal candidate for binary nonvolatile memories, since they have a field-driven switching mechanism. However, the complexity of the traditional perovskite materials and the resulting difficulties to integrate them into the CMOS process has hindered the fast scaling of such devices. The unexpected discovery of ferroelectricity in doped hafnium oxide has completely changed this picture, because hafnium oxide is a standard material in CMOS processing, with manufacturing processes available for different scenarios.

From an electrical parameter point of view, ferroelectric hafnium oxide differs from perovskites in having a much higher coercive field, E _c, and a much lower permittivity. When using the material in the traditional 1T-1C ferroelectric RAM configuration, the high coercive field imposes challenges with respect to low-voltage operation and cycling endurance. Antiferroelectric memory has been proposed to overcome these limitations. When integrating ferroelectric hafnium oxide into the gate stack of a transistor, resulting in a FeFET, both aspects give this material a competitive edge. The developmental progress of FeFETs has significantly increased by the adoption of ferroelectric hafnium oxide.

Nevertheless, achieving a cycling endurance beyond the level of conventional charge-based nonvolatile memories remains a challenge.^{Reference Chernikova, Kozodaev, Markeev, Negrov, Spiridonov, Zarubin, Bak, Buragohain, Lu, Suvorova, Gruverman and Zenkevich39,Reference Yurchuk, Müller, Müller, Paul, Pešić, van Bentum, Schroeder and Mikolajick63} Different strategies have been proposed to overcome these limitations^{Reference Mueller, Polakowski, Muller, Mulaosmanovic, Ocker, Mikolajick, Slesazeck, Flachowsky and Trentzsch64} and encouraging results have recently been published.^{Reference Chatterjee, Kim, Karbasian, Tan, Yadav, Khan, Hu and Salahuddin65} The fact that differently optimized FeFETs are also an option to make steep subthreshold devices by stabilizing the negative capacitance region has further inspired the research and development of such devices. Finally, ferroelectricity is a material property that is accompanied by piezoelectricity and pyroelectricity.^{Reference Mitsui, Martienssen and Warlimont1,Reference Martin, Müller, Schenk, Arruda, Kumar, Strelcov, Yurchuk, Müller, Pohl, Schröder, Kalinin and Mikolajick20,Reference Smith, Kitahara, Rodriguez, Henry, Brumbach and Ihlefeld22} These properties open the path toward integrated sensors, actuators, and energy-harvesting functionalities, and these are only in the early stages of exploration.^{Reference Hoffmann, Schroeder, Künneth, Kersch, Starschich, Böttger and Mikolajick21,Reference Park, Kim, Kim, Moon, Kim and Hwang66,Reference Park, Schenk, Hoffmann, Knebel, Gärtner, Mikolajick and Schroeder67}