Electrical breakdown strength enhancement in aluminum scandium nitride through a compositionally modulated periodic multilayer structure

Artificial intelligence (AI) and the Internet of Things (IoT) have seen rapid advancements in recent years with the development of neuromorphic computing.¹ However, the development of more efficient neuromorphic computing systems is reaching a bottleneck due to bandwidth limitations between the processor and memory.² One way to bypass this limitation is to have non-volatile memory (NVM) monolithically embedded with the processor, significantly reducing the time and energy needed to move data between the processor and memory. Ferroelectric materials are promising for NVM because each unit cell has a non-volatile polarization state that can be written and read to form a memory bit.^3,4 However, the scaling of commercially available ferroelectric memories has stalled at the 130 nm node, largely due to the incompatibility of existing ferroelectric materials with Complementary Metal Oxide Semiconductor (CMOS) fabrication.⁵ In addition, many known ferroelectric materials, such as lead zirconate titanate (PZT), either do not have high enough remanent polarization to produce devices of competitive bit density or do not have high enough coercive field to offer good retention in high bit density devices.^6,7

Aluminum scandium nitride (AlScN) has many desirable properties that could make it useful as a non-volatile memory component. These include a large remanent polarization (80–135 μC/cm²),^8–14 a sharp ferroelectric transition, and deposition temperatures compatible with CMOS integration (<400 °C).¹⁵ However, current AlScN films have a low breakdown field to coercive field (E_BD/E_c) ratio, and thus devices made from the material have a small operating window when compared to other ferroelectric materials.^9,16,17 One method to increase the E_BD/E_c ratio is through the use of a multilayer architecture. Studies on multilayer systems in other ferroelectric materials such as BCT/BZT and PVDF/P(VDF-TrFE-CTFE) have shown that the increased interface density generally leads to an increase in E_BD.^18,19 The goal of the current study is to improve the E_BD and the E_BD/E_c ratios in AlScN by creating interfaces through deposition of multilayers.

In this work, single layer Al_0.72Sc_0.28N and Al_0.64Sc_0.36N were used as control samples, while three-layer, five-layer, and seven-layer multilayer structures were prepared via sequential deposition of Al_0.72Sc_0.28N and Al_0.64Sc_0.36N layers. Six-inch diameter Al(80 nm)/Al_0.80Sc_0.20N(85 nm)/Si wafers were used as substrates. The deposition conditions of these substrates are the same as those described in our prior work⁹ and include the deposition of a highly oriented (111) Al layer upon which to grow AlScN. After the Al(80 nm)/Al_0.80Sc_0.20N (85 nm)/Si substrates were prepared, 45 nm thick AlScN capacitor layers were sputter deposited under 150 kHz pulsed DC bias in an Evatec CLUSTERLINE^® 200 II physical vapor deposition system at 350 °C. The Al target power density was set to 12.7 W/cm², and the Sc target power density was 5.73 W/cm² for the 28% Sc layers and 8.34 W/cm² for the 36% Sc layers. The depositions were carried out with 20 sccm N₂ flow under 8.3 × 10⁻⁴ mbar chamber pressure. The calculated deposition rate was 0.25 nm/s for the 28% Sc layers, and the 36% Sc layer had an approximately 5% higher deposition rate. A capping layer of 30 nm thick Al was sputtered at 150 °C on top of the AlScN without breaking vacuum to prevent surface oxidation of the ferroelectric layer while also serving as the top electrode. Circular Al top electrodes were patterned with a single step of photolithography. An Oxford PlasmaPro 100 Cobra inductively coupled plasma (ICP) etcher was utilized to etch the top Al. The bottom electrodes of all the specimens were exposed by wet etching the AlScN film within 30 mol. % KOH solution.

The cross-sectional TEM sample was prepared in a Plasma-focused Ion Beam (TESCAN S8000X PFIB-SEM) system by using the in situ lift-out technique.²⁰ The sample was coated with electron beam and ion beam deposition of Pt protection layers to prevent charging and to protect the sample surface during FIB milling. A sample with parallel surfaces and a general thickness of ∼50 ± 5 nm was prepared. Scanning/transmission electron microscopy (S/TEM) characterization and image acquisition were carried out on a JEOL F200 TEM operated at 200 kV accelerating voltage. Image analysis and feature extraction were performed using ImageJ. All quantification results presented in this work were calculated with Digital Micrograph software (DM, Gatan Inc., USA).

Electrical measurements, including positive-up negative-down (PUND) measurements, electrical breakdown field (E_BD), capacitance, and dielectric loss tangent, were completed using a Keithley 4200A-SCS analyzer. The PUND measurements consisted of four monopolar trapezoidal pulses each with a rise/fall time of 20 ns, a pulse width of 500 ns, and an inter-pulse delay of 10 μs. The dielectric constant and loss tangent were acquired under −1 to 1 MV/cm DC bias with a tickle AC voltage of 30 mV at 10 kHz. The electric breakdown fields of the films were measured under a 10 kHz triangular DC voltage pulse peaking at 40 V and recorded for the voltage at which the Keithley reached a preset current compliance limit. TheWeibull analysis of the electric breakdown fields used 20 measurements from random positions on the wafer. Breakdown field, capacitance, and dielectric loss tangent tests for the AlScN films were acquired on 50 μm diameter circular top electrodes, while PUND measurements were acquired on 25 μm diameter circular top electrodes to minimize the RC time constant. All electrical tests were completed by voltage driving the capacitors from the bottom electrodes and current/charge sensing on the top electrodes.

A cross-sectional schematic of the single layer control samples and multilayer Al(30 nm)/AlScN (multilayer, 45 nm)/Al(80 nm)/Al_0.8Sc_0.2N(85 nm)/Si samples is presented in Figs. 1(a)–1(f). The ratios of Al_0.72Sc_0.28N and Al_0.64Sc_0.36N were kept the same, at a 3:2 volume ratio, for the multilayer samples. With a constant volume ratio, the findings in this study can be attributed to the number of interfaces and not to the composition or thickness of the samples. A cross-sectional bright-field TEM image of the five-layer sample was taken on the $⟨001⟩$ zone axis of the Si substrate and can be seen in Fig. 2(a). The individual multilayers cannot be clearly imaged due to the large strain contrast present in the image. However, the energy dispersive x-ray spectroscopy (EDS) maps presented in Fig. 2(b) show a clear difference in the relative Sc content in the layers. The interface between the device AlScN layer and the substrate/bottom electrode Al layer consists of a 2.7 nm oxide layer [Fig. 2(c)] for all samples. However, it is noteworthy that this oxide layer does not appear to affect the strong c-axis orientation of the AlScN layers, as seen in the diffraction patterns presented in Fig. 2(d). This finding is consistent with observations of other AlScN films on Al substrates exposed to air and shows that the self-limiting oxide formed on Al does not inhibit the local epitaxial growth of AlScN.⁹ 

Figure 3 presents the Weibull analysis of breakdown field in AlScN. Breakdown measurements were performed using only pristine, unswitched capacitors on the sample to avoid the influence of a wakeup effect, which have been shown to increase domain wall mobility in alternative ferroelectric materials such as HfO₂.^21,22 Breakdowns occurring below 15 V were attributed to hillock formation of the Al substrate and were not included in the data set, which consist of 20 breakdown strength measurements per sample. The characteristic breakdown field of the samples is determined using the Weibull distribution function,

where F is the cumulative probability of the electric failure, E is the experimental breakdown field, E_b is the characteristic breakdown field, which refers to the breakdown field when the cumulative failure probability is 63.2%, and β is a shape parameter, which evaluates the scatter or the consistency of the datapoints. The characteristic breakdown fields of the multilayer samples show a clear trend: breakdown field increases with the number of layers/interfaces. The characteristic breakdown fields of the single layer Al_0.72Sc_0.28N and Al_0.64Sc_0.36N were 5.99 and 5.86 MV/cm, respectively, while those of the three-layer, five-layer, and seven-layer were 6.40, 7.05, and 7.20 MV/cm, respectively.

The change in the characteristic breakdown field can be understood as the multilayer interfaces deflecting the propagation of the electrical tree, which are the gas-filled pathways for charge carriers that cause breakdown once the tree connects the top and bottom electrodes. Tree initiation occurs in localized regions where a higher electric field is produced by imperfections such as hillocks, cracks, and voids at the electrode/AlScN interface. The tree then propagates through partial discharge, which increases the local electric field and damages the materials surrounding the gas-filled tree channels.²³ In a multilayer structure, enhancement of breakdown strength can be explained in two ways. The first is the deflection of the electrical tree by defects at the multilayer interfaces. The second is that the electrical tree growth is retarded in the more compressive layers of the multilayer structure.²⁴ This can be observed in the d-spacing mapping of the seven-layer AlScN film section at the location labeled in Fig. 4(a). The line scan is taken through at least three layers of the multilayer film in Fig. 4(b). The interfaces play a key role in changing the in-plane strain through the thickness, which is in contrast to our prior results²⁵ for a single layer Al_0.64Sc_0.36N film, which simply shows the strain relaxing as the film growth becomes further away from the bottom electrode interface. By comparing the changes in d-spacing/c-axis lattice constant [Fig. 4(b)], we determine that the multilayer structure imposes relative compression and tension in alternating layers. Such enhancement of the breakdown strength in multilayer dielectrics has been found in several other dielectric systems both organic (polymeric)^26,27 and inorganic²⁸ as well as hybrid systems²⁹ with the enhancement mechanism attributed to the same underlying principle as discussed above.

It is also important to determine any impact the increased number of interfaces in the AlScN device layer can have on the ferroelectric properties. As seen in Table I and Fig. 5(a), remanent polarizations from 500 ns pulses measured via the PUND sequence measurements show no significant changes in P_r due to the architecture of the samples. In contrast, DC leakage, as measured from a voltage linear sweep, decreases as the number of layers increases [Fig. 5(b)]. In addition, there is an increase in the coercive field as the number of interfaces is increased, albeit at a smaller rate than the increase in the breakdown field. For the single layer Al_0.72Sc_0.28N and single layer Al_0.64Sc_0.36N, the average coercive field, (E_c+ − E_c−)/2, is 4.36 ± 0.03 and 4.66 ± 0.03 MV/cm. For the three-layer, five-layer, and seven-layer samples, the (E_c+ − E_c−)/2 is 4.88 ± 0.08, 5.0 ± 0.1, and 4.99 ± 0.03 MV/cm, respectively. This increase in the coercive field with an increasing number of layers appears to originate from the multilayer interfaces slightly adding to the driving force required to translate domain switching through the thickness of the samples. The increase in the coercive field with an increasing number of layers could arise from the regions of increased compressive strain at the interfaces between 28% and 36% AlScN layers, shown in Fig. 4(b), as higher compressive strain has been shown to increase the coercive field in AlScN materials.¹⁵ Similar increases in the coercive field have been observed in PZT as grain size decreased.³⁰ This results in E_BD/E_c ratios for single layer Al_0.72Sc_0.28N and single layer Al_0.64Sc_0.36N of 1.37 and 1.26, in comparison to that of the three-layer, five-layer, and seven-layer, which are 1.31, 1.41, and 1.44, respectively. As increased E_BD/E_c has been shown to improve the cycling endurance of alternative ferroelectric materials such as HfO₂,³¹ these results show promise that a multilayered architecture can overcome the endurance limitations of AlScN.⁷ 

In summary, we deposited AlScN multilayers comprised of Al_0.72Sc_0.28N and Al_0.64Sc_0.36N onto orientated Al/AlScN substrates with a constant volume of each composition. The multilayer films were compared to the single layered Al_0.72Sc_0.28N and Al_0.64Sc_0.36N and exhibited minimal changes in remanent polarization and lower leakage currents. Most importantly, the characteristic breakdown field determined by the Weibull analysis showed an increase from 5.99 and 5.86 MV/cm in single layered Al_0.72Sc_0.28N and Al_0.64Sc_0.36N to as high as 7.20 MV/cm in the seven-layered AlScN. In addition, more interfaces resulted in a higher breakdown field to coercive field ratio. These results in conjunction with lattice parameters from a d-space mapping suggest that the enhancement of the breakdown field originates from either the multilayer interfaces deflecting the propagation of the electrical tree or relative compression in the alternating layers of the samples reducing the propagation rate of electrical trees. This increase in breakdown field through the use of device architecture will allow AlScN devices to be ferroelectrically switched at an electric field further from the breakdown field and thereby making such devices more reliable.

This work was supported in part by Semiconductor Research Corporation (SRC) and in part by the Defense Advanced Research Projects Agency (DARPA) Tunable Ferroelectric Nitrides (TUFEN) Program under Grant Nos. HR00112090046 and HR00112090047. The authors would like to thank Dr. Giovanni Esteves of Sandia National Laboratories for providing the Al template substrates. The deposition, patterning, and characterizations of AlScN were performed at the Singh Center for Nanotechnology and supported by the NSF National Nanotechnology Coordinated Infrastructure Program (No. NNCI-1542153). The authors acknowledge the use of facilities and instrumentation supported by NSF University of Pennsylvania Materials Research Science and Engineering Center (MRSEC) (No. DMR-1720530).

The data that support the findings of this study are available from the corresponding author upon reasonable request.

The authors have no conflicts to disclose.