Benedikt
Kersting
and
Martin
Salinga
*
Institute of Physics (IA), Physics of New Materials, RWTH Aachen University, Sommerfeldstr. 14, Aachen 52074, Germany. E-mail: martin.salinga@rwth-aachen.de
First published on 7th November 2018
The market launch of Intel’s 3D XPoint™ proves phase change technology has grown mature. Besides storing information in a fast and non-volatile way, phase change memories (PCMs) may facilitate neuromorphic and in-memory computing. In order to establish PCM as a lasting element of the electronics ecosystem, scalability to future technology nodes needs to be assured. Continued miniaturization of PCM devices is not only prescribed in order to achieve memories with higher data density and neuromorphic hardware capable of processing larger amounts of information. Smaller PCM elements are also incentivized by the prospect of increased power efficiency per operation as less material needs to be heated up for switching. For this reason, a good understanding of the effects of confinement on phase change materials is crucial. Here we describe how miniaturization increases the importance of interface effects and we show how in consequence the crystallization kinetics of phase change materials, when confined into nanometer sized structures, can change significantly. Based on this analysis, the implications of such nanoscale effects are discussed and possible ways of exploiting them proposed.
Attempts to leap forward in scaling the contact cross-section to the nanometer range, e.g. by using carbon nanotubes as electrodes, imply that the effects of scaling may indeed continue along those trends that have been observed in several device generations before.5-6 However, when the actual volume of phase change material is confined more and more narrowly, even properties supposedly characteristic for a material begin to change. Most noticeably, both structural phase transitions, melting and crystallizing, have been reported to take place at significantly different temperatures than in bulk (for this purpose that means in 100 nm thin films).4,7-9
Often in the past, size effects were observed without having sealed the nanostructures with cladding. Noé et al.10 recently demonstrated how easily exposure to an oxygen atmosphere can lead to crucial oxidation of the surface of typical phase change materials and drastically change the crystallization behavior as a consequence of the compositional change. Conversely, careful in situ capping, even with silicon-oxide, can keep the phase change material’s composition unaltered.10 When introducing a certain cladding, however, the material choice itself turns out to have an influence on the crystallization properties.
Traditionally, the interfacial energies between the cladding and the different phases of the phase change material have been seen as the origin for heterogeneous nucleation at the interface, which can largely dominate over homogeneous nucleation within the disordered phase change material, especially when the interface-to-volume ratio is large. Even crystal growth along such an interface can be sped up or slowed down compared to its bulk values depending on the interfacial energies.11
In general, crystal growth becomes more dominant in the crystallization process of nanometric phase change memory cells, not only compared to homogeneous, but even to heterogeneous nucleation. As soon as a crystal-to-glass interface exists only a few nanometers away from any group of disordered atoms, they are much more likely to be incorporated into that growing crystal than they are able to reorganize themselves into a new crystal nucleus first.
Besides reduced nucleation probability in a smaller volume and the altered heterogeneous nucleation at the interface, the confinement of a material can have an impact on the ease with which the atoms can reconfigure within a nanometric volume. Researchers at Yale nicely discussed these competing effects when studying crystallization of metallic glass nanorods of different sizes.12–14 The apparent viscosity is reduced as the confinement is narrowed. As a consequence, crystallization kinetics are slowed down in smaller structures. In the field of phase change materials, this idea has also been proposed to interpret the effects of confinement.11
Others (e.g.ref. 15) put crystallization-induced stress into the center of their interpretation of what happens upon confinement, acknowledging the significant difference in mass density between disordered and crystalline phases. Due to the inability of atoms to freely rearrange along a surface in the presence of a capping, they argue that the necessary viscous flow is more strongly coupled with crystal growth.
A different argument is made by Simpson et al.16 when interpreting their observation of increased crystallization temperature in thinner films confined by an encapsulation material that exerts compressive stress on the initially amorphous phase change material. Based on the observation of vacancies in crystalline phase change materials, they believe that under too high a compressive stress the material might not have sufficient space to realize such a state. In contrast, Eising et al.17 have demonstrated convincingly how compressive stress accelerates crystal growth, which seems intuitive in view of the higher mass density of the crystalline state.
While stress certainly does have an influence on the crystallization kinetics, it is less clear how important this is compared to other effects like the increased viscosity due to confinement or parameters controlling it. In a series of molecular-dynamics simulations Scheidler et al.18,19 investigated cooperative motion in supercooled liquids in close proximity of confining walls. The roughness or smoothness of these surfaces turned out to play a decisive role, pointing to a physical quantity that might turn out to be a relevant specifier for comparing different claddings in the future.
Several of the above-mentioned aspects must be expected to depend (quantitatively) on how the amorphous phase change material and its confinement was created: by deposition from the gas phase and subsequent addition of cladding material or by quenching the melt within its confinement.20 The literature quantifying the crystallization kinetics of phase change materials for melt-quenched amorphous materials is very limited. When it comes to confinement effects, there is an even stronger need for research on the technologically relevant, melt-quenched state.
While subject to the exact choice of cladding material, the general trend in phase change materials is an increased stability of the disordered states, i.e. slowed-down crystallization kinetics, upon confinement. If one does not want to end up with switching speeds and retention times orders of magnitude too high for useful applications, one may not continue to use the same materials that have been designed to work in significantly larger memory cells. Instead, when scaling towards the size of very few nanometers one should consider materials which crystallize much too quickly in larger volumes. They might even be difficult to be amorphized in bulk. Pure antimony fulfills this criterion. In the following we will describe and analyze experiments in which nano-confined phase change memory cells based on pure Sb have been electrically characterized.
In particular, a precise control of transient signals turned out to be crucial (see the section “Limits of programmability” and Fig. 4a). To gain deeper insight into the transient current passing the linecell we performed circuit simulations including setup and sample with Simulink (Simscape). The capacity of the contact pads (60 × 90 mm2, dSiO2 138 nm, εr = 3.9) is estimated to be around 1.42 pF. The series resistor, which is included in order to limit the current passing through the phase change device after electrical breakdown (e.g. upon threshold switching), is a combination of resistor (2300 Ω) and capacitor (∼2378 mm2; dSiO2 = 58 nm; εr = 3.9; 1.35 pF). It is represented as a 1.2 mm long transmission line (C = 1.125 pF mm−1; R = 1917 Ω mm−1). The phase change material in the crystalline state is assumed to behave like a purely resistive element. Coaxial transmission lines leading to the probe-head in contact with the device are simulated with a capacity of 90 pF m−1.
There is a reasonable match between simulated and measured transient signals (Fig. 2c and d). The shape of the 50 ns pulses with edges as short as 3 ns is well preserved. The simulation allows not only the reproduction of the experimental data but also the study the transient signal passing the linecell. It turns out that the trailing edge of the pulse arriving at the linecell depends on which side the signal is applied from (Fig. 2b). If the signal is applied to the series resistor side of the sample, effectively an RC-filter is placed in front of the linecell. Thus, the voltage drop across the phase change material changes faster if the electrical signal is applied to the linecell and the series resistor is behind it.† We performed all experiments in this configuration.
Samples with 10 nm Sb are reset to a high resistive state with trapezoidal voltage pulses (width 50 ns, edges 3 ns). During the plateau of the pulse the device resistance (Rtransient;Scope) is lower than the set resistance (Fig. 3a). The same is true for the resistivity of molten antimony [∼114 μΩ cm]24,25 compared to that of crystalline thin films [rthinfilm(d < 50 nm) > 150 μΩ cm].26 At low temperatures (100–225 K) the resistance of antimony linecells in the crystalline state was found to be almost temperature independent.21 The resistance decrease from Rset = 1850 Ω to Rtransient = 1375 Ω gives a first hint of the applied power being sufficient to melt the phase change material. After the reset pulse the device ends up in a high resistive state (∼35 kΩ). Triangular pulses (200 ns leading and trailing edges) are used to switch the linecell back to the low resistive state. The device can be switched reliably with an ON/OFF ratio of around 20. A significantly increased resistance compared to that of the crystalline state is a first characteristic feature of an amorphized device.
The set process from the high resistive (amorphous) state to the thermodynamically stable crystalline state within a few nanoseconds requires substantial Joule heating to temperatures above the glass transition temperature. It is possible because the amorphous material switches to a low resistive state (on-state) at a threshold voltage.27 Threshold switching is of fundamental importance to operate phase change materials in an electronic memory. The current–voltage characteristics recorded during the set pulse suggest threshold switching between 0.4 V and 0.5 V, indicated by a pronounced kink in the current–voltage curve (Fig. 3b). The following rise of the device current is the result of growing Joule heating after threshold switching under a still increasing applied voltage. However, a gradual transition from high to low resistive states could also appear if a thin crystalline filament remained in the largely amorphized linecell. Hence, the 10 nm thick Sb possibly cannot be quenched fast enough to create an amorphous mark that blocks the complete cross-section of the linecell.
Resistance drift is another distinct characteristic of amorphous phase change materials. The resistivity increases with time as the melt-quenched material relaxes into energetically more favorable glassy states.28,29 The temporal evolution of resistance drift at constant temperature is commonly described by the equation R(t) = R(t0) × (t/t0)ν, with a constant drift exponent ν. A drift measurement over 3500 s shows a continuous resistance increase (Fig. 3c). At 100 K ambient temperature the size of the amorphous mark is stable and does not show indications of crystallization. The drift coefficient of 0.1 ± 0.02 is remarkably similar to that observed for other phase change materials at room temperature and above.30
Programming curves [RReset(PProgramming)] give insight into the power required to change the device state (Fig. 3d). The power applied to the device needs to be sufficient to melt the phase change material in a volume large enough to block the electrical current path through the device even after partial recrystallization in the course of melt-quenching. Above this programming onset the size of the amorphous mark in the device increases continuously with rising programming power; visible as increasing reset resistance.21
Programming onset (1275 ± 40 μW) and reset resistances above onset at different ambient temperatures (100 K, 150 K and 200 K) are comparable within the margin of error. Up to 200 K we detect an amorphous mark in the device. With changing ambient temperature the size of the molten volume (for the same power),31 crystal growth velocity,20,31 resistivity21 and drift coefficient32,33 may alter significantly. For this reason, we must refrain from drawing conclusions about the relative size of the amorphous mark obtained at different temperatures.
At 250 K no resistance change could be identified, which does not necessarily mean that it is no longer possible to quench a glass from the melt. Instead, the amorphous material might recrystallize within the 1.5 s it takes to switch from the AC-path (Reset) to the DC-path (Read).
Extensive programming experiments with various pulse trailing edges show improved programmability for device structures that are confined more narrowly (Fig. 4). A reduction of the antimony thickness from 10 nm to 5 nm decreases the programming onset by a factor of almost 2, from 1300 μW to 675 μW (3 ns trailing edge). Less material needs to be molten and accordingly it requires less power to program the more narrowly confined devices.1,2
Fig. 4 Programming curves of a 5 nm (a) and 10 nm (b) antimony device at 100 K ambient temperature. The ON/OFF ratio is given by RReset/Rset. The programming curves are measured as described previously (see the caption of Fig. 3). The trailing edge of the programming curves is increased in two nanosecond steps until the quench rate is too slow to create a change in linecell resistance; the pulse width is kept constant at 50 ns. Finally, to ensure the device did not degrade during the experiment a programming loop with the fastest trailing edge is repeated and reproduces the first. Error bars denote the standard deviation. |
At the same time, the maximum ON/OFF ratio increases by a factor of 4 (compared at a programming power of around 50% above the programming onset). While the resistivity of crystalline antimony increases with decreasing film thickness,26 we do not have to expect large changes in the resistivity of amorphous thin films with different thicknesses as the typical mean free path in amorphous semiconductors is of the order of interatomic distances.34–36 Hence, the increased ON/OFF ratio is attributed mainly to a larger amorphous mark in the 5 nm Sb sample.
In addition, the quench rate that is necessary to suppress complete recrystallization of the initially molten material may be slower for a more narrowly confined material. Reset pulses with a 5 ns trailing edge do not end abruptly enough to decrease the conductance of the 10 nm Sb device, whereas the 5 nm Sb device can still be programmed even with 7 ns trailing edges. Sb-rich phase change materials are fast-growth materials with already low nucleation rates in much larger volumes.24 A small device volume like ours reduces the probability for nucleation even further and the rim of the small molten volume reaches an interface with crystalline antimony. Thus, no nucleation event is necessary for recrystallization: it can proceed purely by growth, i.e. by progression of the existing crystal boundary into the previously molten volume.
These two observations in more narrowly confined devices, i.e. an increased ON/OFF ratio and amorphization with slower pulse trailing edges, can be attributed either to a slower effective cooling rate (heat remaining longer) in the thicker device or a reduced crystal growth velocity upon confinement. The following section will demonstrate that the second effect is crucial.
To estimate the maximum retention time, we define the threshold resistance (2 × RCryst) below which a reset state can no longer be reliably distinguished from the set state. A retention time merely obtained by a threshold value will, however, be sensitive to its exact value and discards all data except for a single point. By linearly fitting resistance over time the estimated retention time gets less sensitive to noise.§
The retention time shows an Arrhenius temperature dependence over at least four orders of magnitude in time (Fig. 6). Following again the hypothesis of recrystallization taking place only from the rim towards the center of the amorphous mark, without the necessity of any nucleation event, the retention time is dominated by the crystal growth velocity. Samples with 10 nm Sb have an activation energy of 1.09 ± 0.19 eV. Confining the antimony from 10 nm to 5 nm and 3 nm increases the retention times by several orders of magnitude. The thickness reduction to 5 nm shifts comparable retention times to almost 100 K higher temperatures. Even stronger confinement to 3 nm boosts the retention time by another two orders of magnitude. This difference can certainly not be explained merely by a change of the size of the amorphous mark created in devices with different thicknesses of Sb. Instead, it demonstrates that the crystal growth velocity is significantly reduced in more narrowly confined material. The amorphous phase gets stabilized against crystallization.
The retention time of a sample confined to 3 nm is on the order of 32 s at 80 °C, which is 500 times the 64 ms DRAM refresh interval.37 Memory-mapped storage class memories (SCMs) start to become advantageous with retention times that are only ten times longer than that DRAM refresh interval.38 Strongly confined phase change memories of pure antimony are able to satisfy this specification.
The attempt to pursue Moore’s Law will inevitably push device dimensions to scales where bulk properties no longer apply. In this regime materials previously not considered suitable for phase change memory applications can also become valid candidates for device applications. Instead of optimizing material compositions, interface-engineering may open new routes to enhance device performance.
Footnotes |
† The trailing edge of the voltage pulses applied to the linecell defines the electrical quench rate. Devices with 10 nm Sb could only be quenched if pulses with the fastest electrical trailing edges (3 ns) were applied from the linecell side. |
‡ The cell resistance on timescales below 2 s was measured by applying a burst of 1000 triangular voltage pulses (with a peak voltage of 0.25 V). The current–voltage curve of each triangle is fitted linearly to obtain the device resistance. Measurements on longer timescales were realized with an SMU. The delay of the first SMU-based resistance read after reset is approximately 1.5 s. |
§ For the fit only data above the threshold resistance of twice the crystalline resistance (Rthreshold = 2 × Rcryst) are considered. If the resistance drops within the first 10 resistance measurements below the maximum measured resistance the data is fitted from the first measurement point. Otherwise the resistance is fitted from the point at which it drops for the first time below 0.9 times the maximum value. This procedure aims at fitting the experimental data only in a regime in which the resistance continuously decreases. Leaving out regimes where e.g. resistivity drift dominates the devices resistance avoids overestimating the retention time. |
This journal is © The Royal Society of Chemistry 2019 |