Ultralow-power switching via defect engineering in germanium telluride phase-change memory devices

Crystal–amorphous transformation achieved via the melt-quench pathway in phase-change memory involves fundamentally inefficient energy conversion events; and this translates to large switching current densities, responsible for chemical segregation and device degradation. Alternatively, introducing defects in the crystalline phase can engineer carrier localization effects enhancing carrier–lattice coupling; and this can efficiently extract work required to introduce bond distortions necessary for amorphization from input electrical energy. Here, by pre-inducing extended defects and thus carrier localization effects in crystalline GeTe via high-energy ion irradiation, we show tremendous improvement in amorphization current densities (0.13–0.6 MA cm−2) compared with the melt-quench strategy (∼50 MA cm−2). We show scaling behaviour and good reversibility on these devices, and explore several intermediate resistance states that are accessible during both amorphization and recrystallization pathways. Existence of multiple resistance states, along with ultralow-power switching and scaling capabilities, makes this approach promising in context of low-power memory and neuromorphic computation.

Raw tempe e exposed to its linear    At high incident energy of He + ions, the particle energy decreases along the penetration through the alloy, and this is parametrized by stopping power (defined as S=dE/dx: energy lost per unit penetration depth). At low penetration depths, the stopping power is very uniform, suggesting that the very little fraction of this energy loss occurring due to nuclear interactions (knock-on damage) is also very uniform. Using SRIM, we verified that the stopping power of GeTe alloy for 2 MeV He + particles is very uniform in the thickness range of nanowire devices used in this work (from 60 to 200 nm), with only 2% of the incident energy lost in a 200 nm thick nanowire (2 MeV to 1.96 MeV). This very little energy loss also means that almost all the incident He + ions penetrate through the sample without getting implanted in the form of He bubbles. Furthermore, the knock-on damage due to ion-bombardment is homogeneous through the nanowire, without any influence from the abrupt energy loss.

Supplementary Note 2: Extrapolation of saturation resistivity (ρ 0 ) and temperature coefficient of resistivity (TCR) from resistance-temperature plots in the metallic phase
Supplementary Figure 1 shows temperature dependence of resistance of some representative devices in the metallic phase radiated by He + ions at different dosages (0, 300, 700 μC cm -2 ). In all these devices in the metallic phase resistance increases with temperature beyond 30K where phonon-carrier scattering dominates transport, and saturates below 30K where carrier-defect scattering is predominant. So, the resistance extrapolated to 30 K is the saturation resistance (R s ) from which saturation resistivity (ρ o , see Fig 1a) is calculated as ρ o = R s A/l, where A is the cross-sectional area of the nanowire, l is its length. The slope of the resistance-temperature plots when resistance increases linearly with temperature is the temperature coefficient of resistance, from which TCR is calculated by multiplying the slope with a geometric factor A/l. Note that the nanowire resistance is calculated by subtracting the contact resistance extracted from a multi-probe measurement from the device resistance.

Supplementary Note 3: Plasmonic spectroscopy, carrier counting and explanation of TCR trends with dosage in the metallic phase:
As per ref. 1, an approximate expression for TCR in metallic phase considering the weak-localization corrections from defects, is given as TCR  2 m ne 2 Increase in m with dosage is trivial to understand, however, to verify if carrier concentration (n) decreases upon ion irradiation we performed plasmonic spectroscopy using energy electron loss spectroscopy (EELS) in scanning transmission electron microscopy (STEM). The energy of the bulk plasmon resonance is directly dependent on the bound electron density in a material as 2 , and an increase in the energy of plasmonic resonance corresponds to an increase in the bound electron density or a decrease in hole-carrier concentration. In Supplementary Figure 2, we show that upon irradiation at low dosages (50 μC cm -2 ) on 11 different nanowires, the bulk plasmon peak on an average increases from 16.5 eV to 17.1 eV, indicating a decrease in the hole-carrier concentration upon irradiation by 2% (i.e. 17.1 16.5  1        100 ). In GeTe, since Ge vacancies are responsible for large hole concentration (10 19 -10 21 per cm 3 ) 3 , decrease in hole concentration structurally corresponds to a decrease in Ge vacancy concentration, which is to be expected in ion-irradiation where vacancies supersaturate to form extended defects (Fig. 2 in the manuscript).

Supplementary Note 4: Stability of the insulating crystalline state (state 1) engineered by pre-inducing defects
We measured the resistance of the as-engineered insulating crystalline state at 200 o C as a function of time to study the thermal stability of this phase. As shown in Supplementary  Figure 3, there was no change in the resistance of this phase atleast for 36 hours at 200 o C; and this suggests that thermal degradation of this phase via pre-induced defect annealing during memory operation is a non-issue. The SET and RESET cycling data for ~20000 times on a defect-engineered device shown in Fig. 4c of the manuscript, is a further testimony to the stability of defects in the defect-engineered insulating crystalline phase. The SiO 2 conformal coating, we presume plays an important role interms of removing any surface sinks for the extended defects, giving rise to this extraordinary stability. In

Supplementary Note 5: He bubbles and associated voids during He + ion irradiation
It is well known from literature that the implanted He + ions cannot stabilize as such in any material but rather transform to He bubbles and escape from the system, leaving voids behind. Primarily, the reason for using He + ion irradiation is to avoid implantation, and encourage defect-formation via knock-on damage. He bubbles can sometimes get trapped in polycrystalline materials 5 . However, not surprisingly we see no evidence of He bubble trapping from our TEM images. In Supplementary Figure 4, we show some more TEM images of our  p 0.5~N boundelectrons nanowires which clearly demonstrate the non-existence of these bubbles. A general procedure to observe the voids, which are left from He bubbles escaping, is via defocused bright field TEM images 6 . However in our single-crystalline nanowires we do not observe these voids too, and this clearly suggests that there is no He + implantation which subsequently either creates bubbles or voids (see also Supplementary Note 1).

Supplementary Note 6: Profiles of applied voltage pulses and current response
Shapes of the applied voltage pulses and their current response through the devices are measured through a 500 MHz oscilloscope (measurement procedure is described in Methods). Supplementary Figure 5 shows that for a 50 ns applied voltage pulse, the current is mostly at a steady value during the pulse, given by the ratio of applied voltage (V) and the resistance measured after removing the pulse (R). Hence current in the device can be approximated as a rectangular pulse of amplitude V/R. For a 20 ns voltage pulse, however, the reflections corresponding to slight changes in the voltage pulse dominate, and the rectangular current pulse assumption is no longer a valid one. Our discussion in the manuscript, about formation of intermediate states starting from state 1 by applying 20 ns voltage pulses, hence did not contain any quantification of the current values.

Supplementary Note 7: Nanowire switching behavior at dosages where carrier localization effects are not significant (less than 700 μC cm -2 )
As shown in Supplementary Figure 6a, at dosages up to 700 μC cm -2 where carrier localization effects are not significant, we observe the following trends in switching behavior: (i) Switching current density (j s ) increases with increasing l d up to a critical length, l c , demonstrating volume scaling of switching current up to l c , (ii) j s suddenly drops at l c and subsequently increases again with length, and (iii) l c itself decreases with increasing dosage.

Size scaling of switching currents up to l c
In the defect-templated pathway for amorphization, heat shock from a current pulse is responsible for quenching of vacancy clusters into extended defects 4,7 , which migrate with the hole-wind force, and keep piling up at a region of local inhomogeneity up until a local collapse of long-range order (amorphization). The switching currents in this pathway scale with device volumes, and this can be understood by separately considering the effect of a current pulse on devices of same cross sectional area and varying lengths; and devices of varying cross-sectional areas and same length.

Comparison of the effect of a current pulse on nanowire devices of different lengths and same cross-sectional area:
We performed finite element simulations of temperature profiles in a nanowire device upon the application of a current pulse, using COMSOL 8 , to understand the role of heat-shocks in the length scaling behavior of switching currents for devices switched via the defecttemplated pathway. The geometry of the GeTe nanowires (thermal conductivity, κ=0.5 W per mK; thermal diffusivity, α=5x10 -3 cm 2 sec -1 ; and electrical resistivity ρ=0.4 mΩ cm) was simulated as long bars with square cross-section-the diameter of the nanowire being the width and height of the bar. The ends of the nanowire devices were considered to be the heat sinks (electrode regions), and the entire device was embedded in an SiO x dielectric of thickness 30 nm (κ =1.6 W per mK, α=0.1 cm 2 sec -1 ). Current amplitudes were set to the desired values and pulses were defined as rectangular functions of current with rising and trailing edges of width 2.5 ns each. Spatial profile, 120 ns after the application of 100 ns, 0.4 mA pulse (120 ns is the time instant when maximum temperature is reached) on a 2500 nm x 100 nm x 100 nm nanowire device is shown in Supplementary Figure 6b.
To compare the effect of a current pulse (0.4 mA, 100 ns) on devices of different lengths and same cross-sectional area, we calculated the temporal profile of the temperature at the midpoint of several devices (which is spatially the maximum temperature region at every time instant) with varying lengths (900-2700 nm). We find that with increasing the length of the device, quench times for the heat shocks corresponding to a particular current pulse become longer (Supplementary Figure 6c), making their severity lesser and hence the defect creation and migration process less effective. Hence, under a valid assumption that the defect density required for amorphization is size independent, longer devices require higher currents to achieve this critical density than shorter devices, and this explains the length dependence of switching currents.

Comparison of the effect of a current pulse on nanowire devices of different crosssectional areas and the same length:
It is trivial to understand that any particular current pulse will heat a thinner device to a higher temperature than a thicker device of the same length, as the current density is lesser in a thicker device. Defect creation (and subsequent migration) from a current pulse in a thinner device is hence more effective than that in a thicker device, manifesting as higher switching currents for thicker devices, thus explaining scaling of switching currents with cross-sectional area.
Combining both the length scaling as well as cross-sectional area scaling of switching currents, we can conclude that the switching currents scale with the volume of the device in defect-templated amorphization pathway also (just as in melt-quench pathway).

Sudden drop of switching current density at l c
The pre-induced defects and the defects created by the heat shock during electrical pulsing, migrate with the electric wind force and accumulate at a region of local inhomogeneity-defined by structural, morphological or thermal factors-which impedes the motion of defects 4,7 . Beyond a critical concentration of defect pile-up in this local region, an amorphous phase nucleates 7 , and it is easy to argue statistically that longer devices have more of such inhomogeneities than the shorter ones. Hence, longer devices may be treated as many short segments, each containing a defect-templating location. Defect build up towards amorphization at these templates in all the segments happens simultaneously, with the shortest segment determining the ease of switching. Devices just longer than l c have one extra defect-templating location and hence an extra shorter segment than the devices just shorter than l c . This explains the sudden drop of switching current as a function of device length at l c at all the dosages until 700 μC cm -2 . Furthermore, since pre-induced defects act as natural inhomogeneities, it is easy to create multiple jamming locations in devices irradiated at higher dosages (greater concentration of pre-induced defects); and this is reflected as decrease in l c with increasing dosage (Supplementary Figure 6a).

Switching currents increase with dosage at lower dosages (less than, 700 μC cm -2 )
From a structural point of view, in defect-templated amorphization pathway if extended defects are pre-induced, energy expense for both creation and migration of the defects, can be massively reduced, and this translates to reduction in switching currents (and current densities). Contrarily, we note from the data in Supplementary Figure 6a that the switching currents increased with increasing dosage up to 700 μC cm -2 (where defects do not induce any carrier localization effects in transport). This behavior can be understood by noting the trends in programming curve (device steady state resistance as a function of voltage pulse amplitude) of the devices that were ion-irradiated, and those that did not (Supplementary Figure 6d). The ionirradiated devices at low dosages (less than 700 μC cm -2 ) show a clear dip in the resistance as a function of voltage pulse amplitude -not prominent in non-irradiated devices. The dip in resistance from the initial value is a result of pre-induced defect annealing or reorganization in a manner that increases the carrier mobility by reducing carrier-defect scattering. This defect reorganization requires more work to be done on the system (than the case with no pre-induced defects), which manifests itself as higher switching currents. Hence, low-concentration of preinduced defects which electronically do not modify the material is bad in terms of energy consumption for the crystal-amorphous switching process.

Determination of active volumes and active areas:
From previous works on defect-templated amorphization in both GeTe 4 and Ge 2 Sb 2 Te 5 7 we know that amorphous region forms in a local region cutting across the entire cross-section of the nanowire, thus making the active area as the total cross sectional area of the device. The defect evolution responsible for amorphization happens throughout the nanowire device (between two electrodes), and thus the active volume is the total device volume.

Threshold switching via d.c. I-V sweep:
In Supplementary Figure 7, we show threshold switching data to SET state of device D2 (320 nm x 80 nm x 80 nm) exposed to 3600 μC cm -2 , which amorphized at a very low RESET current of 8 μA (0.13 MA cm -2 ; see Fig. 4 of the manuscript).
Supplementary Note 9: Intermediate states, visualization and stability:

Visualization of intermediate states:
The intermediate states are metastable crystalline states, which microstructurally can be described as having a background concentration of defects in most of the nanowire, and a local region where the defect concentration exceeds the background. Supplementary Figure 8 shows TEM images of two of our devices, which were programmed to one of the intermediate states.
However, it must be noted that various intermediate states structurally differ by concentration difference in the defects in the defect-template region. Small changes in defect density in the defect-template are difficult to characterize using diffraction contrast TEM, and hence we rely on transport measurements to deduce both structural and electronic information about these states.

Thermal stability of intermediate states and the amorphous phase
We examined the thermal stability of the amorphous phase and an intermediate state -a limiting factor that determines the data non-volatility, by performing high temperature retention measurements, as practical memory applications warrant high-temperature performance 9 . Isothermal crystallization from an amorphous phase or an intermediate state at high temperatures show an initial incubation regime (no change in resistance) corresponding to the time required for the formation of a critical nucleus of the crystal, followed by a growth regime 9 (Supplementary Figure 9a, b). We considered data retention times as the incubation times, and in both these states (amorphous and intermediate) the incubation time-temperature plots show an Arrhenius behavior (t=Pexp(E a /kT), Supplementary Figure 9c) 9,11 , with the amorphous phase displaying excellent thermal stability, extrapolated to 3.1 years for device operations at 115 o C. The intermediate state is however, not very stable for high temperature operations i.e. for operations at 40 o C, extrapolated stability is 3.1 years, whereas for operations at 70 o C it is barely 20 minutes; and improving the thermal stability of these intermediate electronic states is an interesting problem for future work.

Obtaining intermediate states starting from defect-engineered insulating crystalline state
In Supplementary Figure 10 we show additional data to Fig. 4a on another device (D3, which also amorphizes abruptly at 23 µA with the application of 50 ns pulses, but which proceeds through intermediate states by controlled addition of defects upon the application of 20 ns pulses.

Supplementary Note 10: Model for recrystallization
After amorphization, the nanowires have the following microstructure: most of the nanowire has a background density of pre-induced defects, and a local region(s) has higher defect concentration (defect-template). From previous works on defect-templated amorphization 4,7 we know that the amorphous region cuts across the cross-section of the nanowire in this template wherever the defect concentration exceeded a critical value required for amorphization (Supplementary Figure 11a, b). So in this local region of the defect-template, there is an amorphous region surrounded by a heavily defective crystalline region with the defect concentration exceeding the background concentration (Supplementary Figure 11). The reverse process, the clues for which are obtained from this work for the first time, involves two steps: a) threshold switching followed by recrystallization of the amorphous region, just as it happens in the melt-quench strategy and b) reduction of defect concentration in the rest of the template through homogenization of defects via Joule heating (Supplementary Figure 11c shows recrystallized device, with almost uniform defect contrast suggesting defecthomogenization). The latter process, i.e. degree of defect homogenization, provides access to