John D.
Langhout
a,
Danielle N.
Alverson
a,
Colton
Ginter
a,
Bruce
Ravel
b,
David P.
Adams
c and
Megan M.
Butala
*a
aDepartment of Materials Science and Engineering, University of Florida, Gainesville, FL 32611, USA. E-mail: mbutala@ufl.edu
bMaterial Measurement Laboratory, National Institute of Standards and Technology, Gaithersburg, MD 20899, USA
cSandia National Laboratories, Albuquerque, New Mexico 87185, USA
First published on 8th May 2024
Ge2Sb2Te5 is used in phase change memory, a nonvolatile memory technology, due to its phase change properties. The primary advantage of phase change memory over the state-of-the-art (flash memory) is its simple and small device geometry, which allows for denser nodes and lower power consumption. In phase change memory, resistive heating induces fast switching between the high resistance amorphous and low resistance crystalline phases, corresponding to storage of low and high digital states, respectively. However, the instability of the amorphous phase of Ge2Sb2Te5 presents issues with processing and long-term data storage; such issues can be resolved by C doping, which stabilizes the amorphous phase and raises the crystallization temperature. To better understand the local structural effects of C doping on Ge2Sb2Te5, in situ Ge K-edge X-ray absorption spectroscopy measurements were taken during heating of films with various C doping concentrations. The range of structural transformation temperatures derived from X-ray absorption near-edge structure analysis across the C doping series proved narrower than crystallization temperatures reported in similar in situ X-ray diffraction experiments, which may reflect changes in local structure that precede long-range ordering during crystallization. In addition, rigorous extended X-ray absorption fine structure fitting across and between temperature series revealed effects of C doping on the rigidity of Ge–Te bonds at low (2 at% and 4 at%) C concentrations.
The leading candidate material for optimization of phase change memory is Ge2Sb2Te5 (GST), a chalcogenide with fast crystallization and amorphization rates, as well as high resistivity contrast between its crystalline and amorphous states.1,3–8,14–16 A shortcoming of GST for use in phase change memory is the poor phase stability of its amorphous state. Specifically, the relatively low crystallization temperature of GST (≈140 °C) can result in unwanted crystallization during processing, high-temperature use, or long-term use.6,10,13,15,17 Accordingly, increasing the crystallization temperature of GST is essential for its optimization for phase change memory.
One effective approach for increasing the crystallization temperature of chalcogenide phase change materials is through the incorporation of dopants, typically nitrogen,10,18–21 oxygen,22 and carbon.2,13–17,23–25 However, how these dopants modify structure and properties is not well understood. For GST, carbon is a well-studied dopant, but there is not yet a consensus of how carbon is incorporated in the amorphous or crystalline structures, nor of the mechanism by which carbon modifies phase change behavior, especially crystallization temperature.15,16 Improved understanding of the interdependencies between structure, composition, and properties would not only allow for the optimization of GST for phase change memory applications, but would also have broader implications for designing structure and function in other chalcogenide materials.
It is challenging to study the structure of amorphous materials or characterize light elements using X-rays independently; both are required in combination to understand the role of carbon in GST. X-Ray absorption spectroscopy (XAS) provides access to local structure in the absence of long-range order, however, weakly scattering light elements (i.e., C, N, O) in low concentrations are challenging to probe when their scattering paths overlap with those containing heavier elements such as Ge, Sb, and Te. This issue presents a unique challenge for identifying how carbon modifies the behavior of GST, and more broadly, how light dopants are incorporated in amorphous materials. Indeed, GST is but one material among an abundance of amorphous chalcogenides with promise for memory applications, such as phase change memory1,2,5–7,10,13–16,18,19,23 and ovonic threshold switching,26–30 for which further structural understanding would benefit their development.
Here, we probe local structure changes during crystallization for GST and carbon-doped GST (GST-C) thin films at several compositions [(Ge2Sb2Te5)1−xCx, where x = 0.02, 0.04, 0.06, and 0.12] using in situ Ge K-edge X-ray absorption spectroscopy. We see abrupt changes in the local environment of Ge upon crystallization, and a narrow range of structural transformation temperatures over the composition series relative to X-ray diffraction studies.14 We posit that this difference reflects the different sensitivities of diffraction and spectroscopy to long-range and local atomic structure. From our analysis of in situ extended X-ray absorption fine structure (EXAFS) data, we also find the local environment of Ge changes before crystallization, and that the nature of these changes varies with carbon content. More specifically, we attribute this behavior to a modulation of the Ge–Te bond rigidity with the addition of carbon.
In addition to identifying structural insights specific to GST/GST-C, we develop a rigorous methodology with which to understand its temperature-dependent behavior while avoiding overfitting and other nonphysical fitting results. Of particular note, our methodology recognizes the physical limitations of EXAFS and accounts for parameter correlations by applying parameter fitting constraints and analysis of parameter trends, rather than focusing on parameter values directly. We anticipate that the materials insights and methodological practices we report here will benefit fundamental understanding and technological development for GST and related materials.
XANES data for GST, GST-2%, GST-4%, and GST-6% were fit as linear combinations of their amorphous (lowest temperature) and crystalline (lowest temperature post-transformation) spectra (GST and GST-6% in Fig. 1, GST-2% and GST-4% in Fig. S1, ESI†). Fitting results reflect the step-wise transformation of the spectra, emphasizing the abrupt transformation of GST-2%, and GST-4% between 140 °C and 150 °C, and the relatively gradual transformation of GST spectra between 120 °C and 150 °C and GST-6% between 140 °C and 160 °C.
The change in transformation temperature from in situ XANES data does not align with crystallization temperature (TC) as derived from in situ XRD and resistivity measurements at similar C-doping concentrations. Previous work on similarly prepared films indicates an increase in TC from 125 °C (GST) to 145 °C (GST-6%) based on in situ XRD analysis.14 Literature consistently reports a TC of ≈140 °C for GST,15,17,31 based on observation of resistivity changes and/or the emergence of diffraction peaks during in situ heating. TC measured for carbon-doped GST is typically between 180 °C16,24,31 and 220 °C,15 but values as high as 300 °C have been reported,25 depending on carbon content, film preparation,20 and film thickness.31
The disparity between the transformation temperatures detected using XAS and TC as measured using XRD is most likely a result of the sensitivity to local and long-range structure of XAS and XRD, respectively. Using XRD, the crystallization temperature is defined as the temperature at which diffraction peaks emerge. Meanwhile, XAS is sensitive to the proportion of Ge atoms in local environments resembling the amorphous and crystalline phases. Macroscropic crystallinity is therefore not necessary for a material to ‘look’ crystalline in XAS data, for example, in the case of small crystallization seeds which precede the formation of crystallites.32 In such a case, the material is beginning to crystallize, but is not yet crystalline. Thus, for the sake of accuracy, we will define the first temperature at which the XANES spectra begin to take on crystalline features as the ‘crystallization onset temperature.’
![]() | (1) |
Before crystallization, Fourier transformed EXAFS spectra for all GST/GST-C compositions have one intense feature centered at 2.5 Å, which corresponds to the first atomic coordination shell, and few discernible features beyond 3 Å (Fig. 2). These spectra are consistent with an amorphous material, in which single scattering paths of nearest-neighbor atoms dominate |χ(R)|. At and above the crystallization onset temperature, spectra take a considerably different shape; the intensity of the first coordination feature decreases and the position shifts to a longer distance. Additionally, scattering paths beyond the first coordination shell emerge, indicating a higher degree of crystallinity (Fig. S5, ESI†).
In addition to the pronounced changes in the Fourier transformed EXAFS spectra upon crystallization, there are also relatively subtle temperature-dependent changes in the spectra prior to crystallization. Specifically, in |χ(R)|, the amplitude of the first-shell scattering path gradually dampens with increasing temperature for all compositions, indicating a change in the Ge local environment within the amorphous state. A decrease in the magnitude of |χ(R)| is expected with increasing temperature as a result of thermal vibrations, which broaden features and decrease their amplitude.29,33,35–37 Notably, the extent of this dampening is composition-dependent; by measuring the change in the maximum |χ(R)| of the feature centered at 2.5 Å, we find a decrease in magnitude of 27% for GST and decreases of 10%, 13%, and 15% for GST-2%, GST-4%, and GST-6%, respectively (Fig. 2).
To better understand and quantify the temperature- and composition-dependent structural changes in amorphous GST and GST-C before crystallization, we fit EXAFS spectra using a least-squares method. In doing so, we considered several structural models, favoring those which balanced a low R-factor (the fitting figure of merit) with a reasonably low number of fitting parameters. The most successful models consisted of two single-scattering paths: a first-shell single scattering Ge–Te path and a first-shell single scattering Ge–Ge path.
The contribution of the first-shell single scattering Ge–C path was qualitatively visible near 1.4 Å in the amorphous EXAFS spectra of doped samples (Fig. S4, ESI†). However, modeling this small feature is difficult since it is associated with a dilute and weakly scattering species. Further, modeling such short paths is inadvisable, since data in this region are strongly affected by background removal and can therefore be heavily distorted by the data reduction process.10 For these reasons, we did not include a Ge–C path in fitting models.
EXAFS models consist of several fitting parameters, including the parameter ΔE0, a correction term to the defined edge energy E0, and the amplitude reduction factor S02, which bears little meaning in our discussion as it is empirically derived and not directly related to atomic structure.10,33,34 The value we used for S02 (0.75) was determined from the best-fit model to GST and set constant when fitting other spectra. Three path-specific parameters are also added from each path in a model (e.g., the Ge–Te or Ge–Ge single scattering paths): N, the path degeneracy; r, the half path length; and σ2, the mean square relative displacement parameter (akin to the Debye–Waller factor in diffraction analysis).33,38 The models we used to fit amorphous EXAFS data comprise only single-scattering first-shell paths, such that the degeneracy, N, of each path is essentially the coordination of, in this case, Ge by either Ge or Te. Likewise, the half path length, r, for these first coordination single scattering paths is equivalent to the bond length for Ge–Te or Ge–Ge bonds, and σ2 to the bond length variance.
To determine the origins of differences in EXAFS data as a function of composition and temperature for amorphous GST, we applied several approaches to fitting, which are described following. We first allowed path parameters to vary in unconstrained fitting, which we found to be unstable when applied over the temperature series. Instead, we constrained path parameters to minimize large swings of correlated parameters’ values, from which we evaluated trends within and between each composition.
![]() | ||
Fig. 3 Low temperature (30 °C) |χ(R)| spectra for amorphous (a) GST and (b) GST-2% fit between 1.5 Å and 3.5 Å to models using single-scattering paths for Ge–Ge and Ge–Te with all path parameters unconstrained. Visualization of per-path contributions are provided in Fig. S6 (ESI†). |
To assess the extent of the changes in the local Ge environment with temperature, we applied the model fit to room temperature data to the rest of the amorphous data, tracking the relative goodness of fit, represented by the ‘R-factor.’ A smaller R-factor indicates a better fit, with an R-factor <0.02 generally corresponding to an acceptable fit of the model to the data.33
We used path parameters from unconstrained fits to each lowest temperature spectrum (30 °C for GST, GST-2%, and GST-6%; 40 °C for GST-4%) to generate representative models for each composition. These models were then ‘fit’ to spectra over the rest of the temperature series with all parameters fixed, except for ΔE0. In this ‘fixed fitting’ approach, in which no parameters are refined, the increase of the R-factor with increasing temperature reflects the deviation of spectra from the room temperature model (Fig. 4).
![]() | ||
Fig. 4 R-factor as a function of temperature for amorphous GST, GST-2%, GST-4%, and GST-6% using the ‘fixed fitting’ approach. The increase in R-factor reflects the change in the spectra with increasing temperature; this effect is greatest for GST and GST-6%. Some points are omitted due to inconsistencies in data quality. These omissions are discussed in greater detail in Section S1.2 of the ESI.† |
When the resulting R-factors from these ‘fits’ are plotted as a function of temperature for each composition, temperature- and composition-dependencies become evident (Fig. 4). For room-temperature data, the fits have R-factors <0.02, indicating excellent agreement between the data and the model (Fig. 3). With increasing temperature, the R-factor increases, reflecting a departure of the spectra from their low-temperature models, suggesting that the Ge environment changes significantly with temperature. Furthermore, this behavior is more severe in GST and GST-6%, which show steeper increases in R-factor, than in GST-2% and GST-4%, indicating a composition-dependence of the extent of the local change that precedes crystallization onset.
Visually, the increase in R-factor in Fig. 4 can be attributed to the amplitude of |χ(R)| decreasing with increasing temperature, as seen in Fig. 2. This behavior is commonly observed in amorphous and crystalline materials alike; increasing dynamic atomic disorder due to temperature increases the variance of atomic correlation lengths, dampening the χ(R) signal. More specifically, this behavior is usually best captured using an Einstein model,35–40 which assumes that the change in σ2 is primarily due to the activation of optical phonon modes,35 resulting in a nearly linear increase in σ2 at temperatures above room temperature and across modest temperature ranges.
Testing whether GST and GST-C follow this well-studied behavior proved uniquely challenging. Since GST is amorphous, NGe and NTe cannot be constrained to known coordination values from a crystal structure, so they must instead be refined as variables. Since both N and σ2 contribute significantly to the magnitude of χ(R) of each path, the two parameters are highly correlated, leading to uncertainty in both terms.33 Moreover, since the Ge–Te path and the Ge–Ge path are nearly the same length, their interference with each other causes the fit to be sensitive to their respective ratios.10,35 Finally, N generally contains a high amount of uncertainty due to its correlation to S02, the amplitude reduction factor. Without multiple shells to fit, S02 was fixed to a constant value, and its uncertainty is transferred to N, with which it is statistically indistinguishable, following eqn (1).34 These three factors cause N and σ2 to be highly correlated within and between paths. Fitting the temperature series with all four correlated parameters free to refine causes a high level of uncertainty, and sporadic, nonphysical parameter trends, highlighting the need for an alternate fitting approach.
To account for the change in EXAFS amplitude over each temperature series while avoiding strong parameter correlations and statistical uncertainty, we applied a constrained fitting approach. We fixed σ2 for both paths over each temperature series, while the remaining parameters (ΔE0, N, and r) were refined with no effective constraints. For each temperature series, the value of σ2 for both paths was fixed to the value of the lowest temperature fit. By constraining σ2, intensity changes of |χ(R)| were captured by changes in the values of N for each path.
This constrained fitting approach resulted in more stable, continuous trends in fitting parameters (NGe and NTe) over the temperature series as a result of the reduced parameter correlations (Fig. 5). These fits also do not have the temperature-dependent increase in R-factor that resulted from applying a fixed model over the temperature range, indicating that the resulting models accounted for the temperature-dependent behavior of the EXAFS spectra. Further, the R-factor for fits for all compositions were reduced to near or below 0.02, indicating excellent agreement between the models and the data at all temperatures (Fig. 6).
![]() | ||
Fig. 5 EXAFS fitting results for the constrained fitting method. (a) NGe–Te·σGe–Te2 as a function of temperature for each composition show a linear decrease for all compositions, and steeper slopes for GST and GST-6%. Error bars denote standard error. (b) Resulting R-factors for constrained fits indicate excellent fits relative to fixed fitting (Fig. 4). Some fit values are omitted due to inconsistencies of data quality (see Section S1.2 of ESI†). |
![]() | ||
Fig. 6 High temperature |χ(R)| spectra for amorphous (a) GST (120 °C) and (b) GST-2% (140 °C) fit between 1.5 Å and 3.5 Å using the ‘fixed fitting’ and constrained fitting approaches. |
Plots of the product NGe–Te·σGe–Te2 against temperature prior to crystallization show linear, negative relationships for each composition (Fig. 5a). Note that we report the constrained fitting data as a single value (the product of N and σ2) since these highly correlated parameters must be considered in context of each other, and the precise values of N and σ2 are ultimately arbitrary. Since N is negatively correlated with σ2, this linear behavior can be explained by an increase in σ2 and no change in N. Indeed, this behavior would coincide well with the aforementioned Einstein model for the relationship between σ2 and temperature.36–38,40 The alternative to this would be an increase in N and little-to-no change in σ2, however, this is unlikely given the established thermal dependence of σ2. We observed similar behavior for NGe–Ge·σGe–Ge2 (Fig. S8, ESI†), however, due to the higher signal of the Ge–Te path in |χ(R)|, the parameters for the Ge–Ge path have a higher level of statistical uncertainty.
Linear regressions to NGe–Te·σGe–Te2 as a function of temperature highlight composition-dependent slopes (Fig. 5). The slopes for GST and GST-6% are steeper than those for GST-2% and GST-4%, reflecting the more pronounced temperature-dependent dampening of |χ(R)| in the GST and GST-6% spectra (Fig. 2). Values for the slopes and R2 values for the linear regressions are reported in Table S2 (ESI†).
Since N and σ2 are conflated, the observed difference in slopes can result from either a higher rate of increase in σ2 or constant increase in σ2 paired with a linear decrease in N. A higher rate of increase in σ2 suggests that σGe–Te2 in GST and GST-6% is more susceptible to temperature, which would result from the Ge–Te bonds being less rigid (i.e., weaker) in GST and GST-6% than in GST-2% and GST-4%.29,35,41 This could explain why carbon-doping increases the stability of the amorphous state at low concentrations, specifically: the presence of C strengthens Ge–Te bonds and stabilizes the structure. We speculate that this interpretation of the observed parameter trends is more likely than the alternative, i.e., that Ge coordination decreases with increasing temperature. Simply, a linear increase in σ2 is a commonly observed and well-modeled phenomenon,35–38,40 while a gradual change in atomic coordination, unaccompanied by any indication of a phase change, is not.
Overall, this behavior indicates an effect of C on the strength of the Ge–Te bond, the coordination of Ge by Te, or both when C is in small concentrations (GST-2% and GST-4%). This effect is suppressed for a larger concentration of C (GST-6%), perhaps indicating that a secondary mode of C incorporation is active at higher C concentrations, which stabilizes the amorphous phase through a different mechanism. Literature has suggested the existence of multiple distinct modes of C incorporation into GST, consisting of the formation of Ge–C bonds at low levels,16,23 then Sb–C bonds16 and/or the formation of CC chains15,23,24 at higher doping concentrations.
In the context of the literature, our findings imply a more complex transition from short- to long-range order upon crystallization for GST and GST-C. A decoupling of the short-, medium-, and long-range order of phase change materials has been explored in computational studies17,42,43 and may be key to reconciling the many conflicting reports regarding the structure and performance of GST and GST-C. Certainly, further work consisting of a ensemble of complementary characterization techniques such as in situ XAS, pair distribution function (PDF) analysis, XRD, Raman spectroscopy, and transmission electron microscopy (TEM) could fully explore the structure of GST across length scales, and each structural feature's effect on bulk properties.
For example, in addition to element-specific local structure information provided by XAS, Raman spectroscopy can provide sensitivity to carbon (particularly C–C bonds and their geometry)15,46 and the geometry of Ge–Te and Sb–Te polyhedral units.46 At the mid-range, PDF analysis can provide useful structural information, such as the formation of crystalline seeds, however we note the difficulties of this method on thin film samples.47,48 As for long-range structure, in addition to diffraction studies, TEM can reveal spatially resolved information, such as the size, morphology, and distribution of crystalline and amorphous domains.15,20,23,24,45,49
We find that C content affects the nature of local structure evolution preceding crystallization. Unexpectedly, the effects of C on the evolution of the local structure with temperature do not follow a simple trend. This is opposed to the simple trend between C concentration and phase stability; previous work reports a steady, monotonic increase in crystallization temperature (and thus, amorphous phase stability) as a function of C content up to ≈6 at% C, based on XRD and electronic properties.14 However, as evident in Fig. 5, the Ge environment changes most dramatically with temperature for GST and GST-6%, while very little change is observed for GST-2% and GST-4%. We pose some explanations for the observed spectra changes, especially that C affects either the strength or the coordination of the Ge–Te bond, although further work is needed to more precisely determine how it does so. Further, the nonlinear trend between this behavior and C content may indicate two modes of C incorporation in GST that occur at low and high C concentrations. Indeed, other work has indicated the existence of two distinct models of C incorporation, e.g., distributed at low concentrations and clustered at higher concentrations.15,16,23
Finally, we report an approach to investigate the effects of light dopants on local structure of amorphous materials. While directly probing structural information (e.g., bond coordination) of light dopants remains elusive, we show how one may investigate the indirect effect of the dopant on heavier and/or more concentrated species by modulating the doping concentration and observing structural trends. Furthermore, when using XAS over a temperature series, correlated parameters, such as N and σ2, can be constrained and how they change as a function of temperature can be modeled with greater certainty. We hope this methodology can inspire similar studies of doped amorphous chalcogenides for memory applications and beyond.
Films were sputtered using stoichiometric targets meaning all species deposited as a film were originally present in a single sputter target. New targets were generally sputter conditioned for at least 1 hour prior to first film depositions. In addition, each pre-conditioned target was sputtered for ≈5 minutes to remove adventitious carbon or oxide prior to the start of each deposition. Film compositions were determined by wavelength dispersive spectroscopy to be Ge2Sb2Te5 or (Ge2Sb2Te5)1−xCx (x = 0, 0.02, 0.04, 0.06, and 0.12) within ±0.5 at% (x within ±0.005).14 The electron microprobe used for characterization is a JEOL JXA-8530F field emission microanalyzer. A Charles Taylor multi-element standard 202 was employed for elemental quantification. Reported compositions are generally 50-point averages.
All reported fitting models each contain two paths: the Ge–Te first shell single scattering path and the Ge–Ge first shell single scattering path. The scattering amplitude and phase of the Ge–Ge path was calculated using a cluster derived from diamond-structured pure Ge, whose bond length (2.45 Å) and coordination (four) forms a local electronic environment similar to that of GST. Ge–Te paths were generated the same way, except the Ge atoms surrounding the central Ge were replaced with Te atoms. Both paths were calculated using FEFF53 as a part of the XAS VIEWER interface employing LARCH.51
The Ge–Sb first shell single scattering path was not included in fitting models. While the atomic number of Sb (Z = 51) is similar enough to that of Te (Z = 52) that the scattering contributions of the species are impossible to distinguish,3,15 previous studies have not detected the existence of Ge–Sb bonds using Sb K-edge XAS.3,4,54
The signal contribution of the Ge–C first shell single scattering path is small but qualitatively visible near 1.4 Å in doped samples (Fig. S4, ESI†). Some individual fitting experiments were able to confirm the presence of Ge–C coordinations around 1.85 Å. However, this path could not be reliably fit, so the Ge–C path was omitted from reported models. Instead, the fitting window minimum was increased to 1.5 Å, omitting the signal contributions of the Ge–C path.
For ‘unconstrained’ fits, the path parameters were free to refine within broad, but physically meaningful, bounds. N was constrained to always be positive, and never exceed a sum of 6 for all paths. σ2 was constrained within the typical bounds of this parameter, between 0.002 Å2 and 0.02 Å2.33r was constrained to never deviate by more than 0.1 Å from its original value in the starting model. Incorporating parameters for the third cumulant did not meaningfully improve any fits and was thus avoided to reduce parameter correlations.33,34
Footnote |
† Electronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d4tc01082e |
This journal is © The Royal Society of Chemistry 2024 |