Thermal characterization of Ge-rich GST thin films for phase change memories by Raman thermometry

Phase change memories (PCMs) are enabled by the unique behavior of GeSbTe (GST)-based chalcogenide materials, which can reversibly switch between amorphous and crystalline phases with high resistivity contrast.^1–3 These two resistive states enable data storage in localized phase-changing volumes of the layered material. Among the emerging non-volatile memories (NVMs), PCM stands out as one of the most mature candidates owing to its excellent cycling endurance, good scalability, and low power consumption.^4,5 This is evidenced by its recent commercial realization for both, stand-alone and embedded memory applications.^6,7

Materials engineering has played a crucial role in the realization of PCM as a viable technology. The flag-ship chalcogenide alloy Ge₂Sb₂Te₅ (GST-225) inherited from optical data storage paved the way for early development of PCM, as it showed nanosecond switching times between the two material phases.⁸ However for potential NVM applications, GST-225 had several drawbacks such as low crystallization temperature (150 °C) that limited data retention, poor material reliability, and slow crystallization speed.^9,10 For NVMs, especially embedded memory applications, higher crystallization temperatures are essential to ensure high-temperature data retention.¹¹ From an integration point of view, PCMs embedded in the back-end-of-line (BEOL) should resist the thermal budget of around 400 °C.¹² Also, it is crucial to guarantee data retention after solder reflow of 2 min at 260 °C to preserve the code. These requirements impose a stringent guideline for the development of phase change materials.

In this context, extensive materials engineering has been carried out in the last decade. Specifically, two routes were followed: compositional tuning of GST away from the pseudo-binary line of GeTe–Sb₂Te₃ toward Ge enrichment and further alloying of these stoichiometries with nitrogen or carbon.^13–15 An increase in the number of Ge–Ge bonds resulted in an increase in the crystallization temperature. Furthermore, the formation of Ge–N bonds retarded crystallization kinetics, which partially resolved the persistent issue of Ge segregation and controlled the grain size.^16,17 This sufficed the material requirements for high-temperature tolerance and control over grain size resulted in improved material reliability, which was lacking for GST-225. Stoichiometries such as Ge-rich N-doped GST-212 and GST-225 Ge45%–N4% doping displayed improved data retention, with the capability of 10 years at 210 °C for the later stoichiometry.^12,14 Ge-rich GST-doped alloys have become the material of choice for high-temperature embedded memory applications.

In PCMs, the reversible switching between phases is thermally initiated. Studies report that <1% of energy is utilized for phase change, whereas most of the energy is lost via other heat dissipation pathways in the PCM cell.¹⁸ Here, thermal conductivity is a key physical parameter to understand the temperature distribution in the device and optimize the operation of a PCM device and crosstalk between cells. Therefore, knowledge of the thermal properties of phases over the entire operation temperature range of PCM cells is crucial. The flag-ship GST-225 alloy has been extensively characterized by various techniques,^19–28 but the current state-of-art falls short of thermal characterization of the newly engineered stoichiometries. From a characterization point of view, it is also crucial to have simultaneous thermal and physiochemical characterization of phase change materials.²⁹ Raman thermometry as a contactless characterization technique requiring no microfabrication provides this unique dual characteristic.

In this article, we implement Raman thermometry for the thermal characterization of Ge-rich GST and Ge-rich N-doped alloys for the first time. After a detailed explanation of the method, we demonstrate the successful extraction of temperature and phase-dependent thermal properties of these materials up to higher temperatures (∼350 °C) by Raman thermometry for the first time. The study is supplemented by x-ray diffraction analysis, which agrees with the significant structural changes observed in Raman spectroscopy analysis, confirming the additional benefit of this technique.

Ge-rich GST (GGST) and Ge-rich GST N-doped (GGSTN) 200 nm thin films were deposited by physical vapor deposition on 300 mm silicon (100) wafers using an industrial tool. The Ge content of both these materials is greater than 45%. The thin films were capped right after deposition by a 45 nm silicon nitride thin film or a 10 nm titanium nitride to avoid oxidation. These capping materials were chosen as those are the preferred interfacing materials to GST in a phase change memory cell. Raman thermometry permits the choice of such capping materials.

The silicon nitride thickness was chosen to be 45 nm to optimize power absorption at a wavelength of 473.11 nm. This was evaluated by calculating the coefficients of reflection and absorption of the laser using the transfer matrix formalism. The calculation was performed iteratively from 0 to 150 nm of SiN thickness as presented in Fig. 1. The refractive index used for SiN and GST were extracted from the literature.^30–32  Figure 1 shows that the maximum laser power absorbed in amorphous GST is for SiN thickness ranging from 40 to 50 nm. Maximum absorption of laser light for different stoichiometries of chalcogenides in its amorphous and crystalline phase lies in the same range as shown in Fig. 2. This calculation was performed based on refractive index values from the literature for a general understanding of absorbance of laser light in GST to evaluate the SiN thickness for various GST-based stoichiometries. So, deposition of SiN thickness of 45 nm was chosen to maximize laser absorption. Furthermore, the absorbance of laser in GGST and GGSTN was confirmed by reflectometry as listed in Table S1 in the supplementary material. This configuration of thickness assures maximum laser power absorbed in GST for optimum Raman signal.

The samples were crystallized by annealing ex situ at 400 °C for 30 min at a heating rate of 0.5 °C/s under N₂ atmosphere. For thermal characterization, these samples were heated in situ at the respective temperatures under vacuum. At each temperature step, 5 min of thermalization of samples is performed, and each acquisition is performed after ensuring stable temperature and pressure conditions.

Raman spectroscopy and thermometry measurements were performed using a LabRAM HR confocal system from Horiba Jobin-Yvon as illustrated in Fig. 3. The system is equipped with a continuous wave diode-pumped laser (473.11 nm) and 50× and 100× Mitutoyo Apochromatic NIR objectives. The features of the confocal system to ensure this high spectral resolution to resolve peaks are a spectrometer with a focal length of 800 mm, diffraction grating with 1800 grooves/mm, and spectral resolution at 473.11 nm < 0.7 cm⁻¹. This setup is modified with a continuously variable reflective neutral density filter—THORLABS NDL-25C-4 to enable control of the excitation laser intensity. For sample temperature control, a LINKAM heating stage—LINKAM HFS350-PB4 (−195 to 350 °C, vacuum compatible, liquid nitrogen cooling) with an optical window is used.

X-ray powder diffraction (XRD) was performed on a 9 kW rotated anode Rigaku Smartlab diffractometer working in a Bragg–Brentano geometry using Cu Kα radiation (λ = 1.5418 Å), equipped with a 2D Hypix Detector detector. To avoid saturation of the detector by the (004) peak of the single-crystal silicon substrate, after having verified that possible texture effects were negligible, a 2° misorientation was applied to the sample. Patterns were collected both at room temperature, on ex situ pristine and annealed samples, and in situ during heating. Taking advantage of the high flux of the x-ray beam delivered by the rotated anode and of the high sensitivity, of the 2D detector, patterns were acquired on the fly every 5 °C (heating from RT to 600 °C at 2 °C/min). The EVA 6.1, 2023, Bruker AXS GmbH, Karlsruhe, Germany software was used for phase identification.

Raman thermometry is a steady-state thermal characterization technique that leverages the temperature sensitivity of a Raman scattering peak as a local thermometer. It is a non-destructive optical method that provides sub-micrometer spatial resolution, requires no microfabrication, and offers high material selectivity as its key advantages over other thermal characterization techniques. This technique has been extensively used for semiconductors such as Si,³³ GaN,^34,35 graphene,³⁶ etc. but has not yet been implemented for chalcogenides. The essential material requirements for probing a material using Raman thermometry are (a) the material must be Raman active—meaning that it should produce a significant vibrational signature in the Raman spectra from one of its optical modes and (b) the vibrational modes should be sensitive to temperature.

To analyze the vibrational modes, present in GGST, Fig. 4 illustrates the Raman spectrum of as-deposited GGST with a Lorentzian fit to extract the precise peak position. The spectra can be categorized into Sb–Te, Ge–Te, and Ge–Ge vibrations. Spectra were acquired at a very low incident power (0.5 mW) to avoid any structural modifications. The acquisition time and averaging of a spectrum were optimized for signal-to-noise ratio by trial and error.

The most intense peak (B1) in GGST at ∼158 cm⁻¹ can be attributed to the stretching mode of Sb–Te vibrations in SbTe₃ pyramidal units.^37,38 This is similar to the case of Sb₂Te₃ and GST-225, where the presence of Sb–Te vibrations was evident due to its high polarizability.

In the low-frequency range, a significant peak A1 and a minor peak A2 are evident, one around 100 cm⁻¹ and a second minor peak at 130 cm⁻¹. The first can be assigned to a contribution arising from both, Te–Te stretching mode or Ge–Te stretching mode at 90 and 110 cm⁻¹, respectively. The band at 130 cm⁻¹ associates with the symmetric stretching of Ge–Te vibrations in GeTe_4−nGe_n (n = 1,2) corner-sharing units.^39–41 The vibrations observed in GGST arise from the bonds in a distorted octahedral arrangement, like in GST-225.

In the high-frequency range, peaks related to Ge in tetrahedral units are present. Peak A3 is assigned to Ge–Te vibrations in GeTe₄ tetrahedral units at 217 cm⁻¹. This peak can be considered as a convolution arising from a LO-like (longitudinal optic) at 222 cm⁻¹. Peak A4 at 272 cm⁻¹ is assigned to the TO-like (transverse optic) mode of amorphous Ge.^42,43 The appearance of this peak confirms the presence of excess Ge and assures the amorphous nature of the material.

Figure 5 presents the Raman spectrum of GGST annealed at 400 °C for 30 min. Such annealing conditions are chosen in correspondence to the maximum thermal budget experienced by the PCM during BEOL integration. Due to crystallization, the TO-like (transverse optic) mode of amorphous Ge (A4 peak) transitions to a crystalline Ge tetrahedral peak at 299 cm⁻¹.⁴² The Sb–Te peak shows a minor shift toward higher wavenumbers. Other peaks related to Ge–Te show minor rearrangement. The presence of Ge–Ge vibrations and GST matrix combined with Ge–Te and Sb–Te vibrations hints toward the formation of GST and Ge crystal grains. Sb–Te and Ge–Ge vibrational modes are the most intense and narrow peaks in both phases.

The next step is to analyze the temperature dependence of these vibrational modes. The Raman excitation laser beam is focused on the sample's surface. Since the sample is studied under vacuum and more importantly has a relatively low thermal conductivity, the input power induces a local temperature rise T_H–T₀. Here, T_H denotes the local temperature of the hot spot and T₀ is the sample stage temperature. The sensitivity of the spectral features of Raman peaks acts as the local thermometers providing the local temperature T_H based on the Raman peak shift observed.

First, a calibration of the material is performed to extract the temperature-dependent shift of the Raman peaks. This calibration is performed by changing the sample temperature T₀ while using very low laser power to avoid local heating and local structural changes. The temperature was ramped up at 20 °C/min and at every step a 5 min of thermalization period was maintained to ensure stable temperature and vacuum conditions. Figure 6 shows such calibration curves $( ω T = d ω / dT)$ for amorphous and crystalline GGST, which gives the calibration coefficient $( ω T)$ indicating the change in Raman shift (frequency— $dω$⁠) with respect to temperature $( dT)$⁠.

The Raman spectrum was acquired at each temperature step and followed by the Lorentzian fitting procedure. The Raman shift of all the vibrational modes was analyzed as a function of temperature. The most temperature-sensitive peaks, Sb–Te and c-Ge–Ge, were chosen as “local thermometers.” They also feature the highest intensity and lowest uncertainty in the peak position. Calibration coefficients for all the materials were analyzed as listed in Table S1 in the supplementary material.

Similarly, the change in the spectral peak position as a function of increasing laser power is recorded as $( ω P = d ω / d P abs)$⁠. This coefficient, $( ω P)$⁠, gives the change in Raman shift (frequency— $dω$⁠) with respect to power absorbed in the GGST. $P abs$ is defined as $P in∗A$⁠, where A is the absorbance of the material. The input power $( P in)$ was varied, and the material was held at constant stage temperature T₀. Figure 7 presents the change in Raman shift with increasing power (red axis) for varying stage temperatures. The continuous density filter was used to vary the input power and was regulated by power measurements.

It should be noted that this quantity encompasses all the stack thermal properties in geometry, not only the layer of interest. This experimental process is repeated by ramping up the stage temperature and analyzing the material’s power dependence as presented in Fig. 7. In this way, temperature dependence of thermal conductance can be extracted.

In order to quantitatively extract the GGST thermal conductivity (κ_GGST) from G, a model must be used in order to establish an abacus of G(κ_GGST). The analytical solution of the laser-induced local heating in a semi-infinite medium is well-known.^44,45 In the case of a multilayered thin film, an analytical solution is cumbersome, thus, a finite element model (FEM) was built to quantify the heat transfer in the system induced by the Raman laser beam in COMSOL Multiphysics^®. The problem is axisymmetric, and the material properties are modeled locally dependent on the temperature. Here, the thermal conductivity of GGST (κ_GGST) is parametrized and thermal conductance is extracted as a function of κ_GGST. Temperature-dependent thermal properties are defined for the materials and interfaces. These properties are defined for the relevant thicknesses as SiN⁴⁶ and TiN⁴⁷ show thickness dependence. For thermal boundary resistance, the values are identified for SiN-GST-225 and⁴⁸ TiN-GGST⁴⁹ as available in the literature (listed in Table S2 in the supplementary material).

This provides a correlation between thermal conductance as a function of the thermal conductivity of GGST which can be extracted from the G(κ_GGST) abacus, as presented in Fig. 8.

Next, the crystalline phase thermal conductivity study was conducted on pre-annealed samples. Note that, as Raman thermometry is a steady-state technique, dynamic evaluation of thermal conductivity from amorphous to crystalline is not possible, as in the case of other widely used techniques. Figure 10 displays the thermal conductivity of the pre-annealed–crystalline phases of GGST (with SiN capping) and GGSTN (with SiN and TiN capping) from room temperature to 350 °C. The measured thermal conductivity value of GGST at room temperature is 1.02 W m⁻¹ K⁻¹, higher than GGSTN at 0.8 W m⁻¹ K⁻¹. It is evident that N-doping affects the thermal conductivity by controlling crystallization and nucleation-growth kinetics.

Next, thermal conductivity shows a temperature dependence which is contrary to the observations for GST-based materials. So, the contribution to thermal conductivity from its phonon and electron components is analyzed. Assuming that GGST behaves as a metal in its crystalline state, the electronic contribution is calculated by the Wiedemann–Franz law by using the resistivity data of the materials (listed in Fig. S1 in the supplementary material). The continuous lines shown in Fig. 10 show that the electronic contribution toward resistivity is quite low. Considering the entire range of Lorenz number from 1.5 to 2.44 × 10⁻⁸ W Ω K⁻²,²⁶ the electronic contribution at room temperature is negligible whereas at higher temperatures it is in the range of 2%–5%. This hints that the major contribution arises from the lattice contribution. So, the decreasing thermal conductivity behavior could be due to phonon scattering processes. As the temperature increases, the decreasing trend of κ for GGST and GGSTN is analogous to the behavior of highly crystalline solids. For temperatures above the one tenth of the Debye temperature, $T D =ℏ υ s ( 6 π n a ) 1 / 3/ k B$⁠, κ decreases with increasing temperature due to the onset of the three phonon Umklapp scattering process. For GST alloys, the Debye temperature is estimated to be −18 °C for GST-124 and −53 °C for GST-225.^26,51 Estimating the Debye temperature of GGST for sound velocities ν_l = 3600 m/s and ν_t = 2530 m/s, results in $T D$ as −99 °C, which is well below the temperature range in consideration confirming the possibility of the Umklapp scattering process. This behavior of the alloys with temperature is reversible as the higher room temperature conductivity is attained on cooling back to room temperature.

For other GST alloys, mainly GST-225, extensive analysis shows that lattice contribution does show temperature dependence but is compensated with the increasing electronic contribution owing to the superior electrical conductivity, maintaining constant thermal conductivity.^20,25,27 This electrical conductivity is hampered in the case of GGST due to Ge doping by two to three orders of magnitude. For GGST, as the electronic contribution is low and the lattice component is the majority contributor to thermal conductivity, temperature dependence of the total thermal conductivity is evident.

Structural analysis of GST by XRD patterns shows sharp features of cubic Ge with trigonal GST (t-GST) and cubic-GST (c-GST) phases as shown in Fig. 11. For the same annealing conditions, GGSTN shows broad features with the presence of only cubic Ge and GST phases. This suggests that in GGST, the sharper peaks indicate the growth of the initially small grains, whereas in GGSTN, the grain size remains smaller. Evaluation of Ge grain size (peak Ge [111]) results in 28.96 and 7.35 nm for GGST and GGSTN, respectively. So, N-doping affects crystallization as it bonds primarily to Ge and controls the nature of the growth of crystals. Furthermore, it restricts the formation of a more ordered trigonal structure in GGSTN as is the case for GGST. Also, the thermal budget provided for GGSTN is not sufficient to obtain a material of higher crystallinity reflecting in the thermal conductivity differences.

Similar conclusions can be drawn from the Raman spectrum presented in Fig. 12, where the peak characteristics are broader for GGSTN. The Sb–Te peak characteristics indicate a difference in the structure of GGST and GGSTN. GGST has one sharp peak at 162 cm⁻¹ and a shoulder at 175 cm⁻¹ whereas GGSTN has a rather broad peak at 154 cm⁻¹ as presented in Fig. 12. The shoulder arising at 175 cm⁻¹ could be because of the vibrations from Sb–Sb bonds arising from (Te₂)Sb–Sb(Te₂) units present at 174 cm⁻¹. This indicates a transition to a trigonal-like structure for GGST for these annealing conditions. On the contrary, for GGSTN, the Sb–Te vibrations move toward the SbTe₃ vibrations at 153 cm⁻¹. The Sb–Te vibrations arising from SbTe₃ in GGST or GGSTN are correlated to the cubic structure of GST. The presence of trigonal structural motifs and larger grain size contributes to the higher thermal conductivity of GGST. It is understood from this behavior that GGST transitions from amorphous to cubic and then to the trigonal phase, whereas the thermal budget implemented for GGSTN was not sufficient to foresee the transition to the trigonal phase. For GST alloys, trigonal structural features result in higher thermal conductivity rather than the cubic phase, and in general, for polycrystalline materials, grain size evidently affects thermal conductivity. Furthermore, a clear correlation between Raman and XRD results indicates the complementary structural analysis nature of Raman thermometry.

The thermal conductivity of Ge-rich GST and Ge-rich GST N-doped thin films has been reported using Raman thermometry. The amorphous phase thermal conductivity aligns well with the minimum thermal conductivity model and, the general expectation of thermal conductivity for GST-based alloys. In contrast, the room temperature thermal conductivity of crystalline phase drops from 1.5 W m⁻¹ K⁻¹ of GST-225 to 1.02 W m⁻¹ K⁻¹ for GGST, and further to 0.8 W m⁻¹ K⁻¹ for GGSTN. This demonstrates the significant effect of Ge-enrichment and N-doping. Additionally, the temperature dependence of thermal conductivity shows a decreasing trend, primarily due to Umklapp phonon scattering processes.

We demonstrate the applicability of Raman thermometry as a simultaneous structural and thermal characterization technique for GST. Key structural differences between cubic and trigonal phase occurring in GGST can be identified by observing the respective vibrational modes in Raman spectrum. This dual functionality of Raman thermometry eliminates the need for an additional structural characterization technique. This is key as thermal budget imposed during structural and thermal analysis can vary.

See the supplementary material for additional experimental and simulation details: (1) Calibration coefficients of GGST and GGSTN in their different phase conditions, (2) evaluation of errors for the calculation of thermal conductance, (3) methodology used for evaluating the hotspot temperature, (4) temperature-dependent material properties of materials, and (5) electrical resistivity values used for the evaluation of electronic contribution to thermal conductivity by the Wiedemann–Franz law.

This project has received funding from the ECSEL Joint Undertaking (JU) under Grant Agreement No. 101007321. The JU receives support from the European Union's Horizon 2020 research and innovation programme and France, Belgium, Czech Republic, Germany, Italy, Sweden, Switzerland, Turkey. This project has also received funding from the STMicroelectronics-IEMN common laboratory and is partly supported by the French RENATECH network. The authors would like to thank Roberto Simola, Jury Sandrini, Dominique Vignaud, and Marielle Huve for their contributions to this work.

The authors have no conflicts to disclose.

Akash Patil: Conceptualization (equal); Data curation (lead); Formal analysis (lead); Investigation (equal); Methodology (equal); Visualization (equal); Writing – original draft (lead); Writing – review & editing (equal). Yannick Le-Friec: Investigation (equal); Resources (equal). Pascal Roussel: Data curation (equal); Formal analysis (equal); Investigation (equal). Yves Deblock: Investigation (equal). Simon Jeannot: Funding acquisition (equal); Project administration (equal); Supervision (equal). Philippe Boivin: Conceptualization (equal); Funding acquisition (equal); Project administration (equal); Supervision (equal). Emmanuel Dubois: Project administration (equal); Resources (equal); Writing – review & editing (equal). Jean-François Robillard: Conceptualization (equal); Formal analysis (equal); Funding acquisition (equal); Project administration (equal); Resources (equal); Supervision (equal); Validation (lead); Visualization (equal); Writing – review & editing (equal).

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