Research Article

Femtosecond x-ray diffraction reveals a liquid–liquid phase transition in phase-change materials

See allHide authors and affiliations

Science  14 Jun 2019:
Vol. 364, Issue 6445, pp. 1062-1067
DOI: 10.1126/science.aaw1773

Structural switch for fast switching

Phase-change materials are important for computer memory. They can quickly switch from glassy to crystalline using a thermal pulse and then lock in that structure for a long time at lower temperature. Zalden et al. probed the underlying atomic structure of two phase-change materials during this switching using ultrafast x-rays and simulations (see the Perspective by Rao et al.). A liquid-liquid phase transition in both materials allowed fast switching at high temperatures. The lower-temperature glass locks in the structure, allowing for long-term memory storage.

Science, this issue p. 1062; see also p. 1032


In phase-change memory devices, a material is cycled between glassy and crystalline states. The highly temperature-dependent kinetics of its crystallization process enables application in memory technology, but the transition has not been resolved on an atomic scale. Using femtosecond x-ray diffraction and ab initio computer simulations, we determined the time-dependent pair-correlation function of phase-change materials throughout the melt-quenching and crystallization process. We found a liquid–liquid phase transition in the phase-change materials Ag4In3Sb67Te26 and Ge15Sb85 at 660 and 610 kelvin, respectively. The transition is predominantly caused by the onset of Peierls distortions, the amplitude of which correlates with an increase of the apparent activation energy of diffusivity. This reveals a relationship between atomic structure and kinetics, enabling a systematic optimization of the memory-switching kinetics.

The global amount of data grows exponentially (1) and alternative memory technologies are being considered to satisfy the resulting demands. Phase-change memory promises higher storage densities as well as faster retrieval rates due to the nondestructive readout compared with state-of-the-art memory technology (25). In phase-change memory, information storage occurs by cycling small volumes of the material between its glassy and crystalline states. An optically or electrically induced thermal stimulus allows switching between them. Crystallization is the time-limiting step as glass formation (vitrification) can be facilitated rapidly by a melt-quenching process. Nevertheless, even crystallization can take place within less than a nanosecond (68). Phase-change memory therefore relies on the kinetic contrast of the active materials: A high atomic mobility at elevated temperature allows quick crystallization, whereas low atomic mobility at ambient temperatures allows long-term data retention (7, 913). Good glass-forming materials are generally characterized by a glass transition temperature Tg that is relatively close to the melting temperature Tm. This occurs for values of the Turnbull parameter Tg/Tm ≥ 2/3 (14). Bad glass formers have typical values of Tg/Tm ≤ 1/2, which enables rapid crystallization. Optimum performance of a phase-change material (PCM) cannot be achieved by focusing on any one of these criteria, and the temperature dependence of atomic mobility, i.e., its viscosity, must be understood in detail to enable a systematic optimization of kinetic properties suitable for phase-change memory devices.

The temperature dependence of viscosity differs among various glass-forming liquids as the temperature approaches the glass transition Tg. Although some liquids show an Arrhenius-like behavior and are classified as “strong” (e.g., silica), others display a range of non-Arrhenius behaviors and are classified as “fragile” (e.g., o-terphenyl) (15). In some anomalous liquids, a fragile-to-strong (FTS) crossover may occur, in which the high-temperature fragile liquid is transformed into a low-temperature strong liquid (1618). The FTS crossover is usually associated with a maximum in thermodynamic response functions [e.g., heat capacity (Cp), thermal expansivity (αP), and compressibility (κT)], which may be attributed to a phase transition between two liquid phases characterized by distinct physical properties (e.g., density and entropy) and different atomic structures. The FTS crossover was first proposed to explain the behavior of water, which is a fragile liquid down to the temperature of the homogeneous nucleation limit but behaves kinetically strong when heated from the solid amorphous state above Tg (19). Similar observations were made on the PCM Ag-In-Sb-Te in its liquid and solid amorphous states (9, 12, 20). However, in PCMs, liquid quenching is difficult owing to the rapid onset of crystallization. This limits the accessible supercooling range at common cooling rates (21), making observations of FTS crossovers or liquid–liquid phase transitions (LLPTs) difficult. The apparent FTS crossover in PCMs renders the hypothesis of LLPTs appealing and also implies anomalous structural, thermodynamic, and diffusive properties (2123).

The microscopic mechanism underlying the crossover is unclear. Liquid Ge15Te85 shows an increase of medium-range order (MRO) upon supercooling, compatible with an increase of local atomic distortions (24), yet Ge15Te85 is a good glass former and thus cannot be used as a PCM. More generally, direct experimental evidence for LLPTs in systems that crystallize rapidly is rare because the time scales available to probe the supercooled liquid state before it crystallizes are short. Therefore, the name “no-man’s land” was coined to refer to the highly supercooled region (150 to 236 K) (25) in the phase diagram of water, whose anomalous properties have long been debated (26). Only recently, ultrafast x-ray laser scattering has enabled directly probing supercooled water down to 227 K and revealed evidence of a continuous LLPT (27). From the computational side, LLPTs were observed previously in simulations of ST2 water (28, 29), although the possibility of the existence of a true liquid–liquid critical point (LLCP) in the supercooled liquid has been challenged theoretically (30). Similar data are not yet available for PCMs, but an anomalous breakdown of the Stokes–Einstein relationship (SER) in the equilibrium liquid of GeSb2Te4 was recently observed and was speculated to be caused by an LLPT below Tm (31). We focused our attention on Sb-based PCMs because Ag4In3Sb67Te26 (AIST) (9, 12, 20) and Ge9Sb91 (32) have crossovers in the apparent activation energy of viscosity. Both materials are used as PCMs in optical [AIST (4)] and electronic [Ge15Sb85 (33)] memory devices.

The liquid–liquid phase transition

We used hard x-ray laser pulses from an x-ray free electron laser (the Linear Coherent Light Source, LCLS) to overcome the temporal limitation of probing the atomic structure of highly supercooled PCMs. With the X-ray Pump Probe (XPP) instrument (34), optical laser pulses with 800-nm center wavelength and 50-fs pulse duration are absorbed by a thin film of PCM at time t0. Subsequently, electron–phonon coupling heats the PCM on the few-picosecond time scale (35, 36), leading to melting at sufficiently high excitation fluences. The PCM is quenched by diffusive thermal transport into the supporting membrane of Si3N4, which is of similar thickness as the PCM. To probe the atomic structure during this melt–quench cycle, we collected the diffraction patterns of x-ray pulses with 1.305-Å center wavelength on a two-dimensional area detector (fig. S1) (37). We varied the delay Δt between the optical pump event at t0 and the x-ray probe event at time t, Δt = tt0, between negative values and tens of microseconds. We performed each set of pump–probe events on a new spot of the sample, which was irreversibly modified as a result of the intense optical and x-ray pulses. Nevertheless, the x-ray probe pulse duration of 50 fs is sufficiently short to ensure that the atomic structure does not change during the probe interaction.

We found that the structure factor S(q) of amorphous AIST was dominated by two broad diffraction rings centered at q1 = 2.07 (1) Å−1 and q2 = 3.21 (2) Å−1, clearly visible in Fig. 1, A, B, and D. At Δt < 0, the diffraction pattern of the unpumped, as-deposited amorphous structure is recorded at the initial temperature of 298 K. At Δt of a few picoseconds, the peak intensity of both rings decreases and their radii approach each other. Consistently, the ratio q2/q1 depicted in Fig. 1F decreases within the first picoseconds, but almost recovers to the initial value of the amorphous state after a few nanoseconds. This behavior is strong evidence for a structural transition in a disordered state beyond just a thermally induced reduction of the scattering efficiency (Debye–Waller), which would leave the peak positions unaffected. Further evidence for a phase-transition behavior comes from the highly nonlinear scaling of the peak structural modification with the pump fluence (Fig. 1G). The finding that crystallization eventually occurs at fluences above 14 mJ/cm2 (see fig. S4) enables us to derive a lower bound for the temperature jump several microseconds after optical excitation. At the time of crystallization (5 μs), the temperature in the PCM must be above the temperature at which crystallization sets in at calorimetric heating rates of a few kelvin per minute. In AIST, this temperature is 430 K (38). A comparison with literature data on AIST (vertical gray lines) shows good agreement of the momentum transfer associated with the strongest reflections (Fig. 1E).

Fig. 1 Evolution of the atomic structure of AIST during the melt–quench cycle, resolved in reciprocal space.

(A) Structure factor S(q) of AIST after optical excitation with 24 mJ/cm2 throughout the entire melt–quench cycle from the initial, as-deposited amorphous state (B) to the final crystalline state (C). The atomic structure factors of the initial and final states are in good agreement with reference data from total neutron scattering (ND) (58) and literature values for the crystalline structure (D and E). Femtosecond x-ray diffraction (XD) enables resolving an intermediate structural phase transition in the disordered state of AIST, as seen by the change in the ratio q2/q1 corresponding to a shift of reflections after a few picoseconds and again after a few nanoseconds (F). (G) Fluence dependence of the average q2/q1 after 10 to 100 ps showing the onset of the phase transition at 12 mJ/cm2, also shown in (A) by white horizontal arrows.

To quantify the temperature evolution of our samples more accurately, we performed finite element simulations, which allow temperatures to be associated to the structural information that we obtained at various time delays and fluences. We show the resulting cooling behavior of the 50-nm–thick film of AIST on 50-nm–thick membranes (red curves) based on the normalized temperature Θ=TT0TmT0, where Tm is the melting temperature, 810 K for AIST (39), and T0 is the initial temperature of 298 K (Fig. 2). After the heating by electron–phonon coupling, the heat diffuses from the PCM into the Si3N4 membrane after a few hundreds of picoseconds. Their temperature equilibrates after a few nanoseconds. Subsequently, thermal transport into the frame supporting the membranes becomes the dominant cooling mechanism. The experimental data (Fig. 1F) reflect the impact of the two different length scales of thermal transport associated with the short out-of-plane and the long in-plane axes. The time scale of out-of-plane transport is in good agreement with the change of atomic structure in the disordered state at ~1 ns. We observed in-plane thermal transport only after delays longer than 10 μs, rendering it irrelevant for this work. We derived two temporal intervals ({ta} and {tb}, indicated in Fig. 2), independent of the pump fluence, over which the temperature of the PCM was stable within 10%. During these intervals, we can predict the temperature of the PCM most accurately because it depends predominantly on the specific heat of the materials involved. We found the same intervals for 60-nm–thick Ge15Sb85 [Tm = 860 K (40)] films on 150-nm–thick membranes.

Fig. 2 Temporal evolution of the average temperature inside AIST (red curves) and Ge15Sb85 (blue curves) after optical excitation for different optical excitation conditions.

We identified two intervals (gray areas), one between 10 and 100 ps {ta} and the other between 10 and 30 ns {tb}, during which the temperature is stable within 10%. These time scales are associated with two thermal transport conditions, schematically depicted as insets (yellow arrows denote the main direction of heat flow) and are caused by the high aspect ratio of the samples.

The atomic structure during {ta} and {tb} is described by the radii of the first (r1) and second (r2) coordination shells. We determined the radii from the pair-correlation functions g(r) (Fig. 3A) and they revealed a pronounced increase of r1 upon excitation. The ratio R = r2/r1 provides information about the structural ordering and we measured its evolution with time and fluence (Fig. 3, B and C).

Fig. 3 Average local structure in AIST and Ge15Sb85 during the melt–quench process.

(A) Atomic pair correlation functions g(r) of AIST as a function of time after optical excitation with 24 mJ/cm2. The structural transition is most clearly evidenced by a transient shift of the first coordination shell r1 to longer distances between ≈1 ps and ≈5 ns, well before the onset of crystallization after ≈5 μs. Consequently, the structural parameter r2/r1 for AIST (B) and Ge15Sb85 (C) decreases and stays at the high-temperature value of 1.36 for several nanoseconds when the fluence is sufficiently high. If crystallization is avoided at low pump fluences below 14.5 mJ/cm2, then r2/r1 eventually returns to its initial values (1.51 ± 0.01 in the case of AIST) in the final amorphous solid state for t → ∞, when ambient temperature is restored.

On the basis of these numbers and in combination with the thermal simulations, we determined R for AIST and Ge15Sb85 as a function of inverse temperature Tm/T (Fig. 4A). For each fluence of the pump laser, we derived one value of R from the intervals {ta} and {tb}. Additionally, we derived one average value from the data at negative delays. Direct evidence for a structural transition in the supercooled liquids of AIST and Ge15Sb85 comes from a change in slope of R(T) in the range of Tm/T between 1 and 1.5. This resembles the R(T) behavior in the good glass former Ge15Te85 that also has an FTS crossover and a structural phase transition that correlate (24), albeit at higher temperatures. We interpolated our data with an error function that allowed us to determine the temperature of the structural transition by the maximum slope in r2/r1 occurring at 660 ± 20 K and 610 ± 20 K for AIST and Ge15Sb85, respectively. We estimate the experimental uncertainty as ±20 K SD in our experiment mostly caused by statistical fluctuations in the pump fluence. A normalized version of the same error function fits the prepeak intensity Ipp (Fig. 4B and fig. S3) in S(q), which was located at q = 1.08 ± 0.02 and 1.06 ± 0.02 Å−1 for AIST and Ge15Sb85, respectively. They correspond to the formation of periodic structures in real space with 5.8 and 5.9 Å, which is twice the radius of the first coordination shell r1. This periodicity is due to the formation of alternating long and short bonds on opposite sides of a central atom, which is a characteristic fingerprint of a Peierls distortion (41). The pronounced increase of g(r) around the second coordination shell indicates that the distribution of these interatomic distances, being the hypotenuses of ~90° bond angles, becomes narrower (figs. S7 and S8). The temperature dependence of R and Ipp matches the temperature dependence of the apparent activation energy of inverse diffusivity D–1 (Fig. 4C), which corresponds to the apparent activation energy of viscosity η, assuming that the SER is valid.

Fig. 4 Structural parameter R = r2/r1, the intensity of a pre-peak, and the apparent activation energy of diffusivity correlate in the case of AIST.

(A) Ratio of second and first coordination shell radii, r2/r1, showing a transition for the two PCMs AIST (red) and Ge15Sb85 (blue) in the supercooled regime. Both were refined by an error function as a guide for the eye and to determine the structural transition temperatures TLL (inflexion points) of 660 and 610 K, respectively, visualized by vertical dotted lines. Vertical dashed lines represent the temperature at which ~90% of the low-temperature structure is reached and coincide with the FTS crossover temperature TFTS. Crosses and circles denote the data points obtained from{ta} and {tb}, respectively, and triangles are obtained from the initial structures. A similar transition was reported for Ge15Te85 (gray) at the melting point (24). (B) Simultaneously, a prepeak is formed, which is related to additional MRO caused by a Peierls distortion. Both structural parameters are found to correlate with an increase of the apparent activation energy EA of inverse diffusivity (C). The latter data are derived from measurements of viscosity above Tm [crosses (43)] and from crystal growth velocities in the supercooled liquid [triangles (12)] and glassy regime [circles (9) and squares (59)], as shown in (D). The solid line corresponds to the model proposed by Orava et al. describing the FTS crossover between two liquids (20). Glass transition temperatures (diamonds) are shown for Ge15Te85 (60) and for AIST (45). For the latter, literature values show a wide spread. Error bars in this figure correspond to the SD; errors on temperatures were omitted for clarity.


We derived the diffusivities (Fig. 4D) from previous measurements of the viscosity above Tm for AIST (42) and Ge15Te85 (43). In the case of AIST, we also calculated diffusivities from the crystal growth velocity in the supercooled liquid state (12). Because of the uncertainty regarding the validity of the SER in the supercooled regime, we prefer to provide diffusivity data, which, unlike viscosity data, can be directly calculated from crystal growth velocities (9) and require the assumption of a valid SER only for the equilibrium liquid.

Two different temperatures characterized the correlation of apparent activation energies and atomic structure. The structural transition temperature (TLL) occurs at the inflection points of R and Ipp (Fig. 4, A and B). In the non-PCM Ge15Te85, it occurs at TLL/Tm = 1.05 (24, 44), whereas the LLPT of the PCMs AIST and Ge15Sb85 takes place at TLL/Tm = 0.81 ± 0.01 and 0.71 ± 0.01, respectively. TLL coincides with the temperature at which the apparent activation energy reached ~20% of that for the low-temperature state, as shown in Fig. 4C for AIST and Ge15Te85. For Ge15Sb85, no kinetic data are available from literature. The FTS crossover temperature (TFTS, Fig. 4, C and D) was reported to be 570 K in the case of AIST and in the vicinity of the melting temperature in Ge15Te85. It coincides with the temperature at which the atomic structure reaches ~90% of R for the low-temperature state in the case of AIST and Ge15Te85. In Ge15Sb85, this value of R would correspond to a temperature of 550 ± 20 K.

In this study, we provide direct evidence for the structural transition and emphasize its relevance for the kinetic properties, although the structural transition occurs at higher temperatures than the FTS crossover reported previously. The state for the kinetic data below the FTS crossover is still a matter of debate because of experimental difficulties in determining its kinetic properties. Orava et al. reported a strong liquid state of AIST based on Tg = 378 K for a standard cooling rate of 20 K/min (20). The kinetic data underlying this scenario, however, agree with data reported by Salinga et al. (9) for glass (Fig. 4D). In an alternative scenario in which the glass transition for similar standard rates is located at 450 K (45), the strong liquid is represented hypothetically by the dashed and dotted line in (Fig. 4D), similar to the case of Ge15Te85. In both scenarios, however, the structural transition observed in this study correlates with the kinetic crossover. To show that the structural transition at TLL indeed occurs between two liquid states, it is necessary to discuss the role of the glass transition. Given the reported glass transition of AIST at 450 K (45) for calorimetric heating rates of ~1 K/s, an enormous cooling-rate dependence of Tg is required to explain a transition at 660 K. This requirement, however, is in contradiction to the invariant transition temperatures observed during the out-of-plane cooling during {ta} with a 1011 K/s cooling rate and the in-plane cooling during {tb} with a rate of 106 K/s (fig. S5c). Therefore, the structural transition observed here is an LLPT at temperature TLL.

Microscopic nature of the transition

Liquid PCMs are commonly found to be octahedrally coordinated (46), but their corresponding amorphous states contain several competing structural motifs. Local Peierls distortions of an octahedral environment cause the splitting of the first coordination shell into two subshells with unequal occupation of the first and second subshells (47). Although the Peierls distortion itself leaves R almost unchanged, the lower occupation of the second (larger) subshell will reduce r1, thereby increasing R. This makes it impossible to distinguish this motif from an increase caused by the formation of tetrahedral sites, possibly together with homopolar bonds as reported for GeTe (48, 49). Therefore, we require further information on the local structure to resolve the exact mechanism responsible for the LLPT. We performed ab initio molecular dynamics (AIMD) simulations and verified the simulations by comparing them with the temperature-dependent structure factors (fig. S6). We reproduced with our simulations the peak shift and the reduction of intensity. We found that R was less temperature dependent, which we attributed to the faster quench rate in the simulation (3 × 1012 K/s) (13). At higher quenching rates, the liquid falls out of equilibrium at higher temperatures, kinetically freezing-in an intermediate state. The lack of equilibration time at each temperature step means that the LLPT occurs over a wider temperature range. Nevertheless, the trend observed in AIMD was consistent with our experimental findings.

In the case of Ge15Sb85, 70% of the total number of bonds were formed between Sb atoms, which dominated the local order and the scattering signal due to their higher atomic weight. Histograms of interatomic distances from an Sb atom to its nth nearest Sb neighbor provide further insight. As Sb commonly is sixfold coordinated, the histograms reflect the inner structure within the first coordination shell. In the liquid state (Fig. 5B), all six histograms have equidistant peaks of similar width, indicating a regular octahedral environment with regular fluctuations due to the dynamics in the liquid state. After quenching (Fig. 5C), the n = 1 to 3 histograms shift to shorter distances, becoming narrower and higher. The n = 4 to 6 histograms get separated further and retain their width from the liquid state. This different behavior between the three shorter and three longer interatomic distances is evidence for the onset of a Peierls distortion, also evident from the angular limited bond correlation (ALTBC) plots (Fig. 5, E and F). The large variation of first-neighbor distances limits the discussion of R to a qualitative level, because the sixth-nearest neighbor distance in low-temperature Ge15Sb85 (Fig. 5C) is centered at 3.7 Å, which is closer to the center of mass for the second coordination shell (r2 = 4.2 Å) than the first (r1 = 2.9 Å). This means that the first and second coordination shells can no longer be resolved because of the inherent disorder. Although the Peierls distortion alone can explain the increase in R, the bond angles around Ge change from the octahedral value of 90° in the liquid state to a value of 105° at ambient conditions (fig. S14). This indicates the formation of tetrahedral sites (R = 1.61 for a purely tetrahedral structure), which also increase R. The conditions in AIST are comparable, but tetrahedral sites are formed during the quench only around a fraction of the 4% stoichiometric In atoms (13). Because AIST shows the same structural trend during the quench (Fig. 5A), a Peierls distortion must be the dominant mechanism responsible for the change in kinetic properties. With increasing Peierls distortion in supercooled liquid Ge15Sb85 and AIST, we also observed a tendency for a pseudogap (50) opening in the density of states (Fig. 5D and fig. S27). Supercooled liquids Ge15Te85 (51), As2Se3 (52), and As2Te3 (53) indeed undergo a metal–semiconductor transition during the quench, accompanied by a maximum of the thermodynamic response functions (e.g., Cp, αP, and kT), indicative of a fragile-to-strong crossover. Also, the pseudogap of liquid Ge2Sb2Te5 exhibits an opening tendency during the quench, suggesting that it might also show an LLPT in the supercooled liquid regime (54).

Fig. 5 AIMD simulations of melt–quenching Ge15Sb85 and AIST showing the onset of Peierls distortions upon cooling.

(A) We observed that R increased continuously during the quench, with initial and final values similar to experimental values, except that the transition is wider in the simulation because of the higher quench rate. (B and C) Nearest-neighbor histograms of Sb atoms in Ge15Sb85 showing a continuous sixfold coordination in the high-temperature liquid state, whereas they split into three short and three more widely spaced distances in the quenched phase. (D) We observed the opening of a pseudogap during the quench in Ge15Sb85, but also in AIST (fig. S27). (E and F) ALTBC plots providing evidence for the formation of a Peierls distortion by including only the interatomic distances on opposite sides of Sb atoms.

Discussion and conclusion

We combined our observations and modeling to derive information on the change of the bonding mechanism underlying the LLPT. In the equilibrium liquid state, metallic bonds are nondirectional with a relatively low activation energy of viscosity. The formation of Peierls distortions during the quench localizes electronic charge between the atoms, constraining bond angles and increasing the activation energy of viscosity. This and the correlation between structure and kinetics imply that the LLPT constitutes the structural origin of the kinetic crossover. The tendency to form a Peierls distortion in the crystalline state is one of the characteristic attributes of PCMs, which are classified as incipient metals because of the sensitivity of their electronic localization on the distortion amplitude (55, 56).

Ultrafast x-ray diffraction after short-pulse excitation provides access to the atomic structure of PCMs over the entire melt–quench cycle without interference by crystallization. The resulting diffraction data reveal a structural LLPT in AIST and Ge15Sb85. We used AIMD simulations to show that this LLPT in both PCMs is dominated by the formation of a Peierls distortion, forming distinct short and long bonds and opening a pseudogap in the electronic density of states. Our observation is consistent with earlier predictions that the atomic energy gain by a Peierls distortion ΔEp<kBTm (57), because we show that the distortion is formed at TLL, which implies ΔEpkBTLL. The temperature-dependent structural parameters correlate with the apparent activation energy of inverse diffusivity, which suggests that the LLPT is responsible for the FTS crossover previously reported in AIST. The kinetic transition associated with the FTS crossover has important implications for the application of PCMs in memory devices: The low kinetic activation energy of the high-temperature liquid ensures that the high atomic mobility of the equilibrium liquid is available for crystallization, i.e., for temperatures between Tm and TLL. At temperatures below the LLPT, the Peierls distortion localizes charge and stabilizes the amorphous atomic structure against crystallization, which increases the kinetic activation energy and rapidly reduces the atomic mobility during the quench, enabling the formation of a glass that is stable at ambient conditions. A material with lower TLL/Tm therefore offers a wider temperature window between TLL and Tm, where fast crystallization is possible, and explains why AIST and Ge15Sb85 are PCMs, whereas Ge15Te85 is not. Our results offer a new strategy for the design of improved PCMs for specialized memory applications based on the atomic-bonding mechanism in these materials.

Supplementary Materials

Materials and Methods

Figs. S1 to S27

Tables S1 to S3

References (6282)

References and Notes

  1. See supplementary materials.
Acknowledgments: P.Z. thanks Christophe Bichara for valuable suggestions and discussions. Use of the Linac Coherent Light Source (LCLS), SLAC National Accelerator Laboratory, is supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, under contract no. DE-AC02-76SF00515. Funding: F.Q., A.K., M.N., and K.S.T. gratefully acknowledge financial support from the German Research Council through the Collaborative Research Center SFB 1242 project 278162697 (“Non-Equilibrium Dynamics of Condensed Matter in the Time Domain”), project C01 (“Structural Dynamics in Impulsively Excited Nanostructures”), and individual grant So408/9-1, as well as the European Union (7th Framework Programme, grant no. 280555 GO FAST). M.J.S., R.M., and M.W. acknowledge financial support from the German Research Council through the Collaborative Research Center SFB 917 (“Nanoswitches”) and individual grant Ma-5339/2-1. M.J.S., I.R., and R.M. also acknowledge the computational resources granted by JARA-HPC from RWTH Aachen University under project nos. JARA0150 and JARA0183. M.T., A.M.L., and D.A.R. were supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences, through the Division of Materials Sciences and Engineering under contract no. DE-AC02-76SF00515. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under contract DE-AC52-07NA27344. J.L. acknowledges support from the Swedish Research Council. J.S. acknowledges financial support from the Spanish Ministry of Science, Innovation and Universities through research grant UDiSON (TEC2017-82464-R). P.Z. gratefully acknowledges funding by the Humboldt Foundation. Author contributions: K.S.T. initiated the project. F.Q., K.S.T., and P.Z. conceived the experiment. P.Z. and M.W. were responsible for sample preparation. F.Q., P.Z., A.K., M.N., J.S., M.T., P.A., H.E., M.J.S., T.P., J.L., A.M.L., S.H.R., D.A.R., and K.S.T. performed the experiments. M.J.S., I.R., and R.M. conceived, performed, and evaluated the AIMD simulations. M.C., H.L., and D.Z. operated the XPP instrument. P.Z. analyzed the data with important input from K.S.T., F.Q., A.K., M.N., P.A., H.E., and J.S. H.E.F. and P.Z. conducted and evaluated the neutron total scattering measurements. P.Z. and K.S.T. wrote the manuscript with specific contributions from A.M.L., M.W., and S.W. and comments from all other authors. Competing interests: The authors declare no competing interests; Data and materials availability: The experimental data, scripts for data analysis, molecular dynamics trajectories, and the raw data from Figs. 1 to 5 are available from Zenodo (61).
View Abstract

Stay Connected to Science

Navigate This Article