You are on page 1of 12

Microscopic origin of excess wings in relaxation spectra of deeply supercooled liquids

Benjamin Guiselin*,1 Camille Scalliet*,2 and Ludovic Berthier1, 3


1
Laboratoire Charles Coulomb (L2C), Université de Montpellier, CNRS, 34095 Montpellier, France
2
Department of Applied Mathematics and Theoretical Physics,
University of Cambridge, Wilberforce Road, Cambridge CB3 0WA, United Kingdom
3
Department of Chemistry, University of Cambridge,
Lensfield Road, Cambridge CB2 1EW, United Kingdom
(Dated: March 3, 2021)
We combine the giant equilibration speedup provided by the swap Monte Carlo algorithm to
extensive multi-CPU molecular dynamics simulations to explore the first ten decades of the equi-
arXiv:2103.01569v1 [cond-mat.soft] 2 Mar 2021

librium dynamics of supercooled liquids near the experimental glass transition. We demonstrate in
particular the emergence of a power law in relaxation spectra, in quantitative agreement with the
excess wings observed experimentally. Detailed analysis of particle motion shows that this power law
directly originates from broadly distributed processes experienced by rare, localised regions relaxing
much before the bulk. We explain why these wings appear in “excess” by constructing an empirical
model associating heterogeneous activated dynamics and dynamic facilitation, which appear to be
the two minimal physical ingredients revealed by our simulations.

The formation of amorphous solids results from the universality is debated [4, 5], the presence of an excess
rapid growth of the structural relaxation time τα of the contribution taking the form of a wing is not [12–14].
supercooled liquid [1]. Molecular motion occurs on a Elucidating the nature of molecular motions responsi-
timescale of about 10−10 s at the onset temperature of ble for the small signal in excess wings appears daunt-
glassy behaviour but takes about 100 s at the experi- ing even though experiments managed to characterise its
mental glass transition temperature Tg [2]. Over the last heterogeneous nature [15, 16] and aging properties [17].
decades, dielectric spectroscopy kept developing to probe So far, computer simulations have been unable to ac-
molecular motion over a broader frequency range with cess the required range of equilibration temperatures and
increased accuracy [3]. This progress reveals that the timescales to even address the question. Physical inter-
temperature evolution of τα is just the tip of the iceberg, pretations and empirical models have been proposed to
as relaxation spectra χ00 (ω) measured near Tg exhibit re- explain the shape of relaxation spectra. Some of them
laxation processes taking place over an extremely large couple slow translation to an “additional” degree of free-
frequency window [4–7]. The overall shift of relaxation dom (e.g., rotational) [18–20]. Others invoke spatially
spectra is accompanied by an equivalent broadening of heterogeneous dynamics to construct a broad distribu-
about 12 decades, which is the other side of the same tion of timescales of static [21–25] or kinetic [26] ori-
coin. A microscopic explanation of these slow dynamics gin. The winged asymmetric shape then requires specific
is at the heart of glass transition research [1]. physics, such as geometric frustration [23], lengthscale-
High-temperature spectra reflect near exponential re- dependent dynamics [24], or dynamic facilitation [26].
laxation in the picosecond range, but low-T spectra With specific choices, these approaches yield relaxation
broaden into a two-step process with a stretched expo- spectra comprising excess wings, but direct microscopic
nential relaxation at low frequency ω ≈ 1/τα and a mi- investigations testing the underlying hypotheses are lack-
croscopic peak remaining at the picosecond timescale. In ing.
1990, Nagel and coworkers [6–9] showed that for a num-
Here, we show that computer simulations can now di-
ber of molecular liquids the structural relaxation peak
rectly observe excess wings and assess their microscopic
extends much further at high frequencies ωτα  1 and
origin. We take advantage of the recent swap Monte
transforms into a wide power law, χ00 (ω) ∼ ω −σ , with
Carlo algorithm [27] to efficiently produce equilibrated
a small exponent σ(T ) ∈ [0.2, 0.4] decreasing at low
configurations of a supercooled liquid with τα ≈ 100 s.
T [9]. Using logarithmic scales, this resembles a “wing”
We observe their physical relaxation dynamics over ten
in “excess” of the α-peak. At Tg , the wing extends over
decades in time, up to 20 ms. We are thus able to probe
the mHz-MHz range with an amplitude about 100 times
for the first time the temperature and time regimes where
smaller than the α-peak. A universal scaling compris-
excess wings are observed in experiments. We report the
ing the excess wing was proposed [9], which can be al-
emergence of a power law (a wing) in numerical spec-
tered by additional microscopic processes [10, 11]. While
tra with the same characteristics as in experiments. We
demonstrate that it is caused by a sparse population of
localised regions, whose relaxation times are power law
*Equal contributions. distributed. These relaxed regions then coarsen by dy-
2

namic facilitation. We construct an empirical model to


show that heterogeneous dynamics and dynamic facilita-
tion generically lead to winged relaxation spectra.
We study a size-polydisperse mixture of N soft
spheres [27] of mass m interacting with a repulsive power
law pair potential v(rij )/ = (σij /rij )12 + vc (rij /σij ) in
dimensions d = 2, 3; the fourth-order polynomial vc reg-
ularises the first two derivatives of the potential at the
cutoff rij /σij = 1.25. The diameters σi are drawn from
the distribution P(σ) ∝ σ −3 , and non-additive interac-
tions are used, σij = 0.5(σi + σj )(1 − η|σi − σj |), to
enhance glass-forming ability [27]. The average particle
diameter σ sets the unit length, and η = 0.2. We set
the Boltzmann constant to unity, expressp temperatures
in units of  and times in units of mσ 2 /. We sim-
ulate the system in a cubic box, use periodic boundary
conditions and set the number density to unity.
Using the swap Monte Carlo algorithm designed in
Ref. [28], we can generate ns ∈ [200, 450] independent
equilibrium configurations at temperatures T ranging
FIG. 1. (a) Self-intermediate scattering function Fs (t) at vari-
from the onset temperature down to Tg . Each equilib- ous temperatures. (b) Relaxation time τα rescaled by its value
rium configuration is then taken as the initial condition of τo at the onset temperature. Symbols correspond either to di-
a multi-CPU molecular dynamics (MD) simulation (with- rectly measured data, or are obtained using time temperature
out swap) with time step dt = 0.01. The ns indepen- superposition (TTS). A conservative Arrhenius extrapolation
dent simulations run for up to 2 × 109 MD steps. We locates Tg = 0.056, where τα (Tg )/τo = 1012 (dashed line).
push some of them to 6 × 1010 MD steps, representing (c) Relaxation spectra for the same temperatures as in panel
(a). The dashed lines represent the estimated α-peaks. Close
a total wall time of several months, and a physical time
to Tg , spectra display a power law signal with an exponent
of about 20 ms. Both systems behave similarly so we σ ≈ 0.38 (full line), in quantitative agreement with excess
present quantitative results for d = 3 using N = 1200, wings observed experimentally.
and provide 2d visualisations using N = 104 .
We measure

P the self-intermediate
scattering function
Fs (t) = N1 i cos [q · δri (t)] , where δri (t) is the dis- regime below the mode-coupling crossover Tmct ≈ 0.095.
placement of particle i over time t. The brackets indi- The latter is determined by a power law fit of τα (T ) in the
cate the ensemble average over ns independent runs and range τα /τo < 103 with τo ≈ 3 the value of τα at the onset
an angular average over wavevectors with |q| = 6.9 (the temperature To ≈ 0.20 [27]. At all T , the correlations dis-
location of the first peak in the structure factor). The play a fast initial decay near t ≈ τo , due to thermal vibra-
relaxation time τα is defined by Fs (τα ) = e−1 . In 2d, we tions. At larger times, we observe a much slower decay
define τα using the bond-orientational correlation [29]. to zero. As T decreases, the relaxation time grows and
In the frequency domain we compute a dynamic sus- eventually exits the numerically accessible time window.
ceptibility χ00 (ω) from a distribution of relaxation times At the lowest investigated temperatures near Tg , correla-
G(log τ ) [13, 26] tions appear almost constant over more than 7 decades,
ˆ ∞ suggesting near-complete dynamic arrest. We recall that
00 ωτ
χ (ω) = G(log τ ) d log τ , (1) thanks to the swap algorithm, all measurements reflect
−∞ 1 + (ωτ )2
equilibrium dynamics, even when τα is larger than the
relating the distribution G to time correlation functions, simulated time by many orders of magnitude.
G(log t) ≈ −dFs (t)/d log t. To resolve structural relax- Our strategy allows us to directly measure τα /τo . 107
ation at the particle scale, we compute the bond-breaking down to T = 0.0793, see Fig. 1(b). In this regime, the re-
i laxation is well-described by a stretched exponential with
correlation 0 ≤ CB (t) ≤ 1, which represents the frac-
tion of neighbours lost by particle i since time 0, and an almost-constant stretching exponent β ≈ 0.56. We use
follow known geometric recipes to define neighbours at this time temperature superposition (TTS) property to
times 0 and t [30]. As relaxation proceeds, particles ex- estimate τα for 0.07 ≤ T ≤ 0.0793, where the final decay
perience rearrangements which affect their local environ- of Fs (t) remains visible [31], obtaining τα over 2 more
i decades. We finally use an Arrhenius law to extrapolate
ment. Therefore, CB (t = 0) = 1 and fully relaxed parti-
i τα over 4 more decades to get a safe lower bound for the
cles after t have CB (t) = 0.
We start with equilibrium measurements of Fs (t) in experimental glass temperature, Tg ≈ 0.056 [27].
3d in Fig. 1(a), concentrating on the unexplored low-T The corresponding relaxation spectra are shown in
3

Fig. 1(c). They all display a peak at high frequency


ω ≈ 1/τo , corresponding to the short-time decay of Fs (t).
A low-frequency peak near ω ≈ 1/τα is also visible. As
T decreases, this α-peak shifts to lower frequencies and
eventually exits the accessible frequency window. When
the α-peak is not directly measured, we extrapolate its
shape inserting a stretched exponential fit of Fs (t) into
Eq. (1) using β = 0.56, and τα given by the Arrhenius ex-
trapolation. The resulting α-peaks are shown in Fig. 1(c)
with dashed lines that smoothly merge into the measured
data at the highest temperatures.
As T decreases, the measured susceptibility and the
α-peak deviate increasingly from one another, the data
being systematically “in excess” of the α-peak. Since our
extrapolation (at worst) underestimates τα , this excess is
(at worst) slightly underestimated. At the lowest T when
the α-peak no longer interferes with the measurements,
the spectra follow a power law χ00 (ω) ∼ ω −σ at low fre-
quencies, with an exponent σ ≈ 0.38 slightly decreasing
with T and an amplitude about 100 times smaller than
the α-peak. It is impossible to interpret the signal as an
additive “β-process”, as spectra do not reveal any peak
and the small value of the exponent would lead to a fitted
secondary process much broader than the main α-peak,
which seems unphysical. Therefore, close to Tg , the nu-
merical spectra obey a power law over a similar frequency
range, with a similar exponent and a similar amplitude as
the excess wings obtained experimentally, showing that
even simple glass models display excess wings resembling
observations in molecular liquids.
We take advantage of the atomistic resolution offered
by simulations to explore the microscopic origin of ex-
FIG. 2. Relaxation in the 2d system at T2d = 0.09 with
cess wings and physically understand spectral shapes. In
τα /τo = 108 . Frames are logarithmically spaced between t =
Fig. 2 we show 2d snapshots illustrating how structural 2×10−3 τα (top left) and t = 0.6τα (bottom right) from left to
relaxation proceeds at a temperature T2d = 0.09 (we esti- right and top to bottom. Particles are coloured at their initial
mate Tg,2d ≈ 0.07) for which τα /τo ≈ 108 (corresponding position according to CB i i
(t) from blue [immobile, CB (t) = 1]
i
to about 10 ms in physical time). This temperature is to red [relaxed, CB (t) = 0].
the lowest for which the α-relaxation can be observed
in the numerical window, and is considerably lower than
the mode-coupling crossover near Tmct,2d ≈ 0.12. Im- nearby particles. Also, the slowest regions are typically
ages are shown  at logarithmically
 spaced times t in the “invaded” at t  τα from their boundaries. Dynamic
range t/τα ∈ 10−3 , 1 . Particles are coloured according facilitation has been identified before at high tempera-
i tures above the mode-coupling crossover [32–34]. Our
to CB (t): red particles have relaxed, blue ones have not.
For t  τα , relaxation starts at a sparse population of investigations show that it becomes a central physical
localised regions which emerge independently throughout mechanism for structural relaxation near Tg .
the sample over broadly distributed times. This conclu- We concentrate on the early times where power law
sion holds over a large range of temperatures down to Tg spectra are observed. Visualisation suggests that clus-
in both d = 2, 3. As time increases, newly relaxed regions ters of relaxed particles appear at sparse locations. We
continue to appear, but a second mechanism becomes ap- now establish that these relaxation events are respon-
parent in Fig. 2 as regions that have relaxed in one frame sible for the excess wing. To this end, we define mo-
i i
typically appear larger in the next. The further growth bile (CB < 0.55) and immobile (CB ≥ 0.55) particles;
of relaxed regions in Fig. 2 is the signature of dynamic the threshold value near 0.5 is determined requiring self-
facilitation [32]. More precisely, we observe that from one consistency with alternative mobility definitions based on
frame to the next, relaxation events keep accumulating at displacements. We identify connected clusters of mobile
similar locations, which results in mobile particles under- particles by performing a neighbour analysis, and investi-
going multiple relaxations and mobility propagating to gate the statistical properties of relaxed clusters. In par-
4

FIG. 3. Waiting-time distribution of newly relaxing clusters


in d = 3 for temperatures as in Fig. 1(a). The distributions
develop a power law tail approaching Tg at times much shorter
than τα with a power law that directly accounts for the excess
wings in the spectra of Fig. 1(c).

ticular, we find that the excess wing regime at t/τα  1


is dominated by the appearance of new clusters, whereas
the growth of existing clusters dominates at later times.
In Fig. 3 we report the distribution Π(τ ) of waiting times
for the appearance of new clusters. For T ≤ 0.07, we
cannot measure the entire distribution, which is thus de-
FIG. 4. (a) The liquid is modeled as a collection of traps with
termined up to an uninteresting prefactor.
energies E, distributed according to ρ(E). (b) Relaxation is
At the highest investigated temperature near Tmct , the thermally activated and affects the energy of the other traps
distribution Π(τ ) is already very broad, with clusters by a random amount, proportional to ∆. (c) Dynamic sus-
appearing as early as 10−4 τα . The distribution peaks ceptibility χ00 (ω) with (∆ > 0) and without (∆ = 0) dy-
near 0.1τα , and has a cutoff around 10τα , when dy- namic facilitation for α = 1.1 and T = 0.629. Dynamic facil-
namic facilitation thus dominates. As T decreases below itation compresses the low-frequency part of the underlying
spectrum.
the mode-coupling crossover, a power law tail emerges
at τ  τα . For T ≤ 0.07, the power law extends
over at least 6 decades, with a nearly constant exponent
Π(log10 τ ) ∼ τ 0.38 . The relaxation of localised clusters at spectra, we analyse the simplest version of such a model
early times is extremely broadly distributed, presumably and assume, in a mean-field spirit, that dynamic facili-
stemming from an equally broad distribution of activa- tation equally affects all traps. A more local version was
tion energies. Remarkably, plugging this distribution of designed in Refs. [37, 38], for different purposes.
waiting times in Eq. (1) directly produces a power law We consider N traps with energy levels E > 0 drawn
spectrum χ00 (ω) ∼ ω −0.38 for ωτα  1, in quantitative from a distribution ρ(E), and assume activated dynam-
agreement with Fig. 1(c). This demonstrates that the ics. The energy E of a trap is renewed after a Poisson-
high-frequency power law in χ00 (ω) stems from the relax- distributed timescale of mean hτ (E)i = eE/T . Since deep
ation of a sparse population of clusters characterised by traps take much longer to relax than shallow ones, the
a broad distribution of relaxation times. system is dynamically heterogeneous. Following Ref. [39],
α

This microscopic view of the wing alone does not ex- we use ρ(E) ∝ e−E , with α ∈ [1, 2] to smoothly inter-
plain why it appears in “excess” of the α-peak observed at polate between the much-studied Gaussian [18, 35, 37]
larger times when dynamic facilitation sets in. To explain and exponential [36] distributions. Dynamics at tem-
this point, we construct an empirical model based on our perature T leads to the equilibrium energy distribution
numerical observations. We first imagine that the liq- Peq (T, E) ∝ ρ(E)eE/T . Whenever a trap relaxes, the
uid can be decomposed into independent domains char- energy of all other traps is shifted byh a randomi amount
acterised by a local relaxation time, see Fig. 4(a). This uniformly distributed in the interval − √∆N , √∆N , using a
heterogeneous viewpoint is mathematically captured by Metropolis filter to leave the equilibrium distribution Peq
trap models [35, 36]. To introduce dynamic facilitation unchanged. This coupling between traps mimics dynamic
as the second key ingredient, we construct a facilitated facilitation [37]. We monitorP relaxation through the per-
trap model, assuming that a given local relaxation event sistence function p(t) = h N1 i pi (t)i, where pi (t) = 1 if
may now affect the state of the other traps, see Fig. 4(b). trap i has not relaxed at time t, 0 otherwise. We com-
To provide a qualitative, generic description of relaxation pute p(t) either analytically (∆ = 0), or by simulating
5

the facilitated model (∆ > 0) and use it in place of Fs (t) Contemporary Physics 41, 15 (2000).
to get χ00 (ω) from Eq. (1). [4] U. Schneider, R. Brand, P. Lunkenheimer, and A. Loidl,
The model is specified by two parameters (α, ∆), for Phys. Rev. Lett. 84, 5560 (2000).
[5] P. Lunkenheimer, R. Wehn, T. Riegger, and A. Loidl,
which equilibrium dynamics can be studied at any tem-
Journal of Non-Crystalline Solids 307-310, 336 (2002).
perature T . We have systematically investigated this pa- [6] P. K. Dixon, L. Wu, S. R. Nagel, B. D. Williams, and
rameter space, and find spectra with quantitative dif- J. P. Carini, Phys. Rev. Lett. 65, 1108 (1990).
ferences but generic features. In Fig. 4(c), we select [7] R. L. Leheny and S. R. Nagel, Journal of Non-Crystalline
(α = 1.1, ∆ = 0.05) at T = 0.629 for aesthetic reasons, as Solids 235-237, 278 (1998).
this produces a spectrum close to original experiments on [8] N. Menon, K. P. O’Brien, P. K. Dixon, L. Wu, S. R.
glycerol at Tg [8, 12]. Fitting the α-peak to the frequency Nagel, B. D. Williams, and J. P. Carini, Journal of Non-
Crystalline Solids 141, 61 (1992).
representation of a stretched exponential reveals an ex-
[9] N. Menon and S. R. Nagel, Phys. Rev. Lett. 74, 1230
cess wing at high frequencies. However, imposing ∆ = 0 (1995).
produces the blue spectrum which leaves high frequencies [10] L. Wu, Phys. Rev. B 43, 9906 (1991).
unaffected, but extends much further at low frequencies. [11] K. L. Ngai and M. Paluch, The Journal of Chemical
When ∆ = 0, each trap relaxes independently, and the Physics 120, 857 (2004).
equilibrium distribution Peq dictates the dynamic spec- [12] S. Adichtchev, T. Blochowicz, C. Tschirwitz, V. N.
trum, which is broad and relatively symmetric. In the Novikov, and E. A. Rössler, Phys. Rev. E 68, 011504
(2003).
presence of facilitation, ∆ > 0, shallow traps still relax
[13] T. Blochowicz, C. Tschirwitz, S. Benkhof, and
independently and are essentially unaffected. Crucially, E. Rössler, The Journal of chemical physics 118, 7544
deep traps now receive small kicks whenever a shallow (2003).
trap relaxes, and their energies slowly diffuse towards [14] C. Gainaru, R. Kahlau, E. A. Rössler, and R. Böhmer,
the most probable value, thus accelerating their relax- The Journal of chemical physics 131, 184510 (2009).
ation. As a result, the low-frequency part of the spectrum [15] Phys. Rev. Lett. 110, 107603 (2013).
appears more compressed in presence of facilitation, see [16] K. Duvvuri and R. Richert, The Journal of Chemical
Physics 118, 1356 (2003).
Fig. 4(c), as hinted in [40]. We thus interpret the winged,
[17] P. Lunkenheimer, R. Wehn, U. Schneider, and A. Loidl,
asymmetric spectrum as a broad underlying distribution Phys. Rev. Lett. 95, 055702 (2005).
of relaxation timescales (with a power law shape at early [18] G. Diezemann, The Journal of Chemical Physics 107,
times) compressed by dynamic facilitation at long times. 10112 (1997).
Ironically, in our picture, the α-peak appears in “excess” [19] G. Diezemann, U. Mohanty, and I. Oppenheim, Phys.
of a much broader underlying time distribution with a Rev. E 59, 2067 (1999).
high-frequency power law shape. The wing thus forms [20] M. Domschke, M. Marsilius, T. Blochowicz, and
T. Voigtmann, Phys. Rev. E 84, 031506 (2011).
an integral part of the structural relaxation.
[21] J. P. Sethna, J. D. Shore, and M. Huang, Phys. Rev. B
Our study frontally attacks a central question regard- 44, 4943 (1991).
ing the relaxation dynamics of supercooled liquids near [22] J. D. Stevenson and P. G. Wolynes, Nature physics 6, 62
the experimental glass transition and paves the way for (2010).
many more studies of a totally unexplored territory now [23] P. Viot, G. Tarjus, and D. Kivelson, The Journal of
made accessible to modern computer studies. Chemical Physics 112, 10368 (2000).
We thank G. Biroli and J. Kurchan for discussions, [24] R. V. Chamberlin, Phys. Rev. Lett. 82, 2520 (1999).
[25] J. C. Dyre and T. B. Schrøder, Rev. Mod. Phys. 72, 873
and S. Nagel for detailed explanations about experi-
(2000).
ments. Some simulations were performed at MESO@LR- [26] L. Berthier and J. P. Garrahan, The Journal of Physical
Platform at the University of Montpellier. This work Chemistry B 109, 3578 (2005).
was supported by a grant from the Simons Foundation [27] A. Ninarello, L. Berthier, and D. Coslovich, Phys. Rev.
(#454933, LB), the European Research Council under X 7, 021039 (2017).
the EU’s Horizon 2020 Program, Grant No. 740269 (CS), [28] L. Berthier, E. Flenner, C. J. Fullerton, C. Scalliet, and
a Herchel Smith Postdoctoral Research Fellowship (CS), M. Singh, Journal of Statistical Mechanics: Theory and
Experiment 2019, 064004 (2019).
Sidney Sussex College, Cambridge (Ramon Jenkins Re-
[29] E. Flenner and G. Szamel, Nature communications 6, 1
search Fellowship to CS) and Capital Fund Management (2015).
– Fondation pour la Recherche (BG). [30] R. Yamamoto and A. Onuki, Phys. Rev. E 58, 3515
(1998).
[31] L. Berthier and M. D. Ediger, The Journal of Chemical
Physics 153, 044501 (2020).
[32] D. Chandler and J. P. Garrahan, Annual Review of Phys-
[1] L. Berthier and G. Biroli, Reviews of modern physics 83, ical Chemistry 61, 191 (2010).
587 (2011). [33] A. S. Keys, L. O. Hedges, J. P. Garrahan, S. C. Glotzer,
[2] B. Schmidtke, N. Petzold, R. Kahlau, M. Hofmann, and and D. Chandler, Phys. Rev. X 1, 021013 (2011).
E. Rössler, Physical Review E 86, 041507 (2012). [34] M. Vogel and S. C. Glotzer, Phys. Rev. Lett. 92, 255901
[3] P. Lunkenheimer, U. Schneider, R. Brand, and A. Loid, (2004).
6

[35] J. C. Dyre, Phys. Rev. Lett. 58, 792 (1987). [39] V. I. Arkhipov and H. Baessler, The Journal of Physical
[36] J.-P. Bouchaud, Journal de Physique I 2, 1705 (1992). Chemistry 98, 662 (1994).
[37] C. Rehwald, O. Rubner, and A. Heuer, Phys. Rev. Lett. [40] X. Xia and P. G. Wolynes, Phys. Rev. Lett. 86, 5526
105, 117801 (2010). (2001).
[38] C. Rehwald and A. Heuer, Phys. Rev. E 86, 051504
(2012).
Supplementary Materials: Microscopic origin of excess wings in relaxation spectra of
deeply supercooled liquids
Benjamin Guiselin*,1 Camille Scalliet*,2 and Ludovic Berthier1, 3
1
Laboratoire Charles Coulomb (L2C), Université de Montpellier, CNRS, 34095 Montpellier, France
2
Department of Applied Mathematics and Theoretical Physics,
University of Cambridge, Wilberforce Road, Cambridge CB3 0WA, United Kingdom
3
Department of Chemistry, University of Cambridge,
Lensfield Road, Cambridge CB2 1EW, United Kingdom
(Dated: March 3, 2021)
arXiv:2103.01569v1 [cond-mat.soft] 2 Mar 2021

I. MODELS

1. Glass-forming computer models

We study a non-additive, continuously polydisperse mixture of spherical particles of equal mass m in two and three
dimensions (d = 2, 3). Two particles i and j, at a distance rij from one another interact via the repulsive potential
 12  12  2  4
σij σij rij rij
v(rij ) =  + vc (rij /σij ) =  + c0 + c2 + c4 , (S1)
rij rij σij σij
if their relative distance rij /σij is smaller than a cutoff value of x = 1.25. The constants c0 , c2 , c4 of the fourth-order
polynomial vc (r) are such that the potential v(r) and its first two derivatives are continuous at the cutoff value:
c0 = −28/x12 , c2 = 48/x14 , c4 = −21/x16 . The particles’ diameters σi are drawn from the normalised distribution

 A if σ ∈ [σ , σ
min max ]
P(σ) = σ 3 (S2)

0 otherwise
with A a normalisation constant and σmax /σmin p= 2.219. We use the average diameter σ as unit length,  as unit energy
(the Boltzmann constant is set to unity) and mσ 2 / as unit time. In these units, σmin = 0.73 and σmax = 1.62. We
employ a non-additive cross-diameter rule
σi + σj
σij = (1 − η|σi − σj |) , (S3)
2
with η = 0.2 in the system’s units. The choices of particle polydispersity and non-additivity make the supercooled
fluid robust against fractionation and crystallization at all temperatures numerically explorable, both in two and three
dimensions [S1]. We consider various system sizes: N = 1200, 10000 in 3d and N = 2000, 10000 in 2d. We simulate
this glass-forming model in a cubic box of linear size L using periodic boundary conditions. The number density of
particles ρ = N/Ld is set to unity without loss of generality.

2. Trap model

We consider traps with energy levels E > 0 drawn from the exponential power distribution
α α
ρ(E) = e−(E/E0 ) , (S4)
E0 Γ(1/α)
and we take E0 = 1 in the following. We assume that dynamics at temperature T is thermally activated. The
energy E of a trap is renewed after a Poisson-distributed timescale of mean hτ (E)i = eE/T . The equilibrium energy
distribution at temperature T is
Z ∞
ρ(E)eE/T
Peq (T, E) = where Z(T ) = dEρ(E)eE/T . (S5)
Z(T ) 0

We monitor relaxation in the trap model by computing the average persistence function p(t). In the absence of
dynamic facilitation, the persistence can be directly computed
Z ∞
p(t) = dEPeq (T, E)e−t/hτ (E)i . (S6)
0
2

II. METHODS - PARTICLE SIMULATIONS

1. Preparation of equilibrated configurations

The model glass-forming liquid is efficiently simulated at equilibrium with the swap Monte Carlo algorithm. We
employ the hybrid swap Monte Carlo/Molecular Dynamics algorithm implemented in the LAMMPS package (2d) or
homemade code (3d), with optimal parameters, as described in Ref. [S2]. The algorithm allows to effectively produce
a large number ns of independent equilibrated configurations down to the experimental glass transition temperature
(see Sec. II 5).

2. Molecular dynamics simulations

The equilibrium configurations generated by the swap algorithm are directly used as initial conditions for standard
molecular dynamics (MD) simulations [S3]. In 3d, we perform conventional MD, while we run MD at constant
temperature in 2d using the Nosé-Hoover implementation of the Nosé thermostat. In both cases, the integration time
step equals 0.01. The simulations are either run using a homemade MD code or using the LAMMPS package, the
latter allowing us to run multi-CPU simulations and perform extremely long runs for relatively large systems (e.g.
N = 10000).

3. Dynamic observables

We quantify the relaxation dynamics in two and three dimensions via several observables. In 3d, we measure the
self-intermediate scattering function
* N
+
1 X
Fs (t) = cos [q · δri (t)] , (S7)
N i=1

where δri (t) is the displacement of particle i over time t. The brackets indicate the ensemble average over ns
independent runs and an angular average over wavevectors with |q| = 6.9, which corresponds to the location of the
first peak in the total structure factor. The relaxation time τα is then defined by the condition Fs (τα ) = e−1 .
In 2d, collective long-ranged fluctuations give rise to a spurious contribution to the displacements of particles [S4, S5]
which affects the self-intermediate scattering function. The dynamics is better monitored by studying the evolution of
the local environment of particles, instead of their displacements. We define a bond-orientational correlation function
CΨ (t) [S6, S7]. We first introduce the six-fold bond-orientational order parameter of particle i
ni
1 X
Ψi (t) = ei6θij (t) , (S8)
ni j=1

where the sum is performed over the ni first neighbours of particle i at time t. These neighbours are defined as particles
j with rij < 1.45, which corresponds to the location of the first minimum in the total radial pair correlation function
g(r). We have checked that alternative definitions of first neighbours, e.g. via Voronoi tessellation or a solid-angle
based method [S8], lead to the same quantitative results. The angle θij (t) measures the angle between the x-axis and
the axis connecting the positions of particles i and j at time t. Since these correlations are rotationally invariant, the
choice of x-axis is made without loss of generality. The bond-orientational correlation function is defined as
P ∗
Ψi (t) [Ψi (0)]
CΨ (t) = iP
2
, (S9)
i |Ψi (0)|

where the sums run over the N particles in the system, the brackets denote an ensemble average over the ns in-
dependent simulations, and the conjugate complex is denoted with a star. We define relaxation times τα in the
two-dimensional model via the condition on the bond-orientational function: CΨ (τα ) = e−1 .

4. Mobility at the single-particle level

When analysing the mobility at the single-particle level, we first need a criterion to distinguish between mobile
and immobile particles. In 3d, we have considered several mobility definitions which all give quantitatively similar
3

results. The first mobility definition is based on displacements. To get rid of trivial motion due to thermal vibrations,
we quench the system at t = 0 and at time t into its closest potential energy minima (inherent structures). This
minimisation is done using a homemade code performing a conjugate-gradient method. Particle i is then said to be
mobile at time t if its inherent-state displacement δriIS (t), computed as the difference between its positions in the two
inherent states at times 0 and t [S9], has a magnitude larger than 0.8. This cutoff is between the first minimum and
the second maximum of the self part of the van Hove function
N
1 X
Gs (r, t) = δ(r − |riIS (t)|), (S10)
N i=1

in the time regime where the correlation function Fs (t) is almost constant.
A second mobility definition relies on the detection of changes in the particle’s environment. At time t = 0, we
compute the number, ni , and identity of particle i’s neighbours. Two particles i and j are considered as neighbours
at time 0 if rij /σij < 1.485, which corresponds to the first minimum in the total rescaled pair correlation function
g(rij /σij ). At time t > 0, we define

i ni (t|0)
CB (t) = (S11)
ni
as the fraction of remaining neighbours of particle i, with ni (t|0) the number of particles which were neighbours of
particle i at time 0 and still its neighbours at time t. To avoid a spurious signal that we attribute to particles losing
their farthest neighbours because of thermal vibrations, we define neighbours at time t with a slightly larger cutoff
than at time 0, namely rij /σij < 1.7. The particle i is finally considered as mobile at time t if CB i
(t) < 0.55. This
cutoff is chosen so that the two alternative definitions of mobility give sets of mobile particles with a significant overlap
and hence are consistent with each other.
We then introduce clusters of mobile particles. Two mobile particles at time t belong to the same cluster if their
relative distance rij is smaller than 1.5, which corresponds to the first minimum in the total radial pair correlation
function g(r).
In 2d, we measure a bond-breaking correlation CB i
(t) (see Fig. 2, main text) in the same spirit as in 3d. The cutoff
values employed to define neighbouring particles at time 0 and t > 0 are adapted to the specificities of g(rij /σij ) in
2d. In particular, the cutoff is rij /σij < 1.3 at time 0 and 1.7 at later times.

5. Determination of relevant temperature scales

We determine three temperature scales relevant to the dynamic slowdown. We perform three different fits to the
relaxation time data measured in MD simulations.
First, we fit the high-temperature data to an Arrhenius law. The temperature below which the data starts to
depart from the high-T Arrhenius behaviour locates the onset temperature of glassy dynamics noted To , where the
relaxation time is equal to τo .
The mode-coupling crossover temperature Tmct is obtained by fitting the first three decades of dynamic slowdown
0 ≤ log10 (τα /τo ) ≤ 3 with a power law τα (T ) ∝ (T − Tmct )−γ [S10]. The exponent γ is equal to 2.7 in 2d and 2.5 in 3d.
At the mode-coupling crossover temperature, the dynamics slows down by about 4 decades: log10 (τα /τo ) ≈ 4. The
temperature Tmct therefore marks the limit between the regime T > Tmct in which equilibrium simulations can be
performed with MD without resorting to the swap algorithm to prepare equilibrated samples, and the low-temperature
regime T < Tmct in which dynamics gets difficult to analyse numerically at equilibrium.
In order to compare to experimental results, we extrapolate the experimental glass transition temperature Tg ,
defined by log10 (τα (Tg )/τo ) = 12. Given the width of the simulation time window (max. 9 decades) and the value of
the relaxation time at the onset temperature, we can directly measure relaxation times in the regime log10 (τα /τo ) . 7.
This means that our direct measurements need to be extrapolated over a little more than 5 decades in order to locate
Tg . We perform a better extrapolation by first using a time-temperature superposition (TTS) [S11]. In the regime
where decorrelation to e−1 is reached, we observe that the second step of the relaxation in Fs (t) or CΨ (t) is well-fitted
β
by a stretched exponential of the form F0 e−(t/τ ) , with F0 , τ and β free fitting parameters. The stretching exponent
β ' 0.56 (0.6 for CΨ in 2d) turns out to be almost temperature-independent, and the amplitude F0 increases slightly
and monotonically with decreasing temperature. Fixing β and adapting F0 enables to use TTS in order to estimate the
relaxation time at temperatures where Fs (t) or CΨ (t) do not reach e−1 , yet decorrelates significantly. This allows us to
estimate precisely the relaxation time over 2 additional decades. Note that this relatively safe extrapolation procedure
can only be implemented here thanks to the swap Monte Carlo algorithm which provides us with equilibrated initial
conditions.
4

Finally, we extrapolate the relaxation time data over the 3-4 additional decades using an Arrhenius fit τα (T ) ∝ eEA /T
with an activation energy EA equal to 2.97 in 2d and 2.67 in 3d. This final extrapolation allows us to locate Tg . Since
data at shorter times are super-Arrhenius, we believe that an Arrhenius extrapolation is likely to slightly underestimate
the value of Tg , and the value we get is therefore a conservative one.
The values of the three relevant temperatures in 2d and 3d are reported in Table I.

Model To Tmct Tg
2d 0.2 0.12 0.07
3d 0.2 0.095 0.056

TABLE I. Relevant temperature scales in the two studied models.

6. Computation of relaxation spectra

The computation of relaxation spectra χ00 (ω) first requires to differentiate the correlation function with respect to
the natural logarithm of time. This is done using a first-order finite difference approximation. Namely, if during the
ns simulations, configurations are stored at times {tk }k=1...n logarithmically spaced, we have

dFs (tk ) Fs (tk ) − Fs (tk−1 )


= if k > 1. (S12)
d log t log(tk ) − log(tk−1 )

The integral in Eq. (1) of the main text can then be evaluated using the right-hand (rectangle) rule at any angular
frequency ω

n
X  
dFs (tk ) ωtk tk
χ00 (ω) = − log . (S13)
d log t 1 + (ωtk )2 tk−1
k=2

However, the error on the numerical integration gets larger at small frequencies, when the complete decorrelation of the
self-intermediate scattering function is not observed, as the integral becomes dominated by the long-time behaviour
of Fs (t). The error also increases at large frequencies due to the time discretisation. It is enough to compute χ00 (ω)
only at discrete angular frequencies ωk = 2π/tk for k = 1 . . . n to guarantee a small-enough error and a fine-enough
spacing in the frequency range.
Regarding the α-relaxation peak predicted from the stretched exponential fit of Fs (t) [dashed lines in Fig. 1(c),
main text], we use the previous procedure and we first compute Fs (t) on a sufficiently large time interval [10−40 , 1040 ]
to minimise the error in the computation of χ00 (ω).

7. Experiment/simulation timescale dictionary

We explain how we transform relaxation times τα /τo measured in simulations into their experimental counterparts,
expressed in seconds. Experimental measurements of the temperature evolution of τα in various supercooled liquids
show that at really high temperatures, the relaxation time is about 1 ps. Data measured in this high-temperature
regime follow an Arrhenius law extending over about 2 decades [S12, S13] before the growth of the α-relaxation time
with decreasing temperature becomes super-Arrhenius. Thus, the relaxation time in a large number of molecular
liquids studied in experiments is about τo ≈ 100 ps, i.e. 10−10 s. We use this value to set the dictionary between
the simulations and experiments, so that relaxation times can be expressed in units of τo for both simulations and
molecular experiments.
Consequently, when we extrapolate τα /τo = 108 in simulations (see Fig. 2 of the main text), this corresponds
to a physical time of 10 ms. We also note that as mentioned before, conventional MD simulations are stuck at
τα /τo ≤ 104 − 105 corresponding to a physical time of 1 µs and hence a poor degree of supercooling with respect to
experimental conditions.
5

III. METHODS - TRAP MODEL

1. Relaxation spectra in the absence of dynamic facilitation ∆ = 0

In the absence of dynamic facilitation, the average persistence Eq. (S6) is evaluated using Mathematica (NIntegrate,
working precision 30). We compute the persistence p(t) over a time interval large enough to observe full decorrelation,
[10−10 , 1070 ] for α = 1.1, T = 0.629, and minimise errors in the relaxation spectrum. We then calculate the relaxation
spectrum χ00 (ω) by following the procedure described in Sec. II 6, replacing Fs (t) by the obtained persistence p(t).

2. Simulations of the facilitated trap model ∆ > 0

We describe the numerical method employed to simulate the dynamics of the facilitated trap model. We consider
a system composed of N traps.
We initialise the simulation with an equilibrium condition by sampling the traps’ energies directly from the equi-
librium distribution Peq (T, E). Since the cumulative probability distribution of energies cannot be computed and
inverted explicitly, we use Mathematica to evaluate the cumulative distribution
Z E
Ceq (T, E) = Peq (T, E 0 )dE 0 (S14)
0

on a fine grid from Emin , where Ceq (T, Emin ) = 10 , to Emax , where Ceq (T, Emax ) = 1 − 10−10 . We use a cubic spline
−10

interpolate to numerically construct the reciprocal function E = Ceq −1 (T, E). For each of the N traps, we generate
a random number X from the uniform distribution in [0, 1], and assign the energy E = E(X) to it. By construction,
this procedure generates energies which are distributed according to the distribution Peq . We also assign a relaxation
time exponentially distributed to each trap, with mean eE/T . To do so, we generate another random number Y from
a uniform distribution in [0, 1] and attribute a relaxation time τi = −eE/T log(Y ) to the trap i. We initialise the
persistence pi (t = 0) of all traps to one.
The dynamics proceeds as follows. First, we identify the trap io with the smallest relaxation time τmin , which will
relax first. We update all other traps by subtracting τmin to their relaxation time τi . When the trap io relaxes, its
persistence is set to zero, pio = 0 and we give it a new energy value sampled from ρ(E), and a new relaxation time,
as described above. We sample the probability distribution in Eq. (S4) using the method described in Ref. [S14].
Namely, we generate two variables: X distributed uniformly in [0, 1], and y drawn from a Gamma distribution with
shape parameter 1 + 1/α and rate parameter 1; and we finally compute E = Xy 1/α .
This relaxation event then affects all other traps.
h We attempt
i to displace their energy by a random amount δE
(different for each trap) uniformly distributed in − N , N : E → E 0 = E + δE. The scaling with N ensures that
√∆ √∆

the resulting dynamics is independent on N . We then accept or reject this attempt in order to leave the equilibrium
probability distribution Peq unchanged. To this end, we introduce an effective potential V = −T log Peq , and compute
the change in effective potential δV = T (E 0α − E α ) − δE. We then use the Metropolis filter: if δV < 0, the change in
energy is accepted, otherwise, it is accepted with probability exp(−δV /T ). When accepted, we pick a new relaxation
0
time exponentially distributed with average eE /T . When the move is completed, we again determine which of the
traps is the next one to relax, and proceed as before.
We measure the average persistence p(t) as a function of time by summing over all traps
* N
+
1 X
p(t) = pi (t) , (S15)
N i=0

where the brackets indicate average over independent runs. We simulate the dynamics of the model until the total
persistence is equal to zero.

SUPPLEMENTARY REFERENCES

[S1] Andrea Ninarello, Ludovic Berthier, and Daniele Coslovich, “Models and algorithms for the next generation of glass
transition studies,” Physical Review X 7, 021039 (2017).
[S2] Ludovic Berthier, Elijah Flenner, Christopher J Fullerton, Camille Scalliet, and Murari Singh, “Efficient swap algorithms
for molecular dynamics simulations of equilibrium supercooled liquids,” Journal of Statistical Mechanics: Theory and
Experiment 2019, 064004 (2019).
6

[S3] Shahrazad MA Malek, Richard K Bowles, Ivan Saika-Voivod, Francesco Sciortino, and Peter H Poole, “"swarm relax-
ation": Equilibrating a large ensemble of computer simulations,” The European Physical Journal E 40, 1–11 (2017).
[S4] Skanda Vivek, Colm P Kelleher, Paul M Chaikin, and Eric R Weeks, “Long-wavelength fluctuations and the glass
transition in two dimensions and three dimensions,” Proceedings of the National Academy of Sciences 114, 1850–1855
(2017).
[S5] Bernd Illing, Sebastian Fritschi, Herbert Kaiser, Christian L Klix, Georg Maret, and Peter Keim, “Mermin–wagner
fluctuations in 2d amorphous solids,” Proceedings of the National Academy of Sciences 114, 1856–1861 (2017).
[S6] Elijah Flenner and Grzegorz Szamel, “Fundamental differences between glassy dynamics in two and three dimensions,”
Nature communications 6, 1–6 (2015).
[S7] Elijah Flenner and Grzegorz Szamel, “Viscoelastic shear stress relaxation in two-dimensional glass-forming liquids,” Pro-
ceedings of the National Academy of Sciences 116, 2015–2020 (2019).
[S8] Jacobus A van Meel, Laura Filion, Chantal Valeriani, and Daan Frenkel, “A parameter-free, solid-angle based, nearest-
neighbor algorithm,” The Journal of chemical physics 136, 234107 (2012).
[S9] Thomas B Schrøder, Srikanth Sastry, Jeppe C Dyre, and Sharon C Glotzer, “Crossover to potential energy landscape
dominated dynamics in a model glass-forming liquid,” The Journal of Chemical Physics 112, 9834–9840 (2000).
[S10] Wolfgang Götze, Complex dynamics of glass-forming liquids: A mode-coupling theory, Vol. 143 (OUP Oxford, 2008).
[S11] Ludovic Berthier and Mark D Ediger, “How to “measure” a structural relaxation time that is too long to be measured?”
The Journal of Chemical Physics 153, 044501 (2020).
[S12] B Schmidtke, N Petzold, R Kahlau, M Hofmann, and EA Rössler, “From boiling point to glass transition temperature:
Transport coefficients in molecular liquids follow three-parameter scaling,” Physical Review E 86, 041507 (2012).
[S13] B Schmidtke, N Petzold, R Kahlau, and EA Rössler, “Reorientational dynamics in molecular liquids as revealed by
dynamic light scattering: From boiling point to glass transition temperature,” The Journal of chemical physics 139,
084504 (2013).
[S14] George Marsaglia and Wai Wan Tsang, “A simple method for generating gamma variables,” ACM Trans. Math. Softw.
26, 363–372 (2000).

You might also like