Metastable states in the East model and plaquette spin models Robert Jack (University of Bath) Warwick meeting, Jun 2014 Thanks to:! Juan Garrahan (Nottingham) Ludovic Berthier (Montpellier) Outline − Part I : Metastable states in the East model! − Part II : Plaquette models, Interpreting point-to-set measurements and overlap fluctuations, using metastable state arguments vaporization near Tg. This means that OTP’s landscape is very heterogeneous. The basins sampled at high temperature allow relaxation by surmounting low barriers involving the rearrangement of a small number of molecules. The very large activation energy at T ≈ Tg, on the other hand, corresponds to the cooperative rearrangement of many molecules. These differences between strong and fragile behaviour imply a corresponding topographic distinction between the two Metastable states -7.05 0.0 Figure 6 Mean inherent struct sized Lennard–Jones atoms, a from which the inherent structu dynamics simulations at consta temperatures for 256 Lennard– are of type B. The Lennard–Jon $BB"0.88, $AB"0.8, and %AA temperature, energy and time a $AA(m/%AA )1/2, respectively, with were performed at a density of and 3.33&10#6. When T > 1 entire energy landscape, and th are shallow. Under these condi activation energy for structural and T ≈ 0.45, the activation en ‘landscape-influenced’, and th strongly temperature-depende separating sampled adjacent e (calculations not shown). This i rare jumps over distan and execute Stillinger, 2001] crossover between landscapecorresponds closely with the m from refs 70 and 72.) “Basins” / “minima” / “valleys” on a free energy landscape Transition states Potential energy Basin Ideal glass Coordinates Crystal [Debenedetti Figure 5 Schematic illustration of an energy landscape. The x-axis represents all configurational coordinates. (Adapted from ref. 44.) 0. Regions of configuration space with fast “intra-state” motion and 264slow transitions between states © 2001 Macmillan Magazines Ltd Counting states [Kurchan and Levine, J Phys A, 2011] Regions of configuration space with fast “intra-state” motion and slow transitions between states States with lifetime ⌧ , take a time tobs with 1 ⌧ tobs ⌧ ⌧ Averaging over a time period tobs : same as Boltzmann averaging within the state Diversity (entropy) of averaged properties: … gives number (and properties) of metastable states! … including configurational entropy (complexity) 2 II. East model MODEL, METHODS, AND METASTABLE STATES A. (a) 1 0.8 Model Spins ni = 0, 1, with i = 1 . . . N . The East model [17, 18] consists of L binary spins ni = 0, 1, with i = 1 . . . L, and periodic boundaries. i with 1 We refer to spins with ni = 1 as ‘up’ and those ni = 0 as ‘down’. The notation C = (n1 , n2 , . . . , nL ) indicates a configuration of the system. The key feature of the model is that spin i may flip only if spin i 1 is up. If this kinetic constraint [19, 20] is satisfied, spin i E flips from 0 to 1 with rate c and from 1 to 0 with rate (a) 1 c. We take1c/(1 c) = e where is the inverse temperature, so small c corresponds to low temperature. The rates for0.8 spin flips obey detailed balance, so that the probability ofPconfiguration C at equilibrium is sim0.6 0 i ni /Z, where Z = (1 + e pins ply p (C)C(t) = e )L is the =5 ies. partition function. one has hni i = c, and =4 0.4 At equilibrium, with 3 the regime of interest is= small c (low temperature). 2 of p0 (C), the kinetic conin- Despite the0.2trivial =form = 1 to complex co-operative dyure straint in this model leads 0 -2c. In 0particular, 1 is 2 4 6 8 time 10 namics at small the relaxation 10 10 10 10 10 10 10 in of i the model diverges in a super-Arrhenius fashion, as t 2 /(2 ln 2) rate ⌧0 ⇠ e (b) [18, 33]. The rapid increase in relaxation 2 = 0 1a where erse time is illustrated in bFig. we show b = 1 hn (t)n (0)i hn i i i 1 ure. b=2 i 2 i hn 2i2 hni (0)i hat hni (t)n hni ii b = 3 i C(t) = , (1) b=0 imC(t) c(1 c) the 0.95 and Spin i can flip only if n Flip rates c and 1 C(t) = C(t) 0.6 = 1. =5 =4 =3 =2 =1 0.4 0.2 c gives hn 2 i i =0 c = -2 10 (b) 10 1 01+e2 10 4 10 t b=0 1 . b=1 10 10 8 10 10 b=2 b=3 b=0 C(t) 6 0.95 b=1 =5 =4 0.9 -2 10 10 0 b=3 b=2 2 10 b=4 4 10 t 10 6 10 8 10 FIG. 1: Correlation functiontimes C(t) in the t East model. Characteristic c b ⇡ 10 1 (b ) 2 (a) The inverse temperature is varied from = 1 to = 5, showing a dramatic increase in relaxation time. (b) At the lowest [RLJ, Phys E, 2013] temperatures, plateaus are apparent during Rev the early stages B. Metastable states 0.1 ↵ ciated with a profile C , holds very East model. As discussed by Sollich 0.01 metastable states, we perform a time averalso [34]), motion on a time scale t 0.001 0 50 pins, defining stricted to domains of size100 2b 1 , with [RLJ, Phys Rev E, 2013] b=1 1 immediately to the right of a long-liv 3 Z t 1 1 b b (b ) within suc 0 0 c ⌧ t ⌧ c then spins 0.1 2 n = (1/t) dt n (t ), (2) Characteristic times t ⇡ c i b [ it =5] tories of the East model, starting from a particular initial 1/2 2 hnit i0ic h nit iic flip many times within timeprocesses t, while condition that helps to reveal the physical at 0.01 and l work.likely Fig. 2to shows spins hnof flipthe attime-averaged all. The result this it iic Initial condition: 11.. 1.1. 1..1 1...1 1....1 eraged spin1...profile their variances h( nit )2 iic , where the subscript “ic” indiisthat that nit !was hntaken c if aspin i is with i i = with b=0 1 cates0.001 the average fixed initial con0 50 100 Sollich and [18] used the term dition (in contrast to Evans other averages in this work which b = 2 1 C t = 0.1 (n1t , n2t , . . . , nLt ). (3) are conventional averages over equilibrium trajectories). describe these mobile domains. A m We emphasise the following threeb points: 0.1 lifetime of order c can be identifie 0.01 rvation of [14] is that if the system has • All of the hnit iic are either close to 1 or of order 1/2 position superdomains. 0.01 ates indexed by ↵ = 1, 2, . . . , then each c ⌧ 1. Also,oftheits standard deviation h( nit )2 iic ⌧ 0.001 ↵50 1 0 1 In for all spins. ofThus, for thisthe initial condition ated with a profile C . Further, if100t is much terms Fig. 1b, limit c and for each of these times t, one almost certainly 0.001 b=1 1 0 50 100within a he intrastate relaxation times, but small sponds to choosing a finds a profile C t that is close totime a reference profile ↵ b = 3C 1 = hn i . Each reference profile (one for each it icfunction. t 0.1 heir lifetimes, then the observed profile C t relation However, when ↵ value of t shown) can therefore be identified with a rely be close to one of the C . Hence, the sults to state gainthat information about 0.1 metastable is stable on a time scale t.me Time averaging… 0.01 0.001 0 =2 1 [ b= 5] 0.1 Initial condition: 50 hnit iic 100 1/2 h n2it iic •0.01 For b 1, any spins with hnit iic ⇡ 1 are separated by at least 2b 1 spins with hnit i ⌧ 1. If two up spins are closer than this in the initial condition 0.001 tories of the starting from then rightmost themmodel, is within the mobile 0 the 50ofEast 100 domaincondition associated with leftmost one. Inthe thatp thati the helps to reveal case, the rightmost (nj ) tends to flip many work. Fig. spin 2 shows the time-averag times and njt converges to a value2 close to c. Time-averages (almost) determined bytheir initial condition… 1... 11.. 1.1. 1..1 1...1 1....1 variances h( nit ) iic , where the 0.01 B. Metastable states ↵ 0.1 ciated with a profile C , holds very East model. As discussed by Sollich 0.01 metastable states, we perform a time averalso [34]), motion on a time scale t 0.001 0 50 pins, defining stricted to domains of size100 2b 1 , with [RLJ, Phys Rev E, 2013] b=1 1 immediately to the right of a long-liv Z t 1 1 b b (b ) within suc 0 0 c ⌧ t ⌧ c then spins 0.1 2 nit = (1/t) dt ni (t ), (2) Characteristic times tb ⇡ c flip many times within time t, while 0 0.01 likely to flip at all. The result of this l 1 eraged ñi =spin ⇥(nprofile i is that nit ! hni i = c if spin i is with 2 ) = 0, 1 0.001 0 50 100 Sollich and Evans [18] used the term b = 2 1 C t = (n1t , n2t , . . . , nLt ). (3) describe these mobile domains. A m Measure entropy of ñi b 0.1 lifetime of order c can be identifie rvation of [14] is that if the system has byindexed pattern-counting ` position of its superdomains. 0.01 ates by ↵ = 1, 2, .at . . ,length then each ↵ ated with a profile C . Further, if t is much In terms of Fig. 1b, the limit c1 0.001 P 0 50 100within a he intrastate relaxation times, but small sponds to choosing a time S` = B` p(B` ) ln(1/p(B` )) b = 3 1 heir lifetimes, then the observed profile C t relation function. However, when ↵ rely be close to one of the C . Hence, the sults to gain information about me 0.1 Counting… B` : state of (ñi , . . . , ñi+` 1) 0.01 0.001 0 50 100 i E.g., B`=4 = (1, 0, 1, 0) almost never occurs for b 2 Counting… [RLJ, Phys Rev E, 2013] Convenient to plot S` S` 1 ... converges for large ` to entropy per site [ (a) 0.09 b=0 b=1 b=2 b=3 b=4 theory Sl - Sl-1 0.08 0 =4] 5 10 15 20 (a) ⇤ b Length scale ` ⇡ 2 g (r) 1 t 0.005 For b = 0, expect ñ = n, ‘static’ result Well-defined complexity . . . for states of00 lifetime 2 tb l Theory: up spins are never closer than `⇤(b), 0.03 [ =5] (b) otherwise randomly distributed. . . G4(r,t) b=0 . . 0.04 . “hard rods” 0.02 b=1 4 r Hard rod states [RLJ, Phys Rev E, 2013] 1... 11.. 1x.. 111. 1.1. 1x1x.. 1cx. 1.11. 1x1x1x 1cxx1c 1ccxx 1 : ni ⇡ 1 ⇡0 FIG.0 :3:ni Schematic 1.1.1. b=0 b=1 b=2 T patc a p (ñit eval b=3 c : ni ⇡ c ± O(c ) 1/2 x : n ⇡ c ± O(c illustration of i time evolution) among 3/2 metastable states. Each box represents a state, labelled by its characteristic local density profile, and a value of b that Evolution to “longer rods”, some b deterministic and indicates its lifetime (of order c ). To describe the profile some steps… we stochastic use a notation suggested by Fig. 2, based on the relative 2 1/2 sizes of hnit iic and h nit iic . We distinguish four cases: (i) a whe C˜t , a in C˜ patc A one profi Part I summary Time scale separation means metastable states are welldefined…! Kurchan-Levine method allows counting of metastable states (of given lifetime)…! [Kurchan and Levine, J Phys A, 2011] In the East model, these states correspond with those of “hard rod” systems… ... at least for times tobs ⌧ ⌧↵ (that is, `⇤ ⌧ 1/c) … diverging length scale coupled with diverging time scale [RLJ, Phys Rev E, 2013] Triangular Plaquette Model J. Chem. Phys. 123, 164508 !2005" [Newman and Moore 1999, Garrahan and Newman 2000, Garrahan 2002] barriers only in Ising spins si = ±1 ty of its on vertices of triangular lattice Energy 1 2 X N 2 X s1 s2 s3 = nµ ven in a E = µ square and rhomboid droplets !L 4 showing relation between FIG. 9. Sketch caging 1 #11=is0, marked by a filled = 6#. The spins are on the intersections of grids; µ labels “plaquettes”, n = (1 s s s ) 1 µ 1 2 3 ed than 2 circle. The & symbol marks the plaquette interaction q24 = #24#25#34. Rotaof Ref. tions of 120° leave both Hamiltonian and triangular lattices invariant. HowAt equilibrium, n are uncorrelated (as in KCMs) µ ets with ever, note that the boundaries of the droplets are not all equivalent, since the rhombus shape is not invariant under this rotation. pin-spin Glauber dynamics (single spin flips). . . entropy, 2 c/T ir bulk Relaxation time ⌧ ⇠ e , as in East model H = − !1/2# & #xy#x+1,y#x,y+1 , !31# TPM: spin correlations J. Chem. Phys. 123, 164508 !2005" [Newman and Moore 1999, Garrahan and Newman 2000, Garrahan 2002] barriers only in Ising spins si = ±1 ty of its on vertices of triangular lattice hsi i = 0, hsi sj i = 0 ven in a Selected three-point and higher correlations are non-zero, FIG. 9. Sketch showing relation between square and rhomboid droplets !L caging log 2/ log 3 Correlation length ⇠ ⇠ c = 6#. The spins are on the intersections of grids; #11 is marked by a filled ed than circle. The & symbol marks the plaquette interaction q24 = #24#25#34. Rotaof Ref. tions of 120° leave both Hamiltonian and triangular lattices invariant. Howets with note that the states. boundaries. .of the droplets are not all equivalent, since the Manyever, metastable rhombus shape is not invariant under this rotation. pin-spin Assuming complexity ⇡ simple entropy, entropy, S` ⇡ `2 c ln c + 2` ln 2 [Cammarota and Biroli, EPL, 2012] ir bulk H = − !1/2# & #xy#x+1,y#x,y+1 , !31# approached at the Kauzmann temperature. DOI: 10.1103/PhysRevLett.86.5526 The striking universality of relaxation dynamics in supercooled liquids has remained intriguing for decades. In addition to the overall dramatic slowing of transport as the glass transition is approached, one finds the emergence of a strongly nonexponential approach to equilibrium when a supercooled liquid is perturbed [1]. This contrasts with the behavior of chemically simple liquids at higher temperatures, where only a single time scale for a highly exponential structural relaxation is usually encountered for times beyond the vibrational time scales. Both the range of time scales and the typical magnitude of the relaxation time require explanation. Several theoretical threads lead to the notion that the universal behavior of supercooled liquids arises from proximity to an underlying random first order transition [2–6] which is found in mean field theories of spin glass without reflection symmetry [7–9], and in mode coupling [10,11] and density functional [12–14] approaches to the structural glass transition. This picture explains both the breakdown of simple collisional theories of transport that apply to high temperature liquids at a characteristic temperature TA and the impending entropy crisis of supercooled liquids first discovered by Simon and brilliantly emphasized by Kauzmann [15] at the temperature TK . Furthermore the scenario suggests that the finite range of the underlying forces modifies mean field behavior between TA and TK in a way that leads to an intricate “mosaic” structure of a glassy fluid in which mesoscopic local regions are each in an aperiodic minimum but are separated by more mobile domain walls which are strained and quitepfar from local minima structures, as shown in Fig. 1 [16]. Relaxation of the elements of the mosaic, reconfiguring to other low energy structures, leads to the [RLJ and slow relaxation, and a scaling treatment of theGarrahan, most proba- PACS numbers: 64.70.Pf Relaxation in supercooled liquids is well approximated by the stretched exponential or Kohlrausch-Williams-Watts b (KWW) formula f!t" ! e2!t#tKWW " . A study conducted by Böhmer et al. [18] on over sixty glass formers at about Tg shows b and D are strongly correlated. The smallest b is found for the most fragile liquids. The heterogeneity of time scales suggests possible heterogeneity in space. The existence of spatial heterogeneity received support from recent computer simulations [19–21], although these are at temperatures near TA where the clusters are small, KCM or mosaic? Can think of TPM as a kinetically-constrained model … or as a “mosaic” made up of metastable states BUT: No domain walls, No surface tension between states “Nucleation” dynamics among states: barrier scales with log ` (not ` as in CNT) FIG. 1. An illustration of the “mosaic structure” of supercooled liquids. The mosaic pieces are not necessarily the same size due to the fluctuations in the driving force, configurational [Xia andfrom Wolynes, PRLconfigura2001] entropy. The system escapes a local metastable tion by an activated process equivalent to forming a liquidlike droplet inside a mosaic element. For droplets with size much smaller2005] than the mosaic (as shown with the small circle), the JCP, droplet shape and its surface energy cost are well described by “Cavity” point-to-set To measure “mosaic” tile size… Immobilise system outside a “cavity” of size ` What is the smallest ` for which the cavity has more than 1 state available? [Bouchaud and Biroli, J Chem Phys, 2004] For the TPM: this size is ⇠mos ⇠ c log 2/ log 3 [RLJ and Garrahan, J Chem Phys, 2005] Relaxation on length scale ⇠mos determines bulk relaxation time (consistent with rigorous bound) [Montanari and Semerjian, J Stat Phys, 2006] fi = 1 (a) f = 0.12 pi = 1 pi = −1 (b) f = 0.18 (c) f = 0.25 Random pinning Random pinning: from an equilibrium configuration C0 , ‘pin’ (immobilise) each spin w/prob f ODELSAllow WITH .model .. to relax to configuration C, PHYSICAL REVIEW E 85, 021120 (20 measure overlap hq(C, C0 )i = hsi s0i i 0 Also, measure p = hs s i i fiii= C1 β=3 (a) f = 0.12 (e) (d) 104 105 pi = 1 f = 0.10 (d) Finite length scale ⇠ ⇠ (1/c) (b) f = 0.18 f = 0.15 (e) pi = −1 (f) (c) f = 0.25 f(f)= 0.20 [RLJ and Berthier, PRE 2012] IN MODELS WITH . . . PHYSICAL REVIEW E 85, 021120 (2012) TPM “mosaic” fi = 1 β=3 (a) f = 0.12 (e) (d) 103 104 105 pi = 1 f = 0.10 (d) (b) f = 0.18 f = 0.15 (e) pi = −1 (f) (c) f = 0.25 f(f)= 0.20 [RLJ and Berthier, PRE 2012] FIG. 3. scale (Color⇠ online) of the propensities pi in the Finite length ⇠ (1/c)Snapshots (from numerics) long-time limit. (a)–(c) SPM at β = 3, varying f . At this temperature log 2/ log 3 Natural ∗ mosaic length scale ⇠mos ⇠ (1/c) ∗ 3 for which f f ≈ 0.18. (d)–(f) TPM at fβ ==0.10 f = 0.15≈ 0.15. Pinned f = 0.20 spins are blackidea: while pinning unpinnedreduces spins are number color coded dark blue (dark Mean-field of with available states, FIG.pale 3. (Color online) Snapshots ofgray) the propensities p0 the i in ≈ 1, and blue or green (light for p ≈ and gray) for p i i with fixed surface long-time tension: diverging length scale (at fixed c) limit. (a)–(c) SPM at β = 3, varying f . At this temperature On increasing f , the spins become polarized pi ≈ −1, respectively. [Cammarota and Biroli, EPL 2012] ∗ ∗ f ≈ 0.18. (d)–(f) TPM at β = 3 for which f ≈ 0.15. Pinned spins memost The correlations associated with the?? so that pi ≈ 1 inarespatial black cases. while unpinned spins are color coded with dark blue (dark 3 Coupling 4 5 between domains too weak for this, in TPM 10 10 10 ∗ pale blue or green (light gray) for p ≈ 0 and susgray) for p i ≈ 1,fand i , consistent with χ being maximal. propensity are strongest near Overlap fluctuations For fixed (quenched) C0 , bias C to lie near C0 , ✏N q(C,C0 ) with biasing probability e Allow model to relax to configuration C, measure overlap hq(C, C0 )i = hsi s0i i RANDOM PINNING IN GLASSY SPIN MODELS WITH . . . random pinning… average overlap quenched biased overlap… 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 (a) 1 β=3 0.8 eps=0 0.01 0.05 0.07 0.09 0.11 0.13 0.15 1e+00 1e+01 C(t) 0.6 0.4 f = 0.00 0.2 1e+02 1e+03 1e+04 t [RLJ and Garrahan, unpublished] 1e+05 0 100 14 0.05 0.10 0.15 0.20 101 102 t 103 104 105 [RLJ and Berthier, PRE 2012] Overlap fluctuations For two configurations C, C0 , bias C to lie near C0 , with biasing probability e✏N q(C,C0 ) (annealed case) 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 annealed biased overlap (annealed)… quenched biased overlap (quenched) average overlap average overlap First order phase transition at non-zero ✏⇤ eps=0 0.01 0.05 0.07 0.08 0.09 0.10 0.11 1e-01 1e+00 1e+01 1e+02 1e+03 1e+04 1e+05 1e+06 t 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 eps=0 0.01 0.05 0.07 0.09 0.11 0.13 0.15 1e+00 1e+01 1e+02 1e+03 1e+04 1e+05 t [Garrahan PRE 2014, RLJ and Garrahan, unpublished] Overlap fluctuations [RLJ and Garrahan, unpublished] Why are the quenched/annealed cases so different? 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 annealed biased overlap (annealed)… quenched biased overlap (quenched) average overlap average overlap Defect density hnµ i ⇡ c2 eps=0 0.01 0.05 0.07 0.08 0.09 0.10 0.11 1e-01 1e+00 1e+01 1e+02 1e+03 1e+04 1e+05 1e+06 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 eps=0 0.01 0.05 0.07 0.09 0.11 0.13 0.15 1e+00 1e+01 1e+02 1e+03 1e+04 1e+05 t t Defect density hnµ i ⇡ c Annealed case: system is sampling non-typical metastable states Dimensionality? [RLJ and Garrahan, unpublished] Are two-dimensional models special? If surface tension between metastable states exists, can relate pinned systems to random-field Ising magnets… (hence no phase transitions in 2d) In TPM, no surface tension, so this argument does not apply… preliminary results for 3d generalisation of TPM … consistent with 1st order transition with annealed coupling, no transition for quenched coupling, (as in 2d) Dynamics and metastability Alternate probe of non-typical metastable states:! bias trajectories of a system using time-integrated quantities Let C be a “configuration”. Consider trajectories C(t) of length tobs (sample paths) Bias probabilities as: Prob[C(t); s] = Prob[C(t); 0] e stobs k[C(t)] Z(s) k[C(t)]: time-averaged measurement along trajectory [ Ruelle, Gallavotti-Cohen, Evans, Derrida, Lebowitz, Gaspard, Maes, etc ] Dynamics and metastability Bias probabilities as: Prob[C(t); s] = Prob[C(t); 0] e stobs k[C(t)] Z(s) Effect of field s is to enhance probabilities of long-lived states with small k This can lead to “dynamical phase transitions”: singular responses to field s [ Garrahan et al 2007, Hedges et al 2009, … ] System arrives in non-typical metastable states. . . (cf annealed overlap calculation) Summary KCMs (and the TPM) have well-defined metastable states (with a range of lifetimes…)! Random pinning and quenched overlap calculations reveal correlations among “mosaic tiles” but (apparently) no phase transitions…! No evidence for surface tension between states in TPM, [Relevant for “nucleation” arguments about dynamics, could be a scenario for atomistic systems(?)]! Annealed overlap and s-ensemble calculations probe nontypical metastable states and do lead to phase transitions in KCMs, TPM, …