betaFeedback

WorksPhysics and Astronomy

Rapid microscopic decorrelation precedes sustained collective ordering in quenched two-dimensional colloids

Abstract

Independent Brownian noise can rapidly separate trajectories launched from the same microscopic configuration, but configuration-space distance need not reveal their later assembly outcomes. We studied this distinction in 480 two-dimensional Brownian-dynamics trajectories generated as three independent-noise replicas of 40 parent configurations under four finite-duration quenches. Ninety-two of 160 replica sets sharing one parent and one protocol reached both ordered and non-ordered terminal outcomes. A relative-diffusion calculation relates the free-particle mean-squared replica separation to the time-integrated diffusivity. The measured mean-squared separation remained about one-sixth of the free-particle reference, while its scaling was indistinguishable for replica pairs with the same and different outcomes. In contrast, outcome-resolved global sixfold order, ordered-particle population and coherence, and reciprocal-space translational order separated progressively after quench completion. With quench durations weighted equally, global order, the state score combining global and population order, coherence, and translational order remained informative 0.50 reduced time units before terminal assessment. This result persisted after either changing the terminal window or reducing the integration timestep. Transient threshold crossings were common, and 94.5% of non-ordered crossings subsequently lost the required ordered-particle population. An independent N=128N=128 ensemble showed state-score discrimination at temperatures for which both terminal classes occurred. The assembly dynamics therefore separate generic microscopic decorrelation from the later emergence and finite-window persistence of collective order associated with terminal structure.

main.pdf

Introduction

Nonequilibrium colloidal assembly is shaped jointly by particle interactions, thermal fluctuations, and the history of the driving protocol [1, 2, 3, 4]. Even replicas prepared in the same microscopic configuration can reach different structures because their thermal-noise histories branch. This stochasticity motivates a dynamical question: as independent noise removes microscopic similarity, which components of the resulting motion are relevant to the later assembly outcome [5, 6, 7]?

An isoconfigurational ensemble, formed by launching independent-noise replicas from a shared parent, separates the common initial structure from subsequent stochastic forcing [8]. It also makes two complementary comparisons possible. At the pair level, one can ask whether the magnitude of coordinate or neighbor-network separation distinguishes replicas that later share an outcome from replicas that reach different outcomes. At the trajectory level, one can instead ask when interpretable collective coordinates become associated with the later outcomes. The first comparison probes generic decorrelation; the second probes motion along candidate reaction coordinates [9, 10]. Their relation is nontrivial because a high-dimensional distance accumulates displacements in many directions, whereas assembly depends on a restricted set of collective modes.

Two-dimensional colloids offer experimentally accessible collective coordinates for this purpose. Sixfold order can be separated into global alignment, the population of locally ordered particles, and coherence or connectivity within ordered regions [11, 12, 13, 14]. Bond-orientational order is a well-established crystallization coordinate [15, 16], although local magnitude alone does not capture the full structure [17]. We therefore compare orientational measures with a reciprocal-space translational-order score derived from the first reciprocal shell. Their joint evolution tests whether discrimination between terminal outcomes reflects collective structural ordering beyond the thresholds used to define the terminal state.

Here we analyze 480 Brownian-dynamics trajectories at N=64N = 64: three independent-noise replicas of 40 parents under each of four quench durations. A relative-diffusion reference derived from the Langevin equation sets the diffusive scale for replica separation, while a finite-time outcome-probability decomposition separates future-noise branching from parent-to-parent heterogeneity. Shared parents frequently reach both terminal outcomes, and coordinate and neighbor-graph distances grow without distinguishing the two pair types. Outcome-resolved orientational and translational coordinates instead separate progressively after quench completion. Transient crossings further distinguish short-lived alignment from sustained ordered-particle growth. The resulting picture is a separation between generic noise-driven decorrelation and the slower development and persistence of outcome-associated collective order.

Model and methods

Dynamics, parent ensemble, and quench protocol

We simulated particles in a two-dimensional square periodic box at number density ρ=0.82\rho= 0.82. The primary ensemble used N=64N = 64 and box length L=(N/ρ)1/2=8.8345L = (N/\rho)^{1/2} = 8.8345. Particles interacted through a truncated-and-shifted Lennard–Jones potential,

U(r)={ULJ(r)−ULJ(rc),r<rc,0,r≥rc,(1)U(r) = \begin{cases} U_{\mathrm{LJ}}(r) - U_{\mathrm{LJ}}(r_c), & r < r_c, \\ 0, & r \ge r_c, \end{cases} \tag*{(1)}

with

ULJ(r)=4ε[(σr)12−(σr)6].(2)U_{\mathrm{LJ}}(r) = 4\varepsilon\left[\left(\frac{\sigma}{r}\right)^{12} - \left(\frac{\sigma}{r}\right)^6\right]. \tag*{(2)}

The cutoff was rc=2.5σr_c = 2.5\sigma. Reduced units were used, with σ=ε=μ=kB=1\sigma= \varepsilon= \mu= k_{\mathrm{B}} = 1, where μ\mu is the mobility. The reduced time unit is t0=σ2/(με)t_0 = \sigma^2/(\mu\varepsilon). At Tlow=0.50T_{\mathrm{low}} = 0.50, the free-particle diffusion coefficient is D=μkBT=0.5D = \mu k_{\mathrm{B}}T = 0.5 and the Brownian diffusion time τB=σ2/D=2t0\tau_{\mathrm{B}} = \sigma^2/D = 2t_0. Periodic forces used minimum-image separations. For the primary box, L/2=4.417>rcL/2 = 4.417 > r_c, and no long-range tail correction was applied.

Large-system equilibrium simulations of the same range-limited rc=2.5σr_c = 2.5\sigma interaction place T=0.50T = 0.50, ρ=0.82\rho= 0.82 close to the reported low-temperature solid–fluid coexistence region and show that the phase diagram depends quantitatively on the cutoff [18]. This phase diagram provides equilibrium context for the present finite quenches. Because the N=64N = 64 square periodic cell is incommensurate with a defect-free triangular tiling, “ordered” and “non-ordered” below denote finite-window structural outcomes rather than equilibrium phases.

Following standard Brownian-dynamics constructions [19, 20], we integrated the overdamped Langevin equation by Euler–Maruyama,

ri(t+Δt)=ri(t)+μFi(t)Δt+[2μT(t)Δt]1/2ξi(t),(3)\mathbf{r}_i(t+\Delta t) = \mathbf{r}_i(t) + \mu\mathbf{F}_i(t)\Delta t + [2\mu T(t)\Delta t]^{1/2}\boldsymbol{\xi}_i(t), \tag*{(3)}

where ri\mathbf{r}_i and Fi\mathbf{F}_i are two-dimensional vectors and the Cartesian components of ξi\boldsymbol{\xi}_i are independent standard normal variates. The scalar mobility corresponds to free-draining Brownian dynamics; hydrodynamic interactions were omitted. The primary ensemble used Δt=5×10−5\Delta t = 5 \times10^{-5}. Numerical implementation and trajectory-storage details are given in Supplementary Information.

The Langevin equation also supplies a reference scale for the separation of two independent-noise replicas. For replicas aa and bb, let δri=ri(a)−ri(b)\delta r_i = r_i^{(a)} - r_i^{(b)} and D(t)=μkBT(t)D(t) = \mu k_{\mathrm{B}}T(t). Their relative motion obeys

dδri=μ(Fi(a)−Fi(b))dt+4D(t) dWi.(4)\mathrm{d}\delta r_i = \mu\left(\mathbf{F}_i^{(a)}-\mathbf{F}_i^{(b)}\right)\mathrm{d}t+\sqrt{4D(t)}\,\mathrm{d}\mathbf{W}_i. \tag*{(4)}

At the branching time tbt_b, the coordinates and forces are identical. Defining the translation-corrected mean-square separation by DRMS2=N−1∑i∣δri−δr‾∣2D_{\mathrm{RMS}}^2 = N^{-1}\sum_i\lvert\delta r_i-\overline{\delta r}\rvert^2, the free-particle reference in two dimensions is

⟨DRMS2(t)⟩free=8(1−1N)∫tbtD(s) ds.(5)\left\langle D_{\mathrm{RMS}}^2(t)\right\rangle_{\mathrm{free}} = 8\left(1-\frac{1}{N}\right)\int_{t_b}^{t}D(s)\,\mathrm{d}s. \tag*{(5)}

Interactions subsequently generate force differences that modify the free relative diffusion. We use Eq. (5) as a reference rather than as an interacting-particle fit.

Each primary parent was initialized from a weakly perturbed square arrangement and evolved independently for 2.0 reduced time units at Thigh=1.8T_{\mathrm{high}}=1.8. The 40 parents used distinct initialization and Brownian seeds. We therefore refer to them as independently generated high-temperature parent states, with the finite burn-in defining the sampling scope.

After parent generation, temperature was held at ThighT_{\mathrm{high}} for thold=0.2t_{\mathrm{hold}}=0.2 and then decreased linearly to Tlow=0.50T_{\mathrm{low}}=0.50 over a duration τq\tau_q. The protocol was

T(t)=Thigh,t<thold,T(t)=Thigh−Thigh−Tlowτq×(t−thold),thold≤t≤thold+τq.(6)\begin{aligned} T(t)&=T_{\mathrm{high}}, && t<t_{\mathrm{hold}},\\ T(t)&=T_{\mathrm{high}}-\frac{T_{\mathrm{high}}-T_{\mathrm{low}}}{\tau_q}\times(t-t_{\mathrm{hold}}), && t_{\mathrm{hold}}\leq t\leq t_{\mathrm{hold}}+\tau_q. \tag*{(6)} \end{aligned}

and T(t)=TlowT(t)=T_{\mathrm{low}} thereafter. We used τq=0.15\tau_q=0.15, 0.60, 1.50, and 3.00. Every trajectory continued for 2.0 reduced time units after the quench ended. The quantity τq\tau_q is the quench duration, equivalently the protocol timescale.

For every parent and duration, three replicas reused the same stored pre-quench trajectory. At t=tholdt=t_{\mathrm{hold}}, each replica switched to a distinct Brownian-noise seed. The primary data set therefore contained

Ntraj=40 parents×4 durations×3 replicas=480.(7)N_{\mathrm{traj}}=40\ \text{parents}\times4\ \text{durations}\times3\ \text{replicas}=480. \tag*{(7)}

We refer to the three replicas sharing one parent and one quench duration as a replica set. If qp,cq_{p,c} is the finite-time ordered-outcome probability for parent pp and protocol cc, pc=Epqp,cp_c=E_pq_{p,c}, and dcd_c is the probability that two independent replicas have different outcomes, then

pc(1−pc)=dc2+Var⁡p(qp,c).(8)p_c(1-p_c)=\frac{d_c}{2}+\operatorname{Var}_p(q_{p,c}). \tag*{(8)}

This identity follows from dc=2Ep[qp,c(1−qp,c)]d_c=2E_p[q_{p,c}(1-q_{p,c})]. It separates within-parent future-noise branching from variation in finite-time outcome propensity among parents. Estimation with three replicas and complete-parent resampling is described in Supplementary Information.

The primary design is summarized in Table 1; all primary and comparison ensembles are listed in Supplementary Information, Table S1.

RoleNTemperature(s)τq\tau_qParent blocksReplicas per setRunsΔt\Delta t
Primary ensemble64Tlow=0.50T_{\mathrm{low}}=0.500.15, 0.60, 1.50, 3.004034805×10−55 \times10^{-5}
Fine-step ensemble64Tlow=0.50T_{\mathrm{low}}=0.50same four4034802×10−52 \times10^{-5}
Independent larger-system ensemble128Tlow=0.35T_{\mathrm{low}}=0.35, 0.50, 0.65same four20 shared across 12 conditions12402×10−52 \times10^{-5}

Table 1. Simulation and evaluation design. The primary ensemble also supplies the primary-timestep reference for the timestep comparison

Structural coordinates and finite-window outcomes

Neighbors of particle ii were defined by rij≤rnbr_{ij}\leq r_{\mathrm{nb}} with rnb=1.5r_{\mathrm{nb}}=1.5. This fixed first-shell graph cutoff was chosen independently of terminal outcomes. The complex local sixfold bond-orientational order was defined in the standard form [21] as

ψ6,i=ni−1∑j∈Niexp⁡(6iθij).(9)\psi_{6,i}=n_i^{-1}\sum_{j\in\mathcal{N}_i}\exp(6\mathrm{i}\theta_{ij}). \tag*{(9)}

where nin_i is the coordination number and θij\theta_{ij} is the bond angle. The global magnitude was

Ψ6=∣N−1∑iψ6,i∣.(10)\Psi_6=\left|N^{-1}\sum_i\psi_{6,i}\right|. \tag*{(10)}

A particle was assigned to the locally ordered mask O when 5≤ni≤75\leq n_i\leq7 and ∣ψ6,i∣≥0.6\lvert\psi_{6,i}\rvert\geq0.6; its population fraction was ford=∣O∣/Nf_{\mathrm{ord}}=\lvert O\rvert/N. This mask supplies the population definition used below and the connectivity measures in Supplementary Information [17].

The ordered-particle orientational coherence was

Cord=∣∑i∈Oψ6,i∣/∑i∈O∣ψ6,i∣,(11)C_{\mathrm{ord}}=\left|\sum_{i\in O}\psi_{6,i}\right|/\sum_{i\in O}\left|\psi_{6,i}\right|, \tag*{(11)}

with Cord=0C_{\mathrm{ord}}=0 for an empty ordered set. Additional neighbor-wide coherence and ordered-mask connectivity measures, together with normalization sensitivities, are defined in Supplementary Information.

To test whether the outcome association extended beyond bond-orientational coordinates, we evaluated reciprocal-space translational order through the static structure factor at reciprocal vectors allowed by the square periodic box,

S(k)=N−1∣∑j=1Nexp⁡(ik⋅rj)∣2,(12)S(\mathbf{k})=N^{-1}\left|\sum_{j=1}^{N}\exp(\mathrm{i}\mathbf{k}\cdot\mathbf{r}_j)\right|^2, \tag*{(12)}
k=2πL(nx,ny).(13)\mathbf{k}=\frac{2\pi}{L}(n_x,n_y). \tag*{(13)}

For a triangular reference, a△=[2/(3ρ)]1/2a_{\triangle}=[2/(\sqrt{3}\rho)]^{1/2} and k0=4π/(3a△)k_0=4\pi/(\sqrt{3}a_{\triangle}). The translational-order score sTs_T was the mean of the three largest S(k)S(\mathbf{k}) values in the first-shell annulus around k0k_0. The number three was chosen to match the number of inversion-unique directions in an ideal triangular first shell. Its definition uses neither bond angles, the ordered mask, terminal labels, nor fitted models. The reciprocal-grid construction and an annular-mean sensitivity are described in Supplementary Information.

Terminal statistics were calculated over the last 15% of stored frames, with a minimum of five frames. A trajectory received the finite-window ordered label when all three conditions held within that window:

mean⁡(Ψ6)≥0.55,mean⁡(ford)≥0.75,sd⁡(Ψ6)≤0.08.(14)\begin{aligned} \operatorname{mean}(\Psi_6)&\ge0.55,\\ \operatorname{mean}(f_{\mathrm{ord}})&\ge0.75,\\ \operatorname{sd}(\Psi_6)&\le0.08. \tag*{(14)} \end{aligned}

Every other trajectory received the finite-window non-ordered label. These labels classify the structure reached by the end of the simulated two-unit hold and are used as terminal outcomes throughout the analysis.

These cutoffs define the primary terminal outcome and were held fixed during model training and the primary analyses; only the explicitly identified terminal-window and threshold-sensitivity analyses recomputed terminal outcomes. We also recorded the first time t∗t^* at which a trailing window spanning 7.5% of that trajectory, with a minimum of two stored frames, simultaneously crossed the Ψ6\Psi_6 and fordf_{\mathrm{ord}} thresholds. A finite crossing followed by a terminal non-ordered label was classified as transient. Alternative terminal windows, threshold sensitivities, and the complete crossing-time procedure are provided in Supplementary Information.

Observation times, state score, and replica divergence

We sampled structure at eight predefined observation phases: the shared pre-quench state; 10%, 25%, 50%, and 100% of the trajectory-specific quench; and absolute times 0.25, 0.50, and 1.00 after quench completion. Figures abbreviate these phases as pre, q10, q25, q50, q100, p025, p050, and p100. All observation histories ended before, and did not overlap, the terminal window.

To follow structural development across protocols, quench-relative coordinates aligned equal fractions of each quench, whereas absolute post-quench coordinates aligned equal elapsed times after quench completion. We write the latter as tpost=t−thold−tqt_{\mathrm{post}}=t-t_{\mathrm{hold}}-t_q. The dense structural profiles used 21 quench-relative and 24 absolute post-quench coordinates, giving 45 sampled times in total. Frame selection, history construction, and window-separation checks are given in Supplementary Information.

The threshold-normalized structural score used the latest eligible frame:

sstate=min⁡(Ψ6/0.55,ford/0.75).(15)s_{\mathrm{state}}=\min(\Psi_6/0.55,f_{\mathrm{ord}}/0.75). \tag*{(15)}

Its denominators are the instantaneous structural thresholds in the terminal ordered-label rule. We retain the short name “state score” below; it is a summary relative to this operational outcome definition, not a new thermodynamic order parameter.

For two replicas from the same parent and duration, we measured particle-identity root-mean-square (RMS) separation using periodic minimum-image displacements after subtracting their mean translational displacement. We also formed radius graphs at rnb=1.5r_{\mathrm{nb}}=1.5 and measured the Jaccard distance between undirected edge sets. The pair-level target was whether the two replicas had different terminal labels.

Evaluation and uncertainty

As a flexible representation control, a GraphSAGE encoder and long short-term memory network were also trained on local particle environments and their recent evolution [22, 23]. These graph models probe whether the tested local histories contain outcome information beyond the interpretable collective coordinates; their architectures, inputs, and optimization settings are provided in Supplementary Information.

The graph models were evaluated by parent-grouped cross-validation, with all 12 descendants of a parent kept in the same fold and model selection confined to training and validation data [24, 25]. This procedure yielded held-out scores for every trajectory. Outcome discrimination for both structural coordinates and learned scores was quantified by the area under the receiver operating characteristic curve (AUROC), with equal weight assigned to the four quench durations.

Uncertainty was estimated by resampling complete parent clusters [26, 27]. For learned models, a hierarchical interval additionally propagated variation across repeated fits. Pointwise curves describe structural development, and simultaneous bands account for comparison over sampled times or observables. Fixed-lead measurements at 0.25, 0.50, 0.75, and 1.00 before the terminal-window start hold the distance to terminal assessment constant across protocols. For duration-adjusted correlations, ranks were centered within each quench duration before pooling. Between-parent outcome variation was corrected for the finite-replica Bernoulli contribution. Full resampling procedures and association statistics are given in Supplementary Information.

Timestep and larger-system comparisons

To test sensitivity to time discretization, we repeated the primary design with the same 40 parents, four durations, and three replica identifiers at Δt=2×10−5\Delta t = 2 \times10^{-5}, compared with the primary Δt=5×10−5\Delta t = 5 \times10^{-5}. Because changing the timestep also changes the Brownian-noise grid, this is a distributional comparison rather than a pathwise or continuum test. We repeated the terminal and fixed-lead structural analyses after either reducing the timestep or replacing the fractional terminal window with a fixed final window of 0.50 time units. Numerical checks and paired-comparison procedures are described in Supplementary Information.

The larger-system comparison crossed 20 N=128N = 128 parents generated independently of the N=64N = 64 primary ensemble with Tlow=0.35T_{\mathrm{low}} = 0.35, 0.50, or 0.65 and the same four durations. These trajectories used the fine timestep, so the comparison probes persistence of structural outcome discrimination under a compound change in system size and time discretization. Binary AUROC was averaged over durations containing both classes, with parents as bootstrap clusters. Duration-adjusted correlations with terminal mean Ψ6\Psi_6 supplied a continuous endpoint when the binary label saturated.

Results

Independent noise produces branching from a shared microscopic state

The three replicas in each set inherited exactly the same pre-quench configuration. Their initial coordinates were identical, so all subsequent separation originated from the independent Brownian histories applied after branching.

The 480 trajectories contained 122 ordered terminal outcomes, but these outcomes were distributed broadly within the shared-parent sets. Of the 160 three-replica sets, 66 contained no ordered outcome, 68 contained one, 24 contained two, and 2 contained three. Thus 92 sets (57.5%) contained both outcomes, and 184 of the 480 within-set replica pairs (38.3%) ended with different labels (Figure 1). The observed count distribution was close to the duration-wise Bernoulli expectation (Supplementary Information, Fig. S1).

Figure 1. Stochastic branching under a shared initial configuration and quench protocol. (a) Four temperature protocols and the eight sampled phases. The protocols differ only in quench duration; each reaches Tlow=0.50T_{\mathrm{low}} = 0.50 and then remains there for 2.0 reduced time units. Coloured horizontal segments mark the terminal windows. (b) Three independent-noise descendants of one mixed replica set, shown one time unit after quench completion and labelled by their later outcomes. Particle colour denotes local ∣ψ6,i∣|\psi_{6,i}|. (c) Ordered-outcome counts among the three replicas in all 160 sets; 92 sets contain both outcomes. (d) Left: discordant-pair fractions for the four quench durations. Right: decomposition of the observed parent-level spread into finite-replica sampling and the corrected parent component. The error bar is the 95% complete-parent bootstrap interval, and the dashed line marks zero excess parent variance. Times are in reduced units

Equation (8) formalizes this branching as a competition between future noise and parent-dependent outcome propensity. After correcting the observed spread for the finite-three-replica contribution, the duration-averaged between-parent component was −0.0009-0.0009 [−0.0180-0.0180, 0.01390.0139]. Its non-negative point estimate was zero, and the upper bootstrap percentile of the parent-attributable fraction was 7.5%. The three-replica design cannot resolve small differences among individual parent probabilities, but it provides no evidence that such differences dominate the observed branching. More directly, the 92 mixed sets show that a common microscopic parent and protocol did not uniquely determine the later structure.

Rapid microscopic decorrelation precedes outcome-associated collective order

Independent-noise replicas decorrelated rapidly at the microscopic level. At q10, whose actual elapsed time after branching ranged from 0.008 to 0.296 across the four protocols, the duration-averaged particle-identity RMS separation was approximately 0.55 and the neighbor-graph Jaccard distance approximately 0.36; both continued to grow at later phases (Figure 2(a,Figure 2b)). The same curves were obtained, within uncertainty, for pairs that later shared an outcome and pairs that did not.

Figure 2. Rapid microscopic decorrelation and the later development of structural outcome discrimination. (a,b) Translation-corrected RMS separation and neighbor-graph Jaccard distance for pairs with the same or different outcomes. Grey, blue, and orange backgrounds denote the shared pre-branch state, quench-relative phases, and post-quench phases. (c) AUROC for distinguishing the two pair types from the separation magnitude; all intervals include 0.5. (d) Duration-averaged state-score AUROC after quench completion. The grey region is the pointwise 95% interval, the dotted curve is the simultaneous lower band over all 45 sampled times, and open squares show the estimates at tpost=0.50t_{\mathrm{post}} = 0.50 and 1.00 from the grouped-fold analysis. This lower band remains above 0.5 from tpost=0.60t_{\mathrm{post}} = 0.60 onward. Times are in reduced units

The free-particle reference based on time-integrated diffusivity in Eq. (5) accounted for the scale and protocol dependence of this separation without making it outcome specific. Across q10, q25, q50, and q100, the measured mean-squared separation had a descriptive through-origin slope χ=0.1718\chi= 0.1718 [0.1669, 0.1769] relative to the free-particle prediction (Supplementary Information, Figure S16). The observed mean-squared separation was therefore about one-sixth of the free reference across these checkpoints. More importantly, the slopes were 0.1719 [0.1659, 0.1783] for same-outcome pairs and 0.1717 [0.1652, 0.1783] for different-outcome pairs; their difference was −0.0002 [−0.0087, 0.0085]. The response of the distance to time-integrated diffusivity was thus common to the two pair types.

Figure S16. Replica separation relative to time-integrated diffusivity. (a) Translation-corrected mean-squared replica separation, averaged within each duration-by-checkpoint combination, versus the free-particle reference XX in Eq. (17). The line is a descriptive through-origin fit to all replica pairs; the free-particle relation has unit slope. (b) Fitted slope χ\chi for all pairs and for pairs with the same or different terminal outcomes. Bars show 95% complete-parent bootstrap intervals. The different-minus-same interval includes zero, so rapid separation follows the time-integrated diffusivity similarly for the two pair types

The collective coordinates behaved differently. Immediately after quench completion, the outcome-resolved means were nearly coincident (Figure 3). By tpost=0.55t_{\mathrm{post}} = 0.55, ordered and non-ordered outcomes had mean Ψ6\Psi_6 values 0.502 and 0.444, ordered-particle fractions 0.694 and 0.665, coherences CordC_{\mathrm{ord}} of 0.752 and 0.690, and translational scores sTs_T of 19.84 and 16.82. The gaps continued to grow through tpost=1.20t_{\mathrm{post}} = 1.20. These raw structural curves show that the AUROC rise reflects a progressive shift of the physical order distributions rather than only a change in rank tails.

Figure 3. Outcome-resolved development of collective structure after quench completion. Duration-equal means are shown for trajectories with ordered and non-ordered terminal outcomes: (a) global sixfold order Ψ6\Psi_6, (b) ordered-particle fraction fordf_{\mathrm{ord}}, (c) ordered-particle coherence CordC_{\mathrm{ord}}, and (d) reciprocal-space translational-order score sTs_{\mathrm{T}}. Shading gives pointwise 95% complete-parent bootstrap intervals. Dotted lines in (a,b) mark the corresponding instantaneous structural thresholds; the terminal outcome additionally uses a finite window and a fluctuation criterion. Times are in reduced units

The dense-profile state-score AUROC rose from 0.560 at tpost=0.50t_{\mathrm{post}} = 0.50 to 0.658 [0.602, 0.714] at 1.00 (Figure 2(d)). At tpost=0.55t_{\mathrm{post}} = 0.55, raw Ψ6\Psi_6, the state score, CordC_{\mathrm{ord}}, and sTs_T had AUROCs 0.605, 0.592, 0.599, and 0.631, respectively. The separately defined sTs_T had duration-adjusted rank correlations of 0.838–0.872 with the three orientational measures, consistent with coupled translational and orientational development (Supplementary Information, Table S7).

Fixed-lead sampling quantified how far this structural association extended before terminal assessment. At a lead of 1.00, raw Ψ6\Psi_6, the state score, CordC_{\mathrm{ord}}, and sTs_T had equal-duration AUROCs of 0.599, 0.593, 0.605, and 0.593; at a lead of 0.50, these values increased to 0.660, 0.647, 0.646, and 0.673. Changing the terminal window or reducing the timestep preserved simultaneous lower limits above 0.5 for all four measures at leads 0.75 and 0.50, but not at lead 1.00, under the common 4×44 \times4 orientation-plus-translation family. Because the fixed-window margins at 0.75 were small, we use 0.50 as a conservative common reference for the cross-specification comparison. Complete intervals and duration-resolved results are given in Supplementary Information, Table S6 and Figure S12.

AnalysisLeadRaw Ψ6\Psi_{6}State scoreCordC_{\mathrm{ord}}sTs_{T}
Primary analysis1.000.599 [0.517, 0.680]0.593 [0.511, 0.674]0.605 [0.523, 0.686]0.593 [0.511, 0.675]
Primary analysis0.750.647 [0.565, 0.729]0.628 [0.546, 0.710]0.655 [0.573, 0.737]0.637 [0.555, 0.719]
Primary analysis0.500.660 [0.578, 0.742]0.647 [0.566, 0.729]0.646 [0.564, 0.727]0.673 [0.591, 0.754]
Primary analysis0.250.685 [0.603, 0.766]0.655 [0.573, 0.737]0.701 [0.620, 0.783]0.703 [0.621, 0.785]
Fixed-window-label analysis1.000.551 [0.466, 0.636]0.548 [0.463, 0.633]0.556 [0.471, 0.641]0.560 [0.475, 0.645]
Fixed-window-label analysis0.750.590 [0.505, 0.675]0.590 [0.505, 0.674]0.590 [0.505, 0.675]0.612 [0.527, 0.697]
Fixed-window-label analysis0.500.597 [0.512, 0.682]0.607 [0.522, 0.692]0.596 [0.511, 0.681]0.616 [0.531, 0.700]
Fixed-window-label analysis0.250.692 [0.607, 0.777]0.683 [0.598, 0.768]0.690 [0.605, 0.775]0.691 [0.606, 0.776]
Fine-step analysis1.000.546 [0.460, 0.631]0.542 [0.456, 0.627]0.544 [0.459, 0.630]0.539 [0.454, 0.625]
Fine-step analysis0.750.614 [0.529, 0.700]0.615 [0.529, 0.700]0.605 [0.519, 0.690]0.619 [0.534, 0.705]
Fine-step analysis0.500.618 [0.532, 0.703]0.598 [0.512, 0.683]0.626 [0.541, 0.712]0.627 [0.542, 0.713]
Fine-step analysis0.250.732 [0.646, 0.817]0.717 [0.632, 0.803]0.704 [0.619, 0.790]0.731 [0.645, 0.816]

Table S6. Fixed-lead results under the joint 4×44 \times4 orientation-plus-translation family. Each entry is the equal-duration AUROC followed by its simultaneous 95% parent-cluster interval. The family contains raw Ψ6\Psi_{6}, the structural state score, CordC_{\mathrm{ord}}, and sTs_{T} at all four leads. All intervals use 5000 parent-cluster draws

Figure S12. Quench-duration-specific fixed-lead comparison. (a–c) Duration-specific AUROCs for raw Ψ6\Psi_6, the state score, and Cord\mathit{C}_{\mathrm{ord}} at four leads before the primary terminal window. Error bars are pointwise parent-cluster 95% intervals; the corresponding simultaneous families contain 16 cells per observable and 48 cells jointly. At lead 1.0, all three pointwise lower limits exceed 0.5 only for τq=0.15\tau_q = 0.15, and no duration has this property under the 16-cell correction. Fixed lead maps to different post-quench ages across protocols. At lead 0.50, all three 16-cell lower limits exceed 0.5 for τq=0.15\tau_q = 0.15 and 0.60, but none do so for the two longer durations; Table S7 adds sTs_T and the joint 64-cell family. (d) Selected post-quench time as a function of quench duration for each requested lead. This deterministic time-grid map makes the fixed-lead tradeoff explicit

A scalar potential-energy score and alternative definitions of local and translational order showed the same qualitative development, although their individual time-grid summaries varied (Supplementary Information, Figure S9, Figure S14, and Figure S15). The evidence therefore supports collective outcome-associated ordering more directly than a unique chronology among its orientational, translational, connectivity, and energetic components.

Figure S9. Timestep and structural-definition sensitivity analyses. (a) Paired fine-step-minus-primary fixed-lead AUROC differences for raw Ψ6\Psi_6, the state score, and CordC_{\mathrm{ord}}. Error bars are joint 3×43 \times4 parent-cluster intervals; none excludes zero. (b) Neighbor-phase-coherence curves under three endpoint-amplitude cutoffs and an amplitude-weighted definition. Solid curves are point estimates, and dashed curves are 24-time post-quench observable-wise lower bands. (c) State-score and baseline-GNN AUROC for the cohort defined at p050 by excluding trajectories that already satisfy both instantaneous thresholds and then followed unchanged to p100; error bars show parent-cluster intervals, and the learned-model comparison also uses the hierarchical seed–parent bootstrap interval reported in Table S2. (d) Number of valid seed–fold–duration cells for each explicitly labeled terminal-window or threshold variant. Strict variants change the AUROC evaluation domain. The q25 state estimate is not distinguished from chance because both relevant upper limits exceed 0.5

Figure S14. Non-orientational potential-energy comparison. (a,b) Dense within-duration AUROC for sE=−U/(Nε)s_E = -U/(N\varepsilon) under the primary fractional and fixed final-0.50 terminal labels. Shading gives pointwise parent-cluster intervals; dashed and dotted curves are simultaneous lower bands over the full 45-time and 24-time post-quench families. The full 45-time lower band does not remain above 0.5 for either label through tpost=1.20t_{\mathrm{post}} = 1.20; the primary-label 24-time post-quench lower band does so only at the final 1.20 grid point. (c,d) Fixed-lead AUROCs for raw Ψ6\Psi_6, the state score, CordC_{\mathrm{ord}}, and sEs_E under the two labels. Error bars are joint 95% parent-cluster intervals over all four observables and four leads. At the primary-label lead of 1.0, all three orientational observables remain above chance after correction whereas sEs_E does not. Under the fixed label, none has a lower limit above 0.5 at lead 1.0; the three orientational observables, but not sEs_E, have lower limits above 0.5 at lead 0.50. Under the primary label, sEs_E itself has lower limits above 0.5 at leads 0.50 and 0.25, with AUROCs 0.625 and 0.601. The energy comparison is therefore lead dependent: it does not reproduce the longest-lead or complete dense-time result, but it is not null near the terminal window. All observations strictly precede the applicable label window

Figure S15. Translational-order comparison defined without terminal-label information. (a,b) Dense within-duration AUROC for the top-three first-reciprocal-shell score sTs_T and the annular mean under the primary and fixed-window labels. Shading gives pointwise 95% parent-cluster intervals; dashed curves are joint lower bands over both translational scores and all 45 sampled times. (c) Fixed-lead sTs_T in the primary, fixed-window-label, and fine-step analyses. Error bars are joint 4×44 \times4 intervals over raw Ψ6\Psi_6, the state score, CordC_{\mathrm{ord}}, and sTs_T. All three lower limits exceed 0.5 at lead 0.50; only the primary-analysis lower limit exceeds 0.5 at lead 1.0. (d) Paired fine-step-minus-primary contrasts for both translational scores. Intervals use the joint 5×45 \times4 numerical family and all cross zero. The annular mean is a non-maximizing sensitivity to the top-three summary. The metric geometry and aggregation do not use the terminal labels. All observations precede the applicable terminal window

Transient crossings separate early alignment from sustained ordered-particle growth

The binary terminal labels mark two regions of an underlying continuous structural distribution. The terminal fraction of six-coordinated particles was 0.684 [0.679, 0.690] for ordered outcomes and 0.610 [0.605, 0.615] for non-ordered outcomes. Among the 39 parents that contributed both outcomes, the mean within-parent contrast in this fraction was 0.076 [0.069, 0.083] (Supplementary Information, Figure S13). The bond-orientational pair correlation G6(r)G_6(r) was likewise higher for ordered outcomes over several neighbor shells, whereas differences in g(r)g(r) were modest (Figure 4(c); Supplementary Information, Figure S8). These observations place the two outcome classes at different degrees of sixfold order and orientational coherence.

Figure 4. Transient threshold crossings and sustained terminal order. (a) Fraction of trajectories with a finite short-window crossing time t∗t^{*}. All ordered trajectories and 219 non-ordered trajectories cross. (b) Overlapping terminal-criterion failures among those 219 transient crossings; most end below the ordered-particle population criterion. (c) Outcome-resolved bond-orientational pair correlation G6(r)G_{6}(r). (d) Duration-adjusted associations of current raw Ψ6\Psi_{6}, state score, and CordC_{\mathrm{ord}} with the terminal fraction of six-coordinated particles. Reading left to right moves closer to the terminal window as the lead decreases from 1.00 to 0.25. Error bars in (a–c) are pointwise parent-cluster bootstrap intervals; those in (d) are 95% intervals jointly adjusted across three measures and four leads. Times and leads are in reduced units

Figure S8. First crossings and terminal radial structure. (a) Empirical cumulative distributions of finite first-crossing times relative to quench completion. The outcome distributions overlap, and 219 of 358 terminal non-ordered trajectories cross transiently. (b) Terminal-window radial distribution g(r)g(r) by operational outcome. Curves are means and 95% parent-cluster intervals; the differences are modest. Separation is reported as r/σr/\sigma

Figure S13. Terminal coordination and parent-level contrasts. (a) Full outcome-resolved terminal cutoff-coordination fractions and parent-cluster intervals. (b) Parent-by-outcome means of the terminal fraction of six-coordinated particles. Gray lines connect the 39 parents contributing both labels; horizontal bars show equal-parent outcome means. The annotated ordered-minus-non-ordered paired contrast and interval come from parent-cluster resampling

Short-window threshold crossings were frequently reversible. The first-crossing time t∗t^* was finite for all 122 terminal ordered trajectories and for 219 of the 358 terminal non-ordered trajectories. Within each quench duration, 52.9–67.8% of the non-ordered trajectories crossed the short-window rule before returning to the non-ordered terminal class (Figure 4(a)). Their first-crossing-time distributions substantially overlapped those of the ordered trajectories (Supplementary Information, Figure S8).

The subsequent loss of an ordered-particle population was the dominant feature of these transient passages. Of the 219 non-ordered trajectories that crossed, 207 (94.5% [91.6%, 97.2%]) ended below the terminal fordf_{\mathrm{ord}} criterion. By comparison, 95 (43.4%) ended below the Ψ6\Psi_6 criterion and 75 (34.2%) exceeded the fluctuation cutoff; these categories overlap (Figure 4(b)). The dense fordf_{\mathrm{ord}} profile became persistently outcome-associated later than raw Ψ6\Psi_6, CordC_{\mathrm{ord}}, and sTs_{\mathrm{T}}. Because fordf_{\mathrm{ord}} also enters the terminal label, this difference is descriptive rather than a mechanistic onset sequence. It is consistent with a large ordered-particle population being sustained after orientational and translational order have begun to develop.

The association also extended to the terminal fraction of six-coordinated particles, a continuous endpoint that is not part of the binary-label formula. At the longest tested lead, current raw Ψ6\Psi_6, the state score, and CordC_{\mathrm{ord}} had duration-adjusted Spearman correlations of 0.168, 0.156, and 0.202 with this endpoint. Their joint intervals were [0.060, 0.276], [0.048, 0.265], and [0.094, 0.310], respectively (Figure 4(d)). The same physical ordering process is therefore visible without reducing the terminal structure to two classes.

Protocol, timestep, and larger-system comparisons

Stochastic branching occurred at all four quench durations, with no detectable monotonic change in ordered-outcome probability over the tested protocol timescales (Supplementary Information, Fig. S1). The aggregate separation between microscopic divergence and collective ordering therefore did not simply track a monotonic outcome-frequency trend, although duration-resolved discrimination was most consistently resolved for the two shorter quenches.

The fine-step ensemble retained the same qualitative growth of post-quench outcome discrimination. Although the sampled time at which a simultaneous lower band cleared 0.5 changed, all four structural measures remained informative at a lead of 0.50 (Supplementary Information, Table S6 and Figs. S7 and S15). The physical conclusion therefore persisted without assigning an exact continuum onset.

The same structural state score showed strong outcome discrimination for independently generated N=128N = 128 parents. At Tlow=0.35T_{\mathrm{low}}=0.35, its AUROCs

The fine-step ensemble retained the same qualitative growth of post-quench outcome discrimination. Although the sampled time at which a simultaneous lower band cleared 0.5 changed, all four structural measures remained informative at a lead of 0.50 (Supplementary Information, Table S6 and Figs. S7 and S15). The physical conclusion therefore persisted without assigning an exact continuum onset.

The same structural state score showed strong outcome discrimination for independently generated N=128N = 128 parents. At Tlow=0.35T_{\mathrm{low}}=0.35, its AUROCs were 0.898 [0.811, 0.960] at tpost=0.50t_{\mathrm{post}} = 0.50 and 0.936 [0.880, 0.982] at 1.00. At Tlow=0.50T_{\mathrm{low}}=0.50, the corresponding values were 0.703 [0.563, 0.823] and 0.858 [0.735, 0.946] (Figure 5(a)).

Figure 5. Structural outcome relation for independently generated N=128N = 128 parents. (a) State-score AUROC at tpost=0.50t_{\mathrm{post}} = 0.50 and 1.001.00 for temperatures containing both terminal outcomes. (b) Ordered-outcome fraction across the three low temperatures; error bars are descriptive trajectory-level Clopper–Pearson intervals under a Bernoulli reference model, and no inferential comparison relies on them. (c) Duration-adjusted Spearman correlation between the state score at tpost=1.00t_{\mathrm{post}} = 1.00 and terminal mean Ψˉ6\bar{\Psi}_6. Error bars in (a,c) are 95% complete-parent bootstrap intervals

The ordered-outcome fraction decreased from 51/80 at Tlow=0.35T_{\mathrm{low}}=0.35 to 20/80 at 0.50 and 0/80 at 0.65 (Figure 5(b)). As a continuous complement, the state score at tpost=1.00t_{\mathrm{post}} = 1.00 correlated with terminal mean Ψ6\Psi_6 at all three temperatures, with duration-adjusted correlations 0.722, 0.707, and 0.434, respectively (Figure 5(c)). This relation therefore remained measurable at Tlow=0.65T_{\mathrm{low}}=0.65, where the binary label saturated and AUROC was undefined.

Graph models recovered the late outcome association but did not reveal a comparably supported earlier or independent coordinate. Adding the tested graph scores to state and duration produced no detectable held-out gain, and feature sensitivities pointed back to local sixfold order (Supplementary Information, Figure S3, Figure S4, and Figure S11). They therefore serve as a representation control rather than the physical basis of the conclusion.

Figure S3. Baseline-model feature dependence. (a) Test-time training-mean replacement across phases and node channels. (b) Absolute integrated-gradient shares at q25 and p100. The six plotted channels are zi/6z_i/6, ⟨rij⟩/rnb\langle r_{ij}\rangle/r_{\mathrm{nb}}, SD⁡(rij)/rnb\operatorname{SD}(r_{ij})/r_{\mathrm{nb}}, ∣ψ6,i∣|\psi_{6,i}|, ∣zi−6∣/6|z_i-6|/6, and Δri/rnb\Delta r_i/r_{\mathrm{nb}}. Local ∣ψ6∣|\psi_6| dominates the late p100 dependence. The quantities are reference-dependent model sensitivities rather than causal effects

Figure S4. Current-state and residual-information analyses. (a) Instantaneous threshold rates at p050 and p100. (b) AUROC after excluding trajectories that already meet both structural thresholds; error bars are parent-cluster bootstrap intervals. The p100 targeted-GNN lower limit is 0.501, but its hierarchical seed–parent bootstrap interval [0.481, 0.672] crosses 0.5. (c) Cross-fitted state-adjusted GNN increments. (d) Incremental association of current CordC_{\mathrm{ord}} with later Cmax⁡/NC_{\max}/N after controlling for current state and duration. All intervals for the residual and coherence increments include zero

Figure S11. Exploratory graph-representation tests. (a) Eight-phase comparison of the state score, baseline GNN, and duration-conditioned GNN. The two predefined graph models first pass at p100. (b) Comparison at q100, p025, p050, and p100 with the targeted GNN sensitivity model; p050 is the first sampled phase at which both lower limits exceed 0.5. The gray, blue, and orange backgrounds in panel (a) denote pre-quench, quench-relative, and post-quench checkpoints, respectively. (c) Paired targeted-minus-baseline and targeted-minus-state AUROC contrasts. At p050, these contrasts are +0.085 [0.038, 0.127] and +0.023 [−0.019, 0.069], respectively. (d) Cross-fitted state-adjusted AUROC increments for baseline, duration-conditioned, and targeted GNNs; every interval includes zero at p050 and p100. The targeted GNN comparison changes the added channel, history construction, auxiliary losses, and checkpoint selection together, so it is not a controlled representation ablation. Error bars are parent-cluster bootstrap intervals. In panels (a–c), filled markers have hierarchical seed–parent lower limits above 0.5

Discussion

The shared-parent ensemble separates two processes that are often conflated. Equation (8) shows how a common parent can still have a distribution of future outcomes, and the mixed replica sets demonstrate that this stochastic branching is frequent in the present ensemble. Equation (5) then provides the natural thermal scale for the initial loss of microscopic similarity. Interparticle forces reduce the separation relative to the free reference, but the same exposure response is observed for pairs that later share an outcome and pairs that do not. Generic decorrelation is therefore not itself an indicator of assembly fate.

This result has a simple high-dimensional interpretation. Coordinate RMS and graph Jaccard distances collect displacements or edge changes across all available modes. Most of those changes need not point toward or away from collective order, so an unweighted norm can grow while outcome-relevant information remains diluted. Sixfold alignment, ordered-particle coherence, and first-shell translational order instead project the evolving configuration onto symmetry-sensitive collective modes. Their outcome-resolved curves show a gradual physical separation after the quench, rather than only a classifier-score effect. Intermediate-time reshooting would be required to test whether any of these coordinates approximates a finite-time committor [9, 10, 28, 29].

The later dynamics also distinguish the appearance of order from its persistence. The translational score stst uses neither bond angles nor the terminal-label thresholds, yet evolves with the orientational coordinates. Current structure is associated with the terminal fraction of six-coordinated particles, which is not included as a separate term in the binary-label formula. Moreover, short-window threshold crossings occur in both outcome classes, while 94.5% of transient non-ordered crossings subsequently lose the required ordered-particle population. Successful assembly in this finite observation window is therefore associated with maintaining an extended ordered population, not simply reaching a high instantaneous alignment once.

The fixed-lead analysis places this developing association on a common distance from terminal assessment. A lead of 0.50 corresponds to 0.25τB0.25\tau_B at TlowT_{\mathrm{low}}, although this free-particle conversion is not an interacting relaxation time. All four structural measures remained informative at this lead after either changing the terminal window or reducing the timestep; state-score discrimination also persisted in an independently generated N=128N = 128 ensemble. The available data do not select a universal clock for the collective ordering process, so the result is a finite-protocol dynamical separation rather than a universal onset law. The graph-model controls reach the same late structural information without revealing an independent earlier coordinate.

Scope and limitations

The terminal labels describe degrees of finite-window order during a two-unit post-quench hold in an incommensurate square cell. They do not assign equilibrium phases or establish irreversible nucleation or asymptotic kinetic arrest. Three replicas per parent also limit the resolution of parent-specific outcome probabilities; the 7.5% value is an upper bootstrap percentile of a variance fraction, not a predictive-accuracy bound. The free-draining dynamics omit hydrodynamic interactions.

The timestep comparison changes both the timestep and Brownian-noise grid, and the N=128N = 128 comparison changes size and time discretization together. Matched timestep and size hierarchies would be required for continuum or finite-size extrapolation. Independent measurements of the structural relaxation time τα\tau_{\alpha} would additionally express the protocol durations and leads in relaxation-time units. Because orientational and translational order are accessible in particle-resolved imaging, the separation between rapid decorrelation and developing collective order can also be tested in colloidal experiments [6, 12, 30].

Conclusion

Independent Brownian histories rapidly decorrelate replicas launched from a shared microscopic state. Their mean-squared separation remains below the free-particle reference based on time-integrated diffusivity, but the fitted response is the same for replica pairs with equal and different terminal outcomes. The magnitude of microscopic separation therefore carries little information about the later assembly fate.

Outcome association develops instead in collective structure after quench completion. Sixfold orientational order, ordered-particle population and coherence, and translational order progressively separate the terminal outcomes. A fixed-lead comparison spanning global order, the state score combining global and population order, coherence, and translational order remains informative at a tested lead of 0.50 after changing either the terminal window or the integration timestep. Transient crossings show that instantaneous alignment is insufficient: most non-ordered crossings subsequently lose the extended ordered-particle population. Quenched assembly in this ensemble thus combines generic noise-driven decorrelation with the later emergence and finite-window persistence of collective order.

Supplementary information

The Supplementary Information follows the main article and contains extended methods, additional analyses, 16 supplementary figures, and seven supplementary tables.

Statements and declarations

Competing interests

The authors declare no competing interests.

Data and code availability

The original trajectories, processed trajectory-level data, out-of-fold predictions, bootstrap summaries, Brownian-dynamics and observable implementations, and graph-model training source are archived on Zenodo at https://doi.org/10.5281/ zenodo.22179833. The archive includes relative-path manifests, fixed training seeds, and scripts for reproducing the principal results.

Supplementary Information for Rapid microscopic decorrelation precedes sustained collective ordering in quenched two-dimensional colloids

Ruixin Zhou; Boliang Yu

This Supplementary Information gives the derivations and extended methods that support the physical comparison between microscopic replica separation and collective ordering. It also documents the outcome definition, grouped statistical analysis, graph-model representation control, timestep and terminal-window comparisons, and additional structural observables. All analyses use the trajectories described below; model-based analyses additionally use held-out predictions.

supplementary.pdf

Supplementary methods

Quench-duration analysis

For each quench duration τq∈{0.15,0.60,1.50,3.00}\tau_q \in\{0.15, 0.60, 1.50, 3.00\}, the primary ensemble contained 40 parent states and three noise replicas per parent. Descriptive trajectory-level Clopper–Pearson intervals under a Bernoulli reference model accompany the duration-specific ordered-label fractions; no inferential comparison relies on these intervals. Mixed replica sets contain both terminal outcomes among the three replicas, and pair discordance is the fraction of the three within-set replica pairs with different labels. The duration contrast was specified on log⁡τq\log\tau_q and estimated with parent-cluster uncertainty; its slope was −0.0135-0.0135 with 95% interval [−0.0465,0.0200][-0.0465, 0.0200].

The corresponding duration-level outcomes and replica-set counts are shown in Figure S1; the complete ensemble design is summarized in Table S1.

RoleNNTemperature(s)τq\tau_qParent blocksReplicas per setRunsΔt\Delta t
Primary ensemble64Tlow=0.50T_{\mathrm{low}} = 0.50four durations4034805×10−55 \times10^{-5}
Fine-step ensemble64Tlow=0.50T_{\mathrm{low}} = 0.50four durations4034802×10−52 \times10^{-5}
Initializer comparison, each family64T=1.8T = 1.8, no quench—8 per initializer—8 per initializer5×10−55 \times10^{-5}
Independent larger-system ensemble128Tlow=0.35,0.50,0.65T_{\mathrm{low}} = 0.35, 0.50, 0.65four durations20 shared across 12 conditions12402×10−52 \times10^{-5}

Table S1. Primary and comparison ensembles. Parents are indivisible grouping blocks; the primary ensemble also supplies the primary-timestep reference

Figure S1. Quench-duration design and terminal outcome frequencies. (a) Ordered-label fraction and descriptive trajectory-level Clopper–Pearson interval under a Bernoulli reference model by τq\tau_q; no inferential comparison relies on these intervals. (b) Mixed-set and discordant-pair fractions. (c) Observed distribution of zero to three ordered labels per replica set compared with independent Bernoulli expectations formed within duration; the expected mixed-set count is 90.8, compared with 92 observed. (d) Log-duration contrast. The slope is −0.0135 [−0.0465, 0.0200], and its interval includes zero

Relative diffusion and finite-time branching

Let tb=tholdt_b = t_{\mathrm{hold}} denote the instant at which replicas begin to use independent Brownian noise. Subtracting the overdamped Langevin equations for replicas aa and bb gives

dδri=μ(Fi(a)−Fi(b))dt+4D(t) dWi,D(t)=μkBT(t).(16)d\delta\mathbf{r}_i = \mu\left(\mathbf{F}_i^{(a)}-\mathbf{F}_i^{(b)}\right)dt+\sqrt{4D(t)}\,d\mathbf{W}_i,\qquad D(t)=\mu k_B T(t). \tag*{(16)}

For two non-interacting replicas in two dimensions, each Cartesian relative coordinate has variance 4∫tbtD(s) ds4\int_{t_b}^{t}D(s)\,ds. Removing the common translation subtracts the sample mean of the NN displacement vectors and multiplies the particle-averaged variance by 1−1/N1-1/N. Hence

X(t)≡⟨DRMS2(t)⟩free=8(1−1N)∫tbtD(s) ds.(17)X(t)\equiv\left\langle D_{\mathrm{RMS}}^2(t)\right\rangle_{\mathrm{free}}=8\left(1-\frac{1}{N}\right)\int_{t_b}^{t}D(s)\,ds. \tag*{(17)}

For the linear temperature ramp, at fractional quench coordinate α=(t−tb)/τq∈[0,1]\alpha=(t-t_b)/\tau_q\in[0,1],

X(α,τq)=8(1−1N)μkBτq[Thighα−Thigh−Tlow2α2].(18)X(\alpha,\tau_q)=8\left(1-\frac{1}{N}\right)\mu k_B\tau_q\left[T_{\mathrm{high}}\alpha-\frac{T_{\mathrm{high}}-T_{\mathrm{low}}}{2}\alpha^2\right]. \tag*{(18)}

After quench completion this becomes

X(tpost,τq)=8(1−1N)μkB[Thigh+Tlow2τq+Tlowtpost].(19)X(t_{\mathrm{post}},\tau_q)=8\left(1-\frac{1}{N}\right)\mu k_{\mathrm{B}}\left[\frac{T_{\mathrm{high}}+T_{\mathrm{low}}}{2}\tau_q+T_{\mathrm{low}}t_{\mathrm{post}}\right]. \tag*{(19)}

Equations (S2)–(S4) define a free-particle reference, not an effective theory for the interacting liquid. We fitted the 16 duration-by-checkpoint combinations at q10, q25, q50, and q100 to ⟨DRMS2⟩=χX\langle D_{\mathrm{RMS}}^2\rangle=\chi X through the origin. Each combination averages the available replica pairs. Complete-parent resampling preserved all replica pairs and quench durations. The descriptive slope was χ=0.1718\chi=0.1718 [0.1669, 0.1769] for all pairs, 0.1719 [0.1659, 0.1783] for same-outcome pairs, and 0.1717 [0.1652, 0.1783] for different-outcome pairs. Their difference was −0.0002-0.0002 [−0.0087-0.0087, 0.0085] (Figure S16). The force-difference term in Eq. (16) makes χ\chi history dependent, so these four checkpoints do not establish a universal rescaled clock.

The replica design also yields an exact decomposition of finite-time outcome variation. Let Yp,c(a)∈{0,1}Y_{p,c}^{(a)}\in\{0,1\} be the outcome of replica aa from parent pp under protocol cc, and let qp,c=P(Yp,c(a)=1∣p,c)q_{p,c}=P(Y_{p,c}^{(a)}=1\mid p,c). Conditional independence of two future noise realizations gives

P(Yp,c(a)≠Yp,c(b)∣p,c)=2qp,c(1−qp,c).(20)P\left(Y_{p,c}^{(a)}\ne Y_{p,c}^{(b)}\mid p,c\right)=2q_{p,c}(1-q_{p,c}). \tag*{(20)}

Writing pc=Ep(qp,c)p_c=E_p(q_{p,c}) and averaging the discordance probability over parents as dcd_c gives

dc2=Ep(qp,c)−Ep(qp,c2),(21)\frac{d_c}{2}=E_p(q_{p,c})-E_p(q_{p,c}^2), \tag*{(21)}
pc(1−pc)=dc2+Var⁡p(qp,c).(22)p_c(1-p_c)=\frac{d_c}{2}+\operatorname{Var}_p(q_{p,c}). \tag*{(22)}

Thus the Bernoulli variation at fixed protocol separates into within-parent future-noise branching and between-parent variation in finite-time outcome propensity. With three replicas, if kp,ck_{p,c} are ordered, an unbiased estimate of qp,c(1−qp,c)q_{p,c}(1-q_{p,c}) is

W^p,c=kp,c(3−kp,c)3(3−1).(23)\widehat{W}_{p,c}=\frac{k_{p,c}(3-k_{p,c})}{3(3-1)}. \tag*{(23)}

The finite-replica correction and parent-attributable fraction are specified below with the dense-time uncertainty analysis.

Grouped model evaluation

The baseline and duration-conditioned graph neural networks (GNNs) were evaluated at eight predefined phases. The targeted GNN sensitivity model was evaluated at q100, p025, p050, and p100. Each model–phase combination used five parent-grouped outer folds and three repeated splits. All descendants of one parent remained in the same fold. Within every fold–duration cell there were 24 held-out trajectories, and every original cell contained both outcome classes. An out-of-fold prediction was therefore obtained for each of the 480 trajectories in every repeat. Within-duration AUROCs were given equal weight, calculated separately for each repeat, and then averaged. A cell that contained one class after parent resampling was omitted only within that bootstrap draw.

Checkpoint selection remained internal to the training and validation portion of each outer fold. Baseline and duration-conditioned runs maximized a duration-balanced validation AUROC; targeted-GNN runs maximized the mean of the defined duration-specific validation AUROCs. Because the targeted model also changed its input channel, history sampling, and auxiliary objective, it is a combined sensitivity analysis rather than a single-feature ablation.

For deterministic state-score checkpoint comparisons, the same fold assignments preserved the evaluation domain. Its repeat variation reflects only the fold partition, whereas the learned-model repeat layer also includes training variability. Dense structural profiles do not use fold partitions.

Observation histories and window separation

At every requested cutoff, analysis used the latest stored frame not later than that time. The baseline and duration-conditioned GNNs sampled eight frames from each protocol segment available by the cutoff, giving histories of 8, 16, or 24 selected frames. The targeted sensitivity model instead used the last eight consecutive distinct frames. Repeated frame indices produced zero displacement in the corresponding local input.

For all 32 phase-by-duration cells, the complete observation history was strictly disjoint from the terminal label window. The minimum nominal lead to the terminal-window start was 0.232, and the minimum lead after mapping to the actual stored frame was 0.240. The dense structural grid comprised 21 quench-relative coordinates, α=0,0.05,…,1\alpha= 0, 0.05, \ldots, 1, evaluated at t=thold+ατqt = t_{\mathrm{hold}} + \alpha\tau_q, and 24 absolute post-quench coordinates, s=0.05,0.10,…,1.20s = 0.05, 0.10, \ldots, 1.20, evaluated at t=thold+τq+st = t_{\mathrm{hold}} + \tau_q + s. At tpost=1.20t_{\mathrm{post}} = 1.20, the minimum lead to the primary fractional terminal window was 0.032 both nominally and after stored-frame selection; for the fixed final window, the corresponding leads were 0.300 and 0.304.

Graph representations and training

At each selected frame, particles were nodes of the rnb=1.5r_{\mathrm{nb}} = 1.5 radius graph. The baseline node vector contained six invariant local channels: coordination divided by 6, mean neighbor distance divided by 1.5, neighbor-distance standard deviation divided by 1.5, local ∣ψ6∣|\psi_6|, coordination defect ∣ni−6∣/6|n_i - 6|/6, and the scalar magnitude of the periodic minimum-image displacement from the preceding selected frame, ∣Δri∣/1.5|\Delta r_i|/1.5. The displacement magnitude was zero for the first selected frame; a repeated frame index also produced zero displacement.

Three GraphSAGE layers of width 64 generated node embeddings. Mean pooling produced one frame embedding, and a 64-unit long short-term memory (LSTM) network encoded the selected history. The baseline objective combined weighted binary cross-entropy with SmoothL1 regression of the terminal-window mean disordered-particle fraction 1−⟨ford⟩term1 - \langle f_{\mathrm{ord}}\rangle_{\mathrm{term}} and masked regression of t∗/tmax⁡t^*/t_{\max}, with weights 1, 0.5, and 0.2. Here tmax⁡t_{\max} is the final stored time of the corresponding trajectory. The duration-conditioned model retained this representation and objective but concatenated standardized log τq\tau_q to every pooled frame embedding before the LSTM. Its fixed center and scale were −0.225967-0.225967 and 1.1210611.121061.

The targeted GNN sensitivity model added the local phase-coherence channel c6,ic_{6,i}, used the three 64-wide GraphSAGE layers and 64-unit LSTM, used the last eight consecutive distinct frames, and optimized binary classification alone. It was evaluated at q100, p025, p050, and p100; p050 and p100 were used for the state-adjusted residual analysis. Because the input, history, auxiliary losses, and selection rule change together, the targeted-GNN comparison is not a single-feature ablation.

AdamW used learning rate 10−310^{-3}, weight decay 10−410^{-4}, and batch size 32. Training ran for at most 40 epochs with gradient-norm clipping at 1 and early-stopping patience 8. Class weights came only from the training fold. Baseline and duration-conditioned checkpoints maximized duration-balanced validation AUROC; targeted-GNN checkpoints maximized the macro-average of defined duration-specific validation AUROCs. Selection remained internal to each outer fold.

Terminal labels and local definitions

The primary label used the final 15% of stored frames, whereas the fixed-window label used the final 0.50 absolute time units. The latter produced 121 ordered labels and 91.46% agreement with the primary labels. Local-order sensitivities changed the neighbor cutoff from rnb=1.5r_{\mathrm{nb}} = 1.5 to 1.4 or 1.6 while retaining coordination range 5–7. The 5–7 tolerance accommodates thermal and cutoff-based coordination fluctuations around locally sixfold environments. Requiring exactly six neighbors retained the original ford≥0.75f_{\mathrm{ord}} \ge0.75 threshold without recalibration and produced zero ordered labels; binary AUROC is therefore undefined, not equal to 0.5. This sensitivity tests the joint definition and threshold rather than a competing calibrated phase definition.

The primary numerical cutoffs, mean⁡(Ψ6)≥0.55\operatorname{mean}(\Psi_6) \ge0.55, mean⁡(ford)≥0.75\operatorname{mean}(f_{\mathrm{ord}}) \ge0.75, and sd⁡(Ψ6)≤0.08\operatorname{sd}(\Psi_6) \le0.08 over the final 15% of stored frames, were used to define the terminal label. They remained fixed during training and the subsequent dense-time comparisons and are not interpreted as fitted thermodynamic phase boundaries.

The sensitivity analysis comprised eleven label variants. One-at-a-time variants used Ψ6\Psi_6 cutoffs 0.50 or 0.60, fordf_{\rm ord} cutoffs 0.70 or 0.80, stability cutoffs 0.10 or 0.06, and final-window fractions 0.10 or 0.20. Relaxed and strict joint variants moved all three structural cutoffs together. Only terminal outcomes were recomputed. All predictor scores were held fixed: stored baseline GNN predictions were rescored without retraining, and the baseline state score retained its original denominators 0.55 and 0.75 rather than adopting each variant’s label thresholds. The same above-chance criterion required the point estimate and both the parent-cluster and hierarchical seed–parent lower limits to exceed 0.5. Figure S2 and Table S3 report the label, neighbor-definition, and discrete state/GNN sensitivities.

Table S3: Terminal-label sensitivity with all predictor scores held fixed. Only the terminal outcomes change; the state score retains denominators 0.55 and 0.75, and stored GNN predictions are not retrained. “Valid cells” is the number of seed–fold–duration cells containing both classes at each phase (60 possible). Bold AUROCs meet the full above-chance criterion: the point estimate and the lower limits of both parent–cluster and hierarchical seed–parent bootstrap intervals exceed 0.5. The state score passes at p100 for 10/11 variants but at p050 for 5/11; the baseline GNN passes at p100 for 8/11. Here ww is the fractional terminal-window length

VariantΨ6\Psi_6 cutofffordf_{\mathrm{ord}} cutoffstability cutoffwwn+n_{+}valid cellsstate p050state p100GNN p100
baseline0.550.750.080.1512260/600.5640.6560.624
relaxed joint0.500.700.100.1526560/600.5640.6570.612
strict joint0.600.800.060.151534/600.3910.5970.497
Ψ6\Psi_6 loose0.500.750.080.1512560/600.5610.6540.626
Ψ6\Psi_6 strict0.600.750.080.1511660/600.5640.6570.633
fordf_{\mathrm{ord}} loose0.550.700.080.1521760/600.5740.6550.605
fordf_{\mathrm{ord}} strict0.550.800.080.152241/600.4460.6210.530
stability loose0.550.750.100.1512960/600.5610.6600.624
stability strict0.550.750.060.159158/600.5450.6400.617
window short0.550.750.080.1012060/600.5360.5980.576
window long0.550.750.080.2011360/600.5610.7100.640

Table S3.

Figure S2. Terminal-label and neighbor-definition sensitivity. (a) Predefined state-score AUROC under the primary and fixed absolute terminal windows. (b) State p050 and p100 estimates for the primary neighbor definition and cutoff variants rnb/σ=1.4r_{\mathrm{nb}}/\sigma= 1.4 and 1.61.6, all retaining the 5–7 coordination range. (c) Label prevalence and agreement with the primary definition. (d) Baseline GNN p100 sensitivity. The strict z=6z = 6 definition keeps the original ordered-fraction threshold without recalibration; it yields no ordered labels and hence has undefined AUROC

Additional orientational observables

The main text defines the ordered mask, its population fraction fordf_{\rm ord}, and the within-mask coherence CordC_{\rm ord}. We additionally define Cmax⁡C_{\max} as the number of particles in the largest connected component of the rnb=1.5r_{\rm nb}=1.5 radius graph induced by the ordered mask; Cmax⁡/NC_{\max}/N was set to zero when the mask was empty. The remaining orientational observables and model inputs were defined as follows. To regularize the local phase, we used

u~i=ψ6,imax⁡(∣ψ6,i∣,10−12).(24)\tilde{u}_i = \frac{\psi_{6,i}}{\max\left(\lvert\psi_{6,i}\rvert,10^{-12}\right)}. \tag*{(24)}

The rotation-invariant local phase-coherence input of the targeted GNN was

c6,i=ni−1∑j∈NiRe⁡(u~iu~j∗).(25)c_{6,i}=n_i^{-1}\sum_{j\in\mathcal{N}_i}\operatorname{Re}\left(\tilde{u}_i\tilde{u}_j^*\right). \tag*{(25)}

with c6,i=0c_{6,i}=0 when ni=0n_i=0. The denominator remains nin_i, so a site with ∣ψ6,i∣<10−12\lvert\psi_{6,i}\rvert<10^{-12} contributes with attenuated magnitude rather than an assigned unit phase.

Let EϵE_\epsilon contain every undirected rnbr_{\rm nb} edge whose endpoints both satisfy ∣ψ6∣>ϵ\lvert\psi_6\rvert>\epsilon, with ϵ=10−12\epsilon=10^{-12}. The neighbor-wide phase coherence was

Cnb=∣Eϵ∣−1∑(i,j)∈EϵRe⁡(u~iu~j∗),(26)C_{\rm nb}=\lvert E_\epsilon\rvert^{-1}\sum_{(i,j)\in E_\epsilon}\operatorname{Re}\left(\tilde{u}_i\tilde{u}_j^*\right), \tag*{(26)}

and was set to zero for an empty edge set. Unlike CordC_{\rm ord}, CnbC_{\rm nb} is not restricted to the ordered mask. Endpoint cutoffs 0.1 and 0.3 and an amplitude-weighted edge average proportional to ∣ψ6,i∣∣ψ6,j∣\lvert\psi_{6,i}\rvert\lvert\psi_{6,j}\rvert supplied definition-level sensitivities.

Because CordC_{\rm ord} normalizes within the ordered set, its population-weighted counterpart was

Aord=N−1∣∑i∈Oψ6,i∣=CordN−1∑i∈O∣ψ6,i∣.(27)A_{\rm ord}=N^{-1}\left\lvert\sum_{i\in O}\psi_{6,i}\right\rvert=C_{\rm ord}N^{-1}\sum_{i\in O}\lvert\psi_{6,i}\rvert. \tag*{(27)}

The ordered-set-size distribution and minimum requirements of 3, 5, or 8 ordered particles were checked before this normalization sensitivity was interpreted. These definitions separately probe population, orientational alignment, and graph connectivity; their sampled discrimination profiles do not by themselves establish a mechanistic chronology.

Feature dependence and current-state analyses

Test-time mean replacement substituted one baseline node channel by its training-fold mean and reported intact minus ablated AUROC. Integrated gradients used the trained ordered-label-logit margin; absolute shares sum to unity within phase. Both quantities depend on the trained reference model and correlated input channels and are not interpreted as causal feature effects.

For state-adjusted residual tests, the null logistic stacker contained a cubic B-spline transform of the predefined state score and categorical quench duration. The transform used four uniformly spaced knots, degree 3, and no bias column and was fitted on training parents only. The full stacker added one GNN logit after clipping its probability to [10−5,1−10−5][10^{-5}, 1 - 10^{-5}]. Both stackers used L2-regularized logistic regression (C=1C = 1, lbfgs, at most 2000 iterations), were fitted to out-of-fold (OOF) predictions from the four training parent folds, and were evaluated on the held-out parent fold. The threshold-defined subset excluded trajectories already satisfying both instantaneous thresholds Ψ6≥0.55\Psi_6 \ge0.55 and ford≥0.75f_{\mathrm{ord}} \ge0.75. A separate parent-grouped cross-fitted spline-ridge analysis tested whether current CordC_{\mathrm{ord}} improved prediction of terminal-window mean Cmax⁡/NC_{\max}/N after current Ψ6\Psi_6, fordf_{\mathrm{ord}}, Cmax⁡/NC_{\max}/N, and τq\tau_q were included.

Feature-dependence estimates are shown in Figure S3; analyses of the current state, the threshold-defined subset, residual information, and future growth are shown in Figure S4.

Dense-time bands and finite-replica correction

The extended grid comprised 21 quench-relative sampled times and 24 post-quench times through tpost=1.20t_{\mathrm{post}} = 1.20, for 45 times and 480×45=21 600480 \times45 = 21\,600 trajectory–time observations. Six observables were compared: raw global Ψ6\Psi_6, the state score, CordC_{\mathrm{ord}}, CnbC_{\mathrm{nb}}, fordf_{\mathrm{ord}}, and Cmax⁡/NC_{\max}/N. Let θ^jk\hat{\theta}_{jk} be the within-duration AUROC for observable jj and sampled time kk, and let θ^jk(b)\hat{\theta}_{jk}^{(b)} be its value in a parent-cluster bootstrap draw bb. Pointwise limits were the 2.5th and 97.5th percentiles across bb.

Two correction families were used. The full 45-time family calibrated over all quench-relative and post-quench times, whereas the 24-time post-quench family calibrated over only the absolute post-quench times. For time set K\mathcal{K}, the observable-wise radius was

rj(K)=Q0.95[max⁡k∈K∣θ^jk(b)−θ^jk∣],(28)r_j(\mathcal{K}) = Q_{0.95}\left[\max_{k \in\mathcal{K}}\left|\hat{\theta}_{jk}^{(b)}-\hat{\theta}_{jk}\right|\right], \tag*{(28)}

and the joint radius was

rjoint(K)=Q0.95[max⁡jmax⁡k∈K∣θ^jk(b)−θ^jk∣].(29)r_{\mathrm{joint}}(\mathcal{K}) = Q_{0.95}\left[\max_j\max_{k \in\mathcal{K}}\left|\hat{\theta}_{jk}^{(b)}-\hat{\theta}_{jk}\right|\right]. \tag*{(29)}

The simultaneous bands were θ^jk±rj\hat{\theta}_{jk} \pm r_j or θ^jk±rjoint\hat{\theta}_{jk} \pm r_{\mathrm{joint}}. The same parent draw was used for all observables and times. The earliest persistently discriminative sampled time was the first post-quench grid point whose lower band remained above 0.5 through tpost=1.20t_{\mathrm{post}} = 1.20. AUROC 0.60 was retained only as a descriptive effect-size reference. All headline intervals used 5000 complete-parent draws. Increasing the dense-band calculation to 100 000 draws left the reported structural conclusion unchanged.

For parent pp and duration cc, let zp,c=kp,c/3z_{p,c} = k_{p,c}/3. Using the within-parent estimator in Eq. (23), the raw duration-specific between-parent component was

V^c=var⁡p(zp,c)−mean⁡p(W^p,c)3.(30)\hat{V}_c = \operatorname{var}_p(z_{p,c}) - \frac{\operatorname{mean}_p(\widehat{W}_{p,c})}{3}. \tag*{(30)}

The four raw components were averaged with equal duration weight before one non-negative projection,

V^=14∑cV^c,V^+=max⁡(V^,0).(31)\hat{V} = \frac{1}{4}\sum_c \hat{V}_c,\qquad\hat{V}_+ = \max(\hat{V},0). \tag*{(31)}

The parent-attributable fraction was

fparent=V^+mean⁡cmean⁡p(W^p,c)+V^+.(32)f_{\mathrm{parent}} = \frac{\hat{V}_+}{\operatorname{mean}_c\operatorname{mean}_p(\widehat{W}_{p,c})+\hat{V}_+}. \tag*{(32)}

Every bootstrap draw repeated this operation in the same order. The raw point was −0.0009-0.0009 with interval [−0.0180,0.0139][-0.0180, 0.0139]; the point after projection was zero and the 97.5th percentile of fparentf_{\mathrm{parent}} was 7.5%.

Figure Figure S5 displays the dense bands, lead-time analysis, ordered-set counts, and finite-replica correction. The principal outcome-resolved structural profiles appear in main-text Figure 3.

Figure S5. Dense sampled times, lead time, ordered-set size, and finite-replica correction. Post-quench portions of the 45-point curves with observable-wise (dotted) and joint observable-by-time (dashed) simultaneous lower bands for (a) primary and (b) fixed-window labels. (c) Ordered-particle counts at tpost=0.55t_{\mathrm{post}} = 0.55; thresholds of 3, 5, and 8 remove no trajectory, and the population-weighted amplitude gives the same earliest persistently discriminative sampled time of 0.55. (d) Minimum actual observation-to-primary-label lead for the predefined phases and the tpost=1.20t_{\mathrm{post}} = 1.20 extension, with the corrected parent-variance estimate inset

First-reciprocal-shell translational-order comparison

The translational comparison used only instantaneous coordinates, number density, and periodic-box geometry. For every allowed reciprocal vector k=(2π/L)(nx,ny)\mathbf{k}=(2\pi/L)(n_x,n_y), we calculated

S(k)=1N∣∑j=1Nexp⁡(ik⋅rj)∣2.(33)S(\mathbf{k})=\frac{1}{N}\left|\sum_{j=1}^{N}\exp(i\mathbf{k}\cdot\mathbf{r}_j)\right|^2 . \tag*{(33)}

The ideal triangular reference at density ρ\rho has a△=[2/(3ρ)]1/2a_{\triangle}=[2/(\sqrt{3}\rho)]^{1/2} and k0=4π/(3a△)k_0=4\pi/(\sqrt{3}a_{\triangle}). At ρ=0.82\rho=0.82 and L=8.834522L=8.834522, a△=1.186664σa_{\triangle}=1.186664\sigma, k0=6.113944/σk_0=6.113944/\sigma, and the fundamental reciprocal spacing is 2π/L=0.711208/σ2\pi/L=0.711208/\sigma. We retained one vector from each inversion pair in the annulus ∣∣k∣−k0∣≤π/L||\mathbf{k}|-k_0|\leq\pi/L. This produced 24 vectors on squared integer shells 68, 72, 73, 74, 80, 81, and 82.

For each frame, the primary score sTs_T was the mean of the three largest annular S(k)S(\mathbf{k}) values. Three was fixed to match the number of inversion-unique directions in an ideal triangular first shell, without requiring those directions to be commensurate with the square box. The mean over all 24 annular values was the non-maximizing sensitivity. Both scores are permutation invariant and exactly invariant to a common translation on the allowed periodic reciprocal grid. Neither is exactly invariant to an arbitrary rotation on the finite square reciprocal grid; selecting over the full annulus and retaining its mean provide complementary directional sensitivities. They contain no bond angles, ordered-mask membership, label values, or fitted parameters. The primary analysis used both labels at all 45 sampled times and four fixed leads. The timestep comparison used both timestep settings at the 24 post-quench times and four fixed leads. Dense intervals used one two-score-by-time simultaneous family. Fixed-lead results used an orientation-plus-translation 4×44\times4 family containing raw Ψ6\Psi_6, the state score, CordC_{\rm ord}, and sTs_T. Broader primary families added stored energy and the annular mean; the timestep-comparison family contained the three orientational scores and both translational scores. All uncertainty used 5000 complete-parent draws and equal duration weighting.

Observable association, duration, and fixed-lead analyses

Associations among current observables and with terminal structural quantities used a duration-adjusted Spearman coefficient. Ranks were calculated separately within each τq\tau_q, centered within that duration, and pooled only after centering. Parent clusters were resampled for 5000 percentile draws. This quantity is a finite-window association; it is not an equilibrium autocorrelation function or relaxation-time estimate. Terminal targets were the terminal-window mean of raw Ψ6\Psi_6 for raw Ψ6\Psi_6 and CordC_{\rm ord}, the corresponding structural state score, and the matched terminal structural quantity for the other observables. Selected correlations and quench-duration-specific fixed-lead results are reported below and in Table S7.

Observable paired with sTs_TSpearman ρ\rhoParent-cluster 95% interval
(a) Translational–orientational correlations
Raw Ψ6\Psi_60.872[0.852, 0.891]
State score0.838[0.812, 0.862]
CordC_{\mathrm{ord}}0.855[0.831, 0.877]
τq\tau_qObservableAUROCPointwise 95% interval16-cell lower64-cell lower
(b) Quench-duration-specific AUROC at lead 0.50
0.15Raw Ψ6\Psi_60.714[0.585, 0.823]0.5370.522
State score0.693[0.564, 0.804]0.5210.501
CordC_{\mathrm{ord}}0.685[0.561, 0.788]0.5140.493
sTs_T0.746[0.636, 0.842]0.5730.554
0.60Raw Ψ6\Psi_60.703[0.595, 0.805]0.5250.510
State score0.699[0.594, 0.802]0.5260.506
CordC_{\mathrm{ord}}0.705[0.603, 0.802]0.5340.513
sTs_T0.726[0.629, 0.823]0.5520.534
1.50Raw Ψ6\Psi_60.619[0.489, 0.743]0.4420.427
State score0.618[0.504, 0.724]0.4450.425
CordC_{\mathrm{ord}}0.594[0.470, 0.716]0.4230.402
sTs_T0.567[0.443, 0.690]0.3930.375
3.00Raw Ψ6\Psi_60.605[0.490, 0.715]0.4280.413
State score0.580[0.462, 0.702]0.4070.387
CordC_{\mathrm{ord}}0.598[0.484, 0.706]0.4270.405
sTs_T0.651[0.548, 0.755]0.4770.459

Table S7. Table S7: Translational–orientational coupling and duration-resolved results at lead 0.50. Part (a) gives duration-adjusted Spearman correlations at tpost=0.55t_{\mathrm{post}} = 0.55 with parent-cluster percentile intervals. In part (b), “16-cell lower” is the simultaneous lower limit over four durations and four leads for one observable; “64-cell lower” is the simultaneous lower limit over all four observables, durations, and leads. The same 5000 complete-parent resamples were shared within each family

At tpost=0.55t_{\rm post}=0.55, the separately defined translational score sTs_T had duration-adjusted Spearman correlations of 0.872 with raw Ψ6\Psi_6, 0.838 with the state score, and 0.855 with CordC_{\rm ord}. Their parent-cluster intervals were [0.852, 0.891], [0.812, 0.862], and [0.831, 0.877], respectively (Table S7). These associations show that the translational comparison probes a strongly coupled aspect of the developing collective order rather than a statistically independent signal.

Because the primary last-15% label window has a duration-dependent absolute width, its mean frame counts and starts were tabulated separately for each τq\tau_q. A second label used a fixed final window of 0.50 time units. It changed ordered-label counts from (32, 33, 30, 27) to (28, 31, 32, 30) over τq=(0.15,0.60,1.50,3.00)\tau_q=(0.15,0.60,1.50,3.00), with respective agreements 91.7%, 95.0%, 98.3%, and 80.8%. No duration was normalized by an independently measured equilibrium structural-relaxation time.

The fixed-lead analysis was also repeated relative to the fixed-window start, using the fixed-window outcome. At requested lead 1.0, raw Ψ6\Psi_6, the state score, and CordC_{\rm ord} had equal-duration AUROCs 0.551, 0.548, and 0.556. Their lower limits under the energy-expanded four-observable-by-four-lead family were 0.461, 0.458, and 0.466. At lead 0.50 the corresponding estimates were 0.597, 0.607, and 0.596, with joint lower limits 0.507, 0.517, and 0.507. Thus the fixed-window-label analysis shows a rising fixed-lead profile, but its lead-1.0 lower limits include 0.5. At lead 0.75, the three displayed lower limits round to 0.500, but the unrounded values lie on both sides of chance; that row is not a joint three-observable pass.

For the fixed-lead analysis, the target time for each trajectory was its primary terminal-window start minus ℓ\ell, with ℓ∈{0.25,0.50,0.75,1.00}\ell\in\{0.25, 0.50, 0.75, 1.00\}. The last stored frame not later than that target was selected. Every selected frame strictly preceded the terminal window; the minimum realized separation was 0.256. Within-duration AUROCs and parent-cluster percentile intervals were then recomputed at each lead. Every bootstrap draw used one common resample of the 40 parent clusters for all leads and observables. In addition to pointwise intervals, an observable-wise band controlled the four leads of one observable, and the primary joint band controlled all 3×43 \times4 estimates for raw Ψ6\Psi_6, the state score, and CordC_{\rm ord}. As a broader sensitivity family, we also controlled all 6×46 \times4 estimates after adding CnbC_{\rm nb}, fordf_{\rm ord}, and Cmax⁡/NC_{\max}/N. Each band used the 95th percentile of the bootstrap maximum absolute deviation from the corresponding original-sample estimate. These families were specified before their corrected intervals were inspected.

For the non-orientational comparison, the total potential energy UU stored at every trajectory frame was converted to sE=−U/(Nε)s_E=-U/(N\varepsilon). The sign was fixed so that a more negative potential energy produces a larger ranking score. This scalar contains no bond angle, ordered-mask, or fitted-model input. It was evaluated at the same 45 dense times under both terminal labels. Pointwise parent-cluster intervals and simultaneous lower bands were calculated over the full 45-time and 24-time post-quench families. We also evaluated the four fixed leads relative to each label-window start. Those intervals used one joint family over raw Ψ6\Psi_6, the state score, CordC_{\rm ord}, and sEs_E at all four leads. This comparison used 5000 parent-cluster draws and no trajectory generation or model fitting.

For the primary label at lead 1.0, sEs_E had AUROC 0.489 with joint 4×44 \times4 interval [0.403, 0.574]. Raw Ψ6\Psi_6, the state score, and CordC_{\rm ord} had lower limits 0.513, 0.507, and 0.519 in the expanded family. At leads 0.50 and 0.25, however, sEs_E had AUROCs 0.625 and 0.601, with joint intervals [0.539, 0.710] and [0.516, 0.686]; both lower limits exceeded 0.5 after the same simultaneous correction. Its lead-0.75 interval, [0.476, 0.647], crossed 0.5. The primary-label energy curve rose to 0.591 at tpost=1.20t_{\rm post}=1.20, where the full 45-time lower band was 0.494. Its lower band therefore did not remain above 0.5 in the full 45-time analysis; in the 24-time post-quench analysis this occurred only at the final 1.20 grid point. Under the fixed-window label, neither lower band remained above 0.5 through 1.20. Figure S14 and Table S4 report all estimates.

Terminal labelLeadRaw Ψ6\Psi_{6}StateCordC_{\mathrm{ord}}−U/(Nε)-U/(N\varepsilon)
Primary fractional1.000.599 [0.513, 0.684]0.593 [0.507, 0.678]0.605 [0.519, 0.690]0.489 [0.403, 0.574]
Primary fractional0.750.647 [0.562, 0.732]0.628 [0.542, 0.713]0.655 [0.570, 0.740]0.562 [0.476, 0.647]
Primary fractional0.500.660 [0.575, 0.746]0.647 [0.562, 0.733]0.646 [0.560, 0.731]0.625 [0.539, 0.710]
Primary fractional0.250.685 [0.599, 0.770]0.655 [0.570, 0.740]0.701 [0.616, 0.787]0.601 [0.516, 0.686]
Fixed final 0.501.000.551 [0.461, 0.641]0.548 [0.458, 0.638]0.556 [0.466, 0.646]0.518 [0.428, 0.607]
Fixed final 0.500.750.590 [0.500, 0.680]0.590 [0.500, 0.679]0.590 [0.500, 0.680]0.557 [0.467, 0.646]
Fixed final 0.500.500.597 [0.507, 0.687]0.607 [0.517, 0.697]0.596 [0.507, 0.686]0.557 [0.468, 0.647]
Fixed final 0.500.250.692 [0.602, 0.782]0.683 [0.593, 0.773]0.690 [0.600, 0.780]0.585 [0.495, 0.675]

Table S4. Potential-energy and fixed-window fixed-lead comparison. Each entry is the equal-duration AUROC followed by its joint 4×44 \times4 parent-cluster interval in brackets. The family contains raw Ψ6\Psi_{6}, the state score, CordC_{\mathrm{ord}}, and −U/(Nε)-U/(N\varepsilon) at four leads. All intervals use 5000 parent-cluster draws. Lower limits are compared with 0.5 before rounding; a displayed lower limit of 0.500 need not exceed 0.5 on the unrounded scale

The primary translational-order score changed the specificity interpretation. In the primary analysis, sTs_T had AUROC 0.593 at lead 1.0; its orientation-plus-translation 4×44 \times4 interval was [0.511, 0.675], and its broader all-six interval was [0.502, 0.684]. The fixed-window-label value was 0.560 [0.475, 0.645] in the 4×44 \times4 family, and the fine-step value was 0.539 [0.454, 0.625]; both intervals included 0.5 at lead 1.0. At lead 0.50, the respective values were 0.673 [0.591, 0.754], 0.616 [0.531, 0.700], and 0.627 [0.542, 0.713], all above chance after the same joint correction. They also had lower limits above 0.5 in each analysis’s broader family. The annular mean rose in all three analyses, but its broadest fixed-window interval at lead 0.50 had an unrounded lower limit of 0.4990 and therefore did not pass.

For sTs_T, the earliest persistently discriminative sampled times under the dense two-score joint bands were 0.55 for the primary analysis and 0.85 for both the fixed-window-label and fine-step analyses. The corresponding annular-mean values were 0.55, 1.10, and 0.85. No paired fine-step-minus-primary dense or fixed-lead translational interval excluded zero after simultaneous correction. Figure S15 and Table S5 report the complete summary. The common top-three result at lead 0.50 supports structural-order discrimination, whereas the variation at lead 1.0 and between translational definitions precludes an orientation-specific or uniquely timed claim.

AnalysisTranslational-order scoreLead 1.00Lead 0.50Dense joint sampled time
Primary analysisTop-three sTs_T0.593 [0.502, 0.684]0.673 [0.582, 0.763]0.55
Primary analysisAnnular mean0.583 [0.492, 0.674]0.636 [0.545, 0.727]0.55
Fixed-window-label analysisTop-three sTs_T0.560 [0.466, 0.654]0.616 [0.521, 0.710]0.85
Fixed-window-label analysisAnnular mean0.568 [0.474, 0.663]0.593 [0.499, 0.688]1.10
Fine-step analysisTop-three sTs_T0.539 [0.450, 0.629]0.627 [0.538, 0.717]0.85
Fine-step analysisAnnular mean0.577 [0.488, 0.667]0.610 [0.521, 0.700]0.85

Table S5. First-reciprocal-shell translational-order comparison. Fixed-lead entries are equal-duration AUROCs followed by the broadest available joint interval in brackets. For the two primary-timestep rows, the family contains raw Ψ6\Psi_6, the state score, CordC_{\mathrm{ord}}, −U/(Nε)-U/(N\varepsilon), sTs_T, and the annular mean at four leads (6×46 \times4). For the fine-step rows it contains the three orientational and two translational scores at four leads (5×45 \times4). The final column is the earliest post-quench sampled time whose lower band remains above 0.5 through 1.20 in the joint two-translational-score dense family. A value is a sampled time, not an exact onset. All intervals use 5000 parent-cluster draws

The quench-duration-specific comparison kept the four durations as separate cells. It reported pointwise parent-cluster intervals, one simultaneous family over the 4×44 \times4 duration-by-lead cells of each observable, and a joint family over all 3×4×43 \times4 \times4 primary orientational cells. At requested lead 1.0, the selected post-quench times were 0.642, 0.576, 0.460, and 0.224 for τq∈{0.15,0.60,1.50,3.00}\tau_q \in\{0.15, 0.60, 1.50, 3.00\}. All three pointwise lower limits exceeded 0.5 only for τq=0.15\tau_q=0.15; no duration was thus property under the observable-wise 16-cell correction or the stricter 48-cell correction (Figure S12). Thus fixed lead holds distance to the label window approximately constant but does not hold post-quench age constant.

We then added sTs_T to form the four-measure duration-resolved family reported in Table S7. At lead 0.50, both the pointwise and observable-wise 16-cell lower limits exceeded 0.5 for all four measures at τq=0.15\tau_q=0.15 and 0.60. For τq=1.50\tau_q=1.50 and 3.00, one pointwise lower limit exceeded 0.5 at each duration, but none of the four observable-wise lower limits did so. Under the joint 64-cell family, three of four lower limits exceeded 0.5 at τq=0.15\tau_q = 0.15, all four did so at 0.60, and none did so at 1.50 or 3.00. The equal-duration headline therefore summarizes a signal that is resolved most consistently for the two shorter quench durations in the present sample.

Definition and fixed-cohort sensitivity analyses

The primary neighbor-phase coherence normalized each nonzero local complex order parameter before taking the real phase product over neighbor edges. Its endpoint-amplitude cutoff was varied from 10−1210^{-12} to 0.1 or 0.3. An amplitude-weighted alternative used edge weights ∣ψ6,i∣∣ψ6,j∣|\psi_{6,i}||\psi_{6,j}|. Point curves and 24-time post-quench simultaneous bands were recomputed for all four definitions from the same trajectories.

The cohort analysis fixed membership once at p050 by excluding trajectories that already satisfied both instantaneous structural thresholds. The resulting 393 trajectories, including 95 ordered outcomes, were followed unchanged at p050 and p100. State and baseline GNN scores used their original folds and predictions. Both parent-cluster and hierarchical seed–parent bootstrap intervals were applied, so this analysis does not condition on a different subset at p100. At p050, the fixed and phase-specific definitions select exactly the same trajectories and state statistic. Table S2 therefore reuses the phase-specific state interval instead of presenting the immaterial 10−310^{-3}-level difference from an independent bootstrap realization. In a separate q25 comparison, the apparent below-chance state AUROC was not distinguished from chance because its upper parent-cluster interval exceeded 0.5.

Table S2: Summary of primary and timestep-sensitivity results. Intervals are 95% parent-cluster bootstrap intervals unless a simultaneous family, a second uncertainty level, or a paired contrast is explicitly named

AnalysisSampled timeEstimateParent interval or sampled timeSecond-level or paired resultInterpretation
Mixed replica setsterminal92/160—Bernoulli expectation: 90.8matches duration-wise Bernoulli expectation
Corrected between-parent varianceall τq\tau_q−0.0009[−0.0180, 0.0139]fparentf_{\mathrm{parent}} 97.5th percentile: 7.5% compatible with zero at m=3m = 3
Raw Ψ6\Psi_6/state/CordC_{\mathrm{ord}}: first times45 times0.55/0.60/0.55joint: 0.60/0.80/0.6024-time: 0.55/0.55/0.55depends on multiplicity family
Cnb/ford/Cmax⁡/NC_{\mathrm{nb}}/f_{\mathrm{ord}}/C_{\max}/N: first times45 times0.80/0.85/1.00joint: 0.80/1.00/1.0024-time: 0.55/0.80/0.80Cmax⁡/NC_{\max}/N observable-wise entries use 100 000 draws; others use 5000
Raw Ψ6\Psi_6/state/CordC_{\mathrm{ord}}fixed lead 1.00.599/0.593/0.605joint 3 × 4: [0.519, 0.678] [0.513, 0.672] [0.525, 0.684]joint 6 × 4 lower limits: 0.508/0.502/0.514aggregate lower limits exceed 0.5 after equal duration weighting; not established within every duration
Raw Ψ6\Psi_6/state/CordC_{\mathrm{ord}} versus terminal six-coordinated fractionfixed lead 1.00.168/0.156/0.202joint 3 × 4: [0.060, 0.276] [0.048, 0.265] [0.094, 0.310]duration-adjusted Spearman ρ\rhoexploratory continuous endpoint
Fine-step raw Ψ6\Psi_6/state/CordC_{\mathrm{ord}}fixed lead 1.00.546/0.542/0.544joint 3 × 4: [0.464, 0.627] [0.460, 0.623] [0.463, 0.626]paired fine-step–primary: all corrected intervals include zerointervals include 0.5 at lead 1.0
Fine-step raw Ψ6\Psi_6/state/CordC_{\mathrm{ord}}fixed lead 0.750.614/0.615/0.605joint 3 × 4: [0.533, 0.696] [0.533, 0.696] [0.523, 0.686]paired fine-step–primary intervals include zerolower limits exceed 0.5 at lead 0.75
State scorep050 full cohort0.564[0.513, 0.605][0.504, 0.615]both lower limits exceed 0.5; repeat is the fold partition
Baseline GNNp1000.624[0.560, 0.684][0.525, 0.710]both lower limits exceed 0.5
Duration-conditioned GNNp1000.653[0.592, 0.707][0.579, 0.719]hierarchical seed–parent contrast includes zero
Targeted GNNp0500.588[0.529, 0.635][0.507, 0.653]both lower limits exceed 0.5
State/GNN, cohort defined at p050p0500.551[0.486, 0.596][0.469, 0.611]state interval reused from the identical phase-specific cohort; intervals include 0.5
State/GNN, cohort defined at p050p0500.500[0.450, 0.535][0.408, 0.578]
State/GNN, cohort defined at p050p1000.657[0.595, 0.722][0.581, 0.734]all lower limits exceed 0.5
State/GNN, cohort defined at p050p1000.631[0.558, 0.699][0.509, 0.733]
Finite t∗t^* among non-ordered outcomesterminal219/358—207/219 do not satisfy terminal fordf_{\mathrm{ord}} criterioncrossing is nonabsorbing; the fordf_{\mathrm{ord}} criterion is the most frequently unsatisfied marginal criterion

Table S2.

Finally, each terminal-label variant was checked at the actual seed–fold–duration cell level. A cell was valid only if both labels were present. Because strict variants retained as few as 34/60 or 41/60 cells, their within-duration AUROCs are evaluated on smaller domains and are not directly precision-matched to the 60/60-cell variants. These sensitivity analyses are summarized in Figure S9 and Table S3.

First-crossing time and terminal structure

The auxiliary t∗t^* retained its training-time definition: the first stored time at which a trailing window of 7.5% of the trajectory, with a minimum of two frames, simultaneously met mean Ψ6≥0.55\Psi_6 \ge0.55 and mean ford≥0.75f_{\rm ord} \ge0.75. It was analyzed as a first crossing without an absorbing-state assumption. A finite t∗t^* followed by a terminal non-ordered label was counted as a transient crossing. No right-censored survival model was fitted because the physical event was not shown to be absorbing.

For every transient crossing, the three terminal criteria were evaluated separately: terminal-window mean Ψ6<0.55\Psi_6 < 0.55, terminal-window mean ford<0.75f_{\rm ord} < 0.75, and terminal-window standard deviation of Ψ6>0.08\Psi_6 > 0.08. Marginal failure fractions count a trajectory in every failed criterion, whereas the exclusive breakdown records the exact combination. Parent-cluster draws resampled parent clusters and recomputed each fraction among the transient crossings retained in that draw. The marginal categories therefore need not sum to one.

Terminal pair structure was averaged over the same primary label window. For particle pairs separated by rr, g(r)g(r) used ideal-gas shell normalization in the periodic square cell, and the conditional orientational curve was

G6(r)=⟨Re⁡[ψ6,iψ6,j∗] | rij in the bin around r⟩(34)G_6(r) = \left\langle\operatorname{Re}\left[\psi_{6,i}\psi_{6,j}^{*}\right] \,\middle|\, r_{ij}\ \text{in the bin around } r \right\rangle \tag*{(34)}

Both curves used 48 bins from zero to L/2L/2. Coordination fractions for z=5,6,7z = 5, 6, 7, and values outside 5–7 were calculated with rnb=1.5r_{\rm nb} = 1.5. Outcome-specific means and 95% intervals used complete-parent resampling. We also averaged each coordination fraction within parent and outcome. Forty parents contributed a non-ordered outcome, 39 an ordered outcome, and 39 contributed both. For the paired set, each bootstrap draw resampled parents and averaged the within-parent ordered-minus-non-ordered difference (Figure S13). The terminal fraction of six-coordinated particles was the mean, over the primary terminal window, of N−1∑i1(ni=6)N^{-1}\sum_i 1(n_i = 6). It is not a direct term in the binary terminal label, although it is physically related to the same ordering process. At each fixed lead, its association with current raw Ψ6\Psi_6, state score, and CordC_{\rm ord} was quantified by the duration-adjusted Spearman coefficient defined above. One common parent resample was used over the 3×43 \times4 association family, and simultaneous intervals were constructed from its maximum absolute bootstrap deviation. We treat this endpoint as an exploratory continuous structural check, not as an independent mechanism or a redefined phase label. The terminal snapshots in Figure S10 show the three replicas from one mixed replica set at their final stored frame.

Figure S10. Terminal snapshots from one mixed replica set. The three replicas share the same stored parent and τq=0.15\tau_q = 0.15 but use independent Brownian-noise realizations. Color gives local ∣ψ6,i∣|\psi_{6,i}|; red particle outlines mark coordination outside 5–7. Each panel uses its last stored terminal frame and spans L=8.8345σL = 8.8345\sigma

Numerical and implementation checks

The high-temperature initializer comparison followed eight weakly perturbed square states and eight random-sequential states through time 8 at T=1.8T = 1.8. Both families relaxed without runaway energies, although the random-sequential states began at substantially higher energy (Figure S6).

Figure S6. High-temperature burn-in and initializer comparison. (a) Global Ψ4\Psi_4, (b) global Ψ6\Psi_6, (c) locally ordered fraction, (d) energy per particle, and (e) identity mean-squared displacement through time 8. Points and bars show family means and standard errors across eight seeds. (f) Radial distribution function and (g) low-wave-number structure factor at times 0, 2, and 8 for jittered-square and random-sequential initializations. Shading or error bars show standard errors across seeds. Random-sequential states begin at higher but finite energy and relax rapidly; no step crosses the predefined E/N=100E/N = 100 numerical-instability threshold. Energy, displacement, distance, and wave number are plotted as U/(Nε)U/(N\varepsilon), MSD/σ2\mathrm{MSD}/\sigma^2, r/σr/\sigma, and kσk\sigma, respectively

The fine-step ensemble repeated the 40-parent, four-duration, three-replica design with Δt=2×10−5\Delta t = 2 \times10^{-5} instead of 5×10−55 \times10^{-5}. Both ensembles stored frames every 0.016 time units and evaluated the same force and energy functions. Direct evaluation on identical stored coordinates agreed to floating-point precision. Because the timestep changes the Brownian-noise grid, the trajectory comparison is distributional rather than pathwise; two step sizes also do not constitute a continuum extrapolation.

Post-quench structural discrimination was recalculated over 24 times from tpost=0.05t_{\mathrm{post}} = 0.05 to 1.20 and at four fixed leads. Each ensemble used its own terminal outcomes and terminal-window starts. In the fine-step ensemble, raw Ψ6\Psi_6, the state score, and Cord had AUROCs 0.546, 0.542, and 0.544 at lead 1.0, with simultaneous intervals crossing 0.5. At lead 0.75 the estimates were 0.614, 0.615, and 0.605, with lower limits 0.533, 0.533, and 0.523. Paired fine-step-minus-primary intervals included zero throughout the corrected families. The comparison therefore reproduces the development of structural outcome association without defining a continuum onset (Figure S7 and Figure S9). Exact software requirements and random-number settings are provided with the accompanying reproducibility package.

Figure S7. Parent-matched timestep comparison. (a) Parent-cluster intervals for paired fine-step-minus-primary changes in terminal and branching quantities. The ordered-label rate and the two terminal structural means show small paired differences; the branching contrasts are shown alongside them. (b–d) Dense post-quench AUROC with pointwise intervals for CordC_{\mathrm{ord}}, state score, and Cmax⁡/NC_{\max}/N. Vertical lines mark the earliest persistently discriminative sampled times under the observable-wise bands. The primary and fine-step ensembles use different timesteps and Brownian-noise grids. Their different sampled times do not establish temporal convergence

Supplementary tables

Supplementary figures

Unless stated otherwise, all time quantities in the supplementary figures are reported in the reduced units defined in the main text.

References

  1. [1]Whitesides, G.M., Grzybowski, B.: Self-assembly at all scales. Science 295(5564), 2418–2421 (2002) https://doi.org/10.1126/science.1070821
  2. [2]Glotzer, S.C., Solomon, M.J.: Anisotropy of building blocks and their assembly into complex structures. Nature Materials 6(8), 557–562 (2007) https://doi.org/10.1038/nmat1949
  3. [3]Whitelam, S., Jack, R.L.: The statistical mechanics of dynamic pathways to self-assembly. Annual Review of Physical Chemistry 66, 143–163 (2015) https://doi.org/10.1146/annurev-physchem-040214-121215
  4. [4]Grant, J., Jack, R.L., Whitelam, S.: Analyzing mechanisms and microscopic reversibility of self-assembly. The Journal of Chemical Physics 135(21), 214505 (2011) https://doi.org/10.1063/1.3662140
  5. [5]Auer, S., Frenkel, D.: Prediction of absolute crystal-nucleation rate in hard-sphere colloids. Nature 409(6823), 1020–1023 (2001) https://doi.org/10.1038/35059035
  6. [6]Gasser, U., Weeks, E.R., Schofield, A., Pusey, P.N., Weitz, D.A.: Real-space imaging of nucleation and growth in colloidal crystallization. Science 292(5515), 258–262 (2001) https://doi.org/10.1126/science.1058457
  7. [7]Royall, C.P., Charbonneau, P., Dijkstra, M., Russo, J., Smallenburg, F., Speck, T., Valeriani, C.: Colloidal hard spheres: Triumphs, challenges, and mysteries. Reviews of Modern Physics 96(4), 045003 (2024) https://doi.org/10.1103/RevModPhys.96.045003
  8. [8]Widmer-Cooper, A., Harrowell, P., Fynewever, H.: How reproducible are dynamic heterogeneities in a supercooled liquid? Physical Review Letters 93(13), 135701 (2004) https://doi.org/10.1103/PhysRevLett.93.135701
  9. [9]Bolhuis, P.G., Chandler, D., Dellago, C., Geissler, P.L.: Transition path sampling: Throwing ropes over rough mountain passes, in the dark. Annual Review of Physical Chemistry 53, 291–318 (2002) https://doi.org/10.1146/annurev.physchem.53.082301.113146
  10. [10]E, W., Vanden-Eijnden, E.: Towards a theory of transition paths. Journal of Statistical Physics 123(3-4), 503–523 (2006) https://doi.org/10.1007/s10955-005-9003-9
  11. [11]Halperin, B.I., Nelson, D.R.: Theory of two-dimensional melting. Physical Review Letters 41(2), 121–124 (1978) https://doi.org/10.1103/PhysRevLett.41.121
  12. [12]Zahn, K., Lenke, R., Maret, G.: Two-stage melting of paramagnetic colloidal crystals in two dimensions. Physical Review Letters 82(13), 2721–2724 (1999) https://doi.org/10.1103/PhysRevLett.82.2721
  13. [13]Bernard, E.P., Krauth, W.: Two-step melting in two dimensions: First-order liquid-hexatic transition. Physical Review Letters 107(15), 155704 (2011) https://doi.org/10.1103/PhysRevLett.107.155704
  14. [14]Kapfer, S.C., Krauth, W.: Two-dimensional melting: From liquid-hexatic coexistence to continuous transitions. Physical Review Letters 114(3), 035702 (2015) https://doi.org/10.1103/PhysRevLett.114.035702
  15. [15]Kawasaki, T., Tanaka, H.: Formation of a crystal nucleus from liquid. Proceedings of the National Academy of Sciences 107(32), 14036–14041 (2010) https://doi.org/10.1073/pnas.1001040107
  16. [16]Russo, J., Tanaka, H.: The microscopic pathway to crystallization in supercooled liquids. Scientific Reports 2, 505 (2012) https://doi.org/10.1038/srep00505
  17. [17]Mickel, W., Kapfer, S.C., Schröder-Turk, G.E., Mecke, K.: Shortcomings of the bond orientational order parameters for the analysis of disordered particulate matter. The Journal of Chemical Physics 138(4), 044501 (2013) https://doi.org/10.1063/1.4774084
  18. [18]Li, Y.-W., Pica Ciamarra, M.: Phase behavior of Lennard–Jones particles in two dimensions. Physical Review E 102(6), 062101 (2020) https://doi.org/10.1103/PhysRevE.102.062101
  19. [19]Ermak, D.L., McCammon, J.A.: Brownian dynamics with hydrodynamic interactions. The Journal of Chemical Physics 69(4), 1352–1360 (1978) https://doi.org/10.1063/1.436761
  20. [20]Brańka, A.C., Heyes, D.M.: Algorithms for brownian dynamics simulation. Physical Review E 58(2), 2611–2615 (1998) https://doi.org/10.1103/PhysRevE.58.2611
  21. [21]Steinhardt, P.J., Nelson, D.R., Ronchetti, M.: Bond-orientational order in liquids and glasses. Physical Review B 28(2), 784–805 (1983) https://doi.org/10.1103/PhysRevB.28.784
  22. [22]Hamilton, W.L., Ying, R., Leskovec, J.: Inductive representation learning on large graphs. In: Advances in Neural Information Processing Systems, vol. 30, pp. 1024–1034 (2017). https://proceedings.neurips.cc/paper/2017/hash/5dd9db5e033da9c6fb5ba83c7a7ebea9-Abstract.htmlarxiv.org/abs/1706.02216
  23. [23]Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735–1780 (1997) https://doi.org/10.1162/neco.1997.9.8.1735
  24. [24]Roberts, D.R., Bahn, V., Ciuti, S., Boyce, M.S., Elith, J., Guillera-Arroita, G., Hauenstein, S., Lahoz-Monfort, J.J., Schröder, B., Thuiller, W., Wharton, D.I., Wintle, B.A., Hartig, F., Dormann, C.F.: Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography 40(8), 913–929 (2017) https://doi.org/10.1111/ecog.02881
  25. [25]Varma, S., Simon, R.: Bias in error estimation when using cross-validation for model selection. BMC Bioinformatics 7, 91 (2006) https://doi.org/10.1186/1471-2105-7-91
  26. [26]Efron, B.: Bootstrap methods: Another look at the jackknife. The Annals of Statistics 7(1), 1–26 (1979) https://doi.org/10.1214/aos/1176344552
  27. [27]Field, C.A., Welsh, A.H.: Bootstrapping clustered data. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 69(3), 369–390 (2007) https://doi.org/10.1111/j.1467-9868.2007.00593.x
  28. [28]Chen, H., Roux, B., Chipot, C.: Discovering reaction pathways, slow variables, and committor probabilities with machine learning. Journal of Chemical Theory and Computation 19(14), 4414–4426 (2023) https://doi.org/10.1021/acs.jctc.3c00028
  29. [29]Contreras Arredondo, S., Tang, C., Talmazan, R.A., Megías, A., Chen, C.G., Chipot, C.: Learning the committor without collective variables. Nature Computational Science 6, 350–357 (2026) https://doi.org/10.1038/s43588-026-00958-2
  30. [30]Lavergne, F.A., Aarts, D.G.A.L., Dullens, R.P.A.: Anomalous grain growth in a polycrystalline monolayer of colloidal hard spheres. Physical Review X 7(4), 041064 (2017) https://doi.org/10.1103/PhysRevX.7.041064

Paper details

vlnc:W_6yjs4gne
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents