Structural Predictors of Glassy Mobility as Ancestry-Biased Tags

When a liquid is cooled toward the glass transition, its particles stop moving uniformly. Some neighborhoods rearrange freely while others stay frozen, even though snapshots of the structure look almost the same everywhere. Physicists have built many "structure-based predictors" that score each particle's local arrangement and use the score to predict how mobile that particle will be. These predictors work well in some model liquids and fail in others. The pattern resembles a familiar failure in human genetics. Risk scores built from European cohorts lose accuracy in other ancestries, because the genetic markers they use only tag the true causal variants, and the tag-to-cause correlation differs between populations.

We propose to take that analogy literally as a study design, not as a metaphor for the genome. The hypothesis is that local structure is a tag for the actual cause of mobility, which is the height of local energy barriers on the liquid's energy landscape. The tag-to-barrier link changes from one interaction potential to another. The barrier-to-mobility link should be much more stable. If that holds, then landscape-level predictors should transfer across liquids where structural ones do not. The way the structural predictors fail should also be predictable from how the tag-to-barrier link changes.

Core Claim

Structure–dynamics correlation and its link to fragility and dynamic heterogeneity correlated a density-functional "caging" structural order parameter (SOP) with isoconfigurational mobility across several model liquids. They found the correlation strong in attraction-dominated (enthalpic) liquids and collapsing in purely repulsive (entropic) ones. They explained this through thermodynamic regime. Our claim is different and testable. The collapse is mostly a change in how well structure tags local barriers, not a change in whether local barriers control mobility.

The prediction comes from Local vs. Cooperative: Unraveling Glass Transition Mechanisms with SEER. That work extracted about 200 elementary rearrangements ("excitations") per configuration and showed that local barrier changes track the change in activation energy, across five orders of magnitude of relaxation time. It did this only in one repulsive polydisperse model. If barriers are the stable cause, a mobility predictor built on per-particle barriers and calibrated on one liquid should lose little accuracy on another. A predictor built on the SOP should lose accuracy roughly in proportion to how much the SOP–barrier correlation drops.

There is a second, conditional payoff. Suppose the barrier–mobility link itself fails to transfer, most likely into attractive liquids. That would be direct evidence that something beyond local barriers governs mobility there. Such evidence would bear on the local-versus-cooperative dispute that SEER left open.

Research Program

The landscape side reuses inherent-structure mapping. An inherent structure is the local energy minimum reached by quenching an instantaneous configuration with energy minimization. For each particle we compute two landscape features. The first is local basin depth, following Relation between local diffusivity and local inherent structures in the Kob-Andersen Lennard-Jones model, which found mobility to be a steep nonlinear function of that depth. The second is a local barrier: the lowest barrier among extracted excitations in which the particle participates.

SEER's roughly 200 excitations per 2000-particle configuration will not cover every particle. We will therefore seed its path-penalized search locally around under-covered particles. We will also report barrier coverage as a quantity in its own right, since uncovered particles bias any correlation.

The dynamic target is isoconfigurational propensity: each particle's mean squared displacement over the relaxation time, averaged across many runs from identical positions with fresh velocities. We use 100 runs, as Sharma et al. did.

The genetics side supplies the evaluation logic. The two-step reasoning of Mendelian randomization separates the instrument–exposure link from the exposure–outcome link, as in Causal associations between risk factors and common diseases inferred from GWAS summary data. Here the instrument is structure, the exposure is barrier height, and the outcome is mobility. For a predictor $f$ calibrated on a source liquid $A$ and tested on a target liquid $B$, we define its portability as

$$P_{A\to B}(f) = \frac{\rho_B(f, p)}{\rho_A(f, p)},$$

where $\rho_X(f,p)$ is the Spearman rank correlation between predictor and propensity $p$ in liquid $X$. This mirrors how risk-score portability is reported across ancestries.

Under the tag hypothesis, and if structure affects mobility only through barriers, the SOP's portability should be close to the ratio of its SOP–barrier correlations in $B$ and $A$. The barrier predictor's portability should stay near one. This mediation check is the central analysis. It is informative even when absolute correlations are modest.

The "populations" are the benchmark liquids in the file. These are the Kob-Andersen binary Lennard-Jones mixture at densities 1.1, 1.2, 1.4 and 1.6; its purely repulsive Weeks-Chandler-Andersen truncation; the modified $(q,p)$ potentials (12,11), (12,6) and (8,5); and the swap-equilibrated polydisperse soft spheres used by SEER.

Comparing liquids at equal temperature would be meaningless. We will compare them at matched relaxation time measured relative to each liquid's onset temperature. That onset temperature is defined by the entropy crossing in Determination of onset temperature from the entropy for fragile to strong liquids, and it marks where landscape-dominated hopping begins, as described in Crossover to potential energy landscape dominated dynamics in a model glass-forming liquid.

The Lennard-Jones/WCA pair matters most. The two differ only in the attractive tail, so they act as a controlled "admixture" contrast: one ingredient changed, everything else held fixed.

Baselines

The do-nothing reference assigns every particle the same predicted mobility. It has zero rank correlation and portability undefined, and it anchors how much any feature recovers. A second floor uses particle species and the radial-distribution neighbor count. This checks that gains do not come from trivial composition effects.

The strongest existing structural method is the Sharma et al. SOP. We compute it exactly as they define it, on the same configurations, with the same 100-run propensity and the same matched state points. The file reports only qualitative results for it. Correlation was highest for Lennard-Jones at density 1.1 and for the (8,5) model, and it collapsed for WCA. Microscopic barriers grew linearly with inverse SOP. No numeric Spearman value was found on Lacuna for the SOP–mobility correlation, so we will reproduce it ourselves before any comparison.

The landscape baseline is La Nave et al.'s local inherent-structure depth. It is far cheaper than barrier extraction, and it is the natural first competitor to the barrier feature.

The pooled baseline is a learned structural predictor: a regression on rich local descriptors trained jointly on all source liquids. It is tested leave-one-liquid-out. It follows the genetics practice of pooling ancestries. In Multi-ancestry polygenic risk scores using phylogenetic regularization, naive pooling gave BMI correlation 0.03 in non-White groups, against 0.10 for single-population and 0.14 for tree-regularized models. Pooling is therefore not automatically safe. All learned models get the same descriptors, the same configurations and an equal hyperparameter search budget, tuned only on source liquids.

The key ablation removes the barrier feature from the landscape predictor and keeps only basin depth. This isolates whether barriers, rather than inherent-structure information in general, carry the transferable signal.

"Beating" the strongest baseline has a fixed meaning: the barrier-based predictor must show leave-one-liquid-out portability above the SOP's on the WCA and (8,5) targets. Confidence intervals come from resampling configurations, and the result must hold at both a moderately and a deeply supercooled matched state.

Why not just…?

Why not just train a more expressive structural model on all the liquids at once? This is the most serious threat. A flexible model trained across potentials may learn whatever local geometry correlates with barriers in each. The leave-one-liquid-out protocol settles it. If pooled structural models match barrier-based portability on a held-out potential, then structure is not a liquid-specific tag in any useful sense, and we would report that the tag hypothesis fails. The DendroPRS pooling result suggests pooling can hurt, but that is a genetics result and does not decide the physics question.

Why not just use local basin depth, which needs only a minimization? La Nave et al. show it already tracks mobility steeply in Kob-Andersen. If depth alone matches barriers in transfer, the expensive excitation search is unnecessary. We would adopt depth as the cheap transferable feature, which is still a useful result. The ablation decides this directly. We expect depth to fail where basin depth and escape barrier decouple, which is exactly the regime SEER's flash-reheating results point to.

Why not just recalibrate the SOP on a little target-liquid data? Local re-weighting works in genetics. Leveraging the local genetic structure for trans-ancestry association mapping borrows information only where local cross-population correlation is high and gains 17–24% in prediction $R^2$. It requires target dynamics, though, and target dynamics are the expensive thing we want to avoid simulating. We will report the SOP with recalibration on 5% and 20% of target configurations. If barrier features with zero target data match that, they win on cost. If recalibrated SOP is clearly better, then a small amount of target data is the practical answer and the barrier result becomes mainly explanatory.

Why not accept Sharma et al.'s regime explanation as sufficient? Their enthalpic-versus-entropic account describes when the SOP works. It does not say whether mobility in repulsive liquids is still set by something structure fails to capture. The mediation check distinguishes two cases. If the SOP–barrier link drops while the barrier–mobility link holds, decoupling is a tagging failure. If both drop, their regime explanation stands and barriers are not the missing cause.

Decisive Checks

The proposal succeeds if three things hold at matched relaxation times. First, the barrier–propensity rank correlation varies much less across the seven liquids than the SOP–propensity correlation. Second, the observed SOP portability on each target is predicted, within resampling error, by the change in the SOP–barrier correlation. Third, on the held-out WCA and (8,5) targets, the barrier predictor exceeds both the SOP and the pooled structural model in portability.

The proposal fails informatively if the barrier–mobility link breaks specifically in the attractive Lennard-Jones liquids. That outcome would extend SEER's single-model conclusion only to repulsive systems. It would locate where cooperative effects, such as the elastic avalanches of Scaling Description of Dynamical Heterogeneity and Avalanches of Relaxation in Glass-Forming Liquids, take over.

Barrier coverage and propensity noise from finite isoconfigurational ensembles are the main confounds. Both will be reported per liquid, so that a portability gap cannot come from missing barriers or noisier targets alone.