A Landau-de Gennes Energy for Cage Shape in Glass-Forming Liquids

When a liquid is cooled toward a glass, its relaxation slows by many orders of magnitude while standard structural snapshots, such as the pair distribution of interparticle distances, barely change. Graph neural networks now predict which particles will move from a single static snapshot, but they do so with tens of thousands of opaque weights. They therefore cannot say which structural quantity sets the energy barrier a particle must cross to escape the cage formed by its neighbors. This proposal borrows the main object of liquid-crystal theory, the Landau-de Gennes Q-tensor, and applies it to the shape of each particle's cage. The Q-tensor is a symmetric, traceless 3×3 matrix that records how strongly, and along which axes, a set of directions is aligned, without distinguishing a direction from its opposite. A cage squashed along one axis is a head-tail symmetric object of exactly this kind. We build a smooth Q-tensor field of cage shape from simulation snapshots, write the standard Landau-de Gennes energy density for it (a few polynomial invariants plus gradient terms), and fit only its handful of coefficients so that the local energy predicts particle mobility.

The claim is that a Landau-de Gennes cage-shape energy with roughly ten fitted coefficients recovers most of the predictive power of the best graph networks for glassy dynamics. We further claim that its gradient (elastic) terms carry information that per-particle orientational order cannot. Because the model has the Landau-de Gennes form, its coefficients define a static correlation length $\xi=\sqrt{L/a}$, where $L$ is the elastic coefficient and $a$ the quadratic bulk coefficient. Tracking $\xi$ against temperature gives a direct, structure-based test of whether the slowdown is driven by growing cooperative regions or by rising local barriers. The program has three parts: interpretable propensity prediction, a length-scale measurement, and a test on true nematic particles.

Core Claim and What Is New

The closest existing work comes from two directions. EIGNN shows that the direction in which a particle sits off-center in its cage matters a great deal for predicting its mobility. It encodes that direction with equivariant features and Clebsch-Gordan tensor products, which are the standard way to combine rotating vector and tensor features while preserving rotation behavior. Geo-GNN shows that three-body bond angles fix a "low-pass filter" failure of distance-only networks. Both motivate the proposal, since they find that directional and angular structure beyond pair distances carries dynamical information. Both treat that structure as raw input to a learned network rather than as an order-parameter field with an energy.

On the liquid-crystal side, Glass Transition in Monolayers of Rough Colloidal Ellipsoids is the only work in the file that ties nematic alignment to glassy arrest. It finds that the local alignment $S$ predicts rotation in smooth ellipsoids, and that slight surface roughness destroys this correlation. That study is in 2D and offers no field theory.

What is new here is the combination of three things. First, the Q-tensor is computed from cage geometry, so it applies to spheres as well as rods. Second, the full Landau-de Gennes invariant set is used, including biaxiality and gradients. Third, the result is a closed-form, few-parameter energy whose coefficients can be read physically.

A reviewer should note one subtlety up front. The rotation-invariant magnitude of the rank-2 part of the neighbor-bond distribution is essentially the classic Steinhardt $q_2$ parameter. Per-particle magnitude alone is therefore not new. The proposal stands or falls on whether the parts that $q_2$ cannot express add predictive power. Those parts are the cubic invariant $\mathrm{tr}\,Q^3$, which separates uniaxial from biaxial cage shapes; the gradient terms, which penalize misalignment between neighboring cages; and the coupling between the cage axes and the particle's off-center displacement.

Research Program

The first step is the field construction. For each particle $i$ in a snapshot, we take its neighbors $j$ and the unit bond vectors $\hat r_{ij}$. The cage tensor is $$Q_i=\frac{1}{z_i}\sum_{j}w(r_{ij})\left(\hat r_{ij}\hat r_{ij}^{T}-\tfrac13 I\right),$$ where $z_i$ is the weighted neighbor count, $w$ is a smooth radial cutoff, and $I$ is the identity. This is the discrete version of the Landau-de Gennes definition, with the bond-direction distribution playing the role of the molecular orientation distribution. We smooth $Q_i$ with a Gaussian kernel of width $\sigma$ to obtain a field $Q(\mathbf x)$ and its spatial derivatives. We build the field on both inherent structures (local energy minima) and cage centers. The inherent-structure version does not require a well-defined cage center. That matters at $T=0.64$, the temperature at which EIGNN breaks down.

The second step is the energy model. The local energy density is $$f_i=a\,\mathrm{tr}Q^2-b\,\mathrm{tr}Q^3+c\,(\mathrm{tr}Q^2)^2+L_1|\nabla Q|^2+L_2|\nabla\!\cdot Q|^2+\gamma\,\hat d_i^{T}Q\,\hat d_i,$$ evaluated at particle $i$. The last term couples the cage axes to $\hat d_i$, the unit direction of the particle's offset from its cage center. It is inversion-symmetric, so a particle leaning toward the open end of an elongated cage is treated the same whichever end it leans toward. The term with $L_2$ is the unequal-elastic-constant term that On a Divergence Penalized Landau-de Gennes Model analyzes in the continuum. We reuse its functional form, not its boundary-value analysis.

The coefficients are fit by regressing log dynamic propensity on $f_i$, with separate coefficient sets for each particle species. Dynamic propensity is a particle's mean squared displacement averaged over many runs that start from the same positions with random velocities. The resulting model has about a dozen parameters per temperature and timescale.

The third step is analysis. We report how $a$, $L_1$, and $\xi=\sqrt{L_1/a}$ vary with temperature. We also correlate $f_i$ with the local energy barriers extracted by the SEER algorithm (Systematic Excitation ExtRaction) of Local vs. Cooperative: Unraveling Glass Transition Mechanisms with SEER. That paper finds that the barrier rise alone accounts for the growth of the activation energy, while the number of particles per excitation shrinks on cooling. If cage-shape energy predicts those barriers and $\xi$ stays flat as the barriers rise, the Landau-de Gennes picture supplies a structural reason for "local" barrier growth, which the local camp currently lacks. If $\xi$ grows with cooling, the result instead lends structural support to the cooperative picture of Random field Ising-like effective theory of the glass transition. II and point-to-set arguments. Either outcome is informative, which is why the length measurement is worth doing. Does a Growing Static Length Scale Control the Glass Transition? constrains the interpretation, because it argues that a growing static length need not control $\tau_\alpha$. We will therefore treat $\xi$ as evidence only if it tracks the barriers, not merely the temperature.

The fourth step tests real nematics. Following the coarse-grained Gay-Berne ellipsoids with surface bumps from the rough-ellipsoid study, we simulate 3D systems with smooth and rough particles. Each particle then carries two Q-fields: one from its molecular orientation and one from its cage shape. We add a coupling term $\kappa\,\mathrm{tr}(Q_{\mathrm{mol}}Q_{\mathrm{cage}})$. The prediction, motivated by the 2D finding that roughness merges the rotational and translational transitions, is that the fitted $\kappa$ is small for smooth particles and large for rough ones. This would give a field-level account of the "geometrical interlock." Separately, the 2D study reports that the correlation between alignment and rotation vanishes with roughness. A Landau-de Gennes model with a small number of fitted terms should reproduce that loss of correlation and identify which term absorbs it. This also answers, in 3D, the question the 2D monolayer study leaves open: whether a separate orientational glass survives.

Baselines

The naive reference is a per-particle linear model on scalar descriptors: local density, potential energy, and Steinhardt $q_2$ and $q_4$. It is fit with the same regression, data split, and per-species treatment. It answers whether Landau-de Gennes structure adds anything beyond the magnitude of the rank-2 bond order. Effect of total and pair configurational entropy suggests that pair-level scalars should be weak, since it reports liquids with nearly identical pair structure but very different relaxation times. No reported result was found on Lacuna for this exact linear baseline on GlassBench-3D, so we will compute it ourselves.

The strongest existing methods are the graph networks. BOTAN (BOTAN) predicts relative motion between neighboring particles in addition to single-particle motion. It reports a Pearson correlation of about 0.9 at short, β-relaxation times on GlassBench-3D. Geo-GNN reports a Pearson correlation of about 0.798 in the caging regime at $S(q,t)=0.8$, against about 0.515 for a standard GNN. EIGNN reports qualitatively that its cage variant beats GlassMLP, CAGE, and BOTAN at $T=0.44$, but the file gives no numerical value for it. We will retrain all three on the identical 400/100 configuration split at all four temperatures and all timescales. Each gets a matched hyperparameter search budget of 50 trials, and we report mean and standard deviation over five seeds.

We do not expect to beat the networks in raw correlation. A win is defined as two outcomes together. The first is retaining at least 85% of the best network's Pearson correlation with fewer than 20 parameters. The second is a clear improvement over the network at $T=0.64$ when using the inherent-structure field.

Two ablations isolate the key ingredients. The first removes the gradient terms ($L_1=L_2=0$), leaving a purely local Q model. The second replaces $Q$ with the polar displacement vector and its outer-product-free invariants. If the first ablation matches the full model, the field-theoretic part of the proposal is empty. If the second matches it, head-tail symmetry is not doing work. A third check feeds $Q_i$ and $\nabla Q$ as extra node features into EIGNN, to see whether networks already extract this information.

Why not just…?

Why not just use Steinhardt parameters or SOAP descriptors with a linear model? This is the real threat. The magnitude of the traceless part of $Q_i$ is $q_2$ up to normalization, so a large share of any gain could come from information that $q_2$ and $q_4$ already hold. The naive baseline and the gradient-free ablation settle this directly. Machine learning of microscopic structure-dynamics relationships gives some prior reason for doubt: purely structural spherical-harmonic descriptors there merged dynamically distinct environments. If the local-only Q model equals Steinhardt regression and gradients add nothing, we will report that the Landau-de Gennes framing adds no information for spheres. The proposal would then reduce to the ellipsoid study.

Why not just add Q-tensor features to EIGNN or BOTAN and report a higher correlation? That would probably raise accuracy slightly, and we run it as a check. It would not answer the scientific question, however, because a network's weights do not yield a correlation length or a barrier law. The point of the few-parameter energy is that $\xi$ and the $\kappa$ coupling are measurable quantities that can be compared with SEER barriers and with the 2D roughness result. A network gives neither.

Why not just use the displacement vector, which EIGNN already shows is informative? For spheres the vector and the tensor coupling $\hat d^{T}Q\hat d$ may perform similarly, and the vector ablation measures that. For ellipsoids, a vector representation of particle orientation distinguishes a particle from its 180°-flipped copy, which is the same physical state. That is unphysical, and as Scholich et al. and liquid-crystal theory generally hold, it breaks the symmetry. In the anisotropic experiments the tensor is therefore required, not optional.

Proposed Experiments

The first experiment runs on GlassBench-3D, the 4,096-particle Kob-Andersen 80:20 Lennard-Jones mixture at $T=0.44$, 0.50, 0.56, and 0.64. Using the BOTAN-style protocol of 32 isoconfigurational runs per configuration, we predict propensity at timescales from caging to the α relaxation and report Pearson correlation. The decisive checks are three. Gradient terms must improve correlation beyond the local model at long times. If they help only at short times, that contradicts the Geo-GNN smoothness result and is itself worth reporting. The inherent-structure Q field must degrade less than EIGNN at $T=0.64$. And a model fit at $T=0.50$ must transfer to $T=0.44$, matching the transfer EIGNN reports.

The second experiment reproduces the SEER system: 2,000 polydisperse soft particles equilibrated by swap Monte Carlo. We correlate $f_i$, evaluated at each excitation's core, with the extracted local barrier, and we plot $\xi(T)$ against the activation energy, which that paper reports roughly doubles across the studied range. The decisive check is whether cage-shape energy explains a significant fraction of barrier variance, and whether its mean shift tracks $\Delta E$.

The third experiment uses 3D Gay-Berne ellipsoids with aspect ratio 3.8 and 0 to 10 surface bumps, across packing fractions bracketing the transitions. We fit $\kappa$ against roughness and test whether the model reproduces the loss of the alignment-rotation correlation and the merging of rotational and translational relaxation.

The project fails cleanly if all three of the following hold together: the gradient terms add nothing, $\xi$ shows no relation to the barriers, and $\kappa$ shows no dependence on roughness.