Quantifying Gibbs measures of disordered crystals up to the solid-liquid phase transition

The paper addresses the long-standing problem of quantifying Gibbs measures for thermally disordered condensed matter systems, specifically focusing on how to define these measures over non-compact configuration spaces like crystals, which is essential for accurately simulating macroscopic electronic properties. The central conjecture proposes that a crystalline phase is uniquely encoded in the stable facets of Voronoi cells throughout its entire phase diagram up to the melting line.

The gist: The Gibbs measure of a material's classical degrees of freedom is fully encoded in the statistics of the pointed Voronoi cell geometries, specifically by observing four large facets that remain stable up to the melting line for crystalline silicon.

Pattern Recognition and Pattern Analysis

The paper tackles the challenge of removing labels from indistinguishable atoms by viewing a thermally disordered configuration as a pattern in the physical space R3. The authors propose using topological methods, such as endowing compact subsets of R3 with the Hausdorff metric or employing one-point compactification, to formalize this. This leads to modeling the configuration space as the orbit of one Delone set in an abstract topological compact space of patterns, which is identified with a probability measure over the space of Delone subsets by Jean Bellissard. The practical obstacle remains finding an explicit characterization of this configuration space that supports the Gibbs measure.

Encoding the Configuration Space using Pointed Voronoi Cells

The authors present a solution by engaging the geometry of Voronoi cells. A pointed Voronoi cell is defined as a cell plus its unique enclosed atom, and it can be quantified by at most 3N¯ real numbers encoding the positions of N facets. The key insight is that these Voronoi cells do not need special labels; an unordered set of cells can be used to reconstruct the crystal by matching facets between different cells—gluing together these two cells along that specific facet and repeat the process. This reconstruction works for both crystalline and liquid phases.

Unique and Complete Markers for Crystalline Phases

For the silicon crystal, they demonstrate that the existence of exactly four facets that remain stable for temperatures up to silicon’s melting line is a signature of the crystalline phase. While 0K Voronoi cells have 16 facets requiring prohibitive parameters, thermal fluctuations introduce more facets. However, analyzing histograms of facet numbers and areas reveals that below the melting temperature, a peak corresponding to exactly four large facets endures all the way to the solid-liquid transition. This empirical observation assures that an atom of crystalline silicon continues to have precisely four first nearest-neighbors up to the melting line.

Quantifying the Gibbs Measure

The Gibbs measure is supported by a sub-manifold of the configuration space defined by these stable facets. The authors show that when restricted to this manifold, the measure has a density:

**/dP(q) = ρ(q) d 3q, where q represents the four vectors carried by each atom in primitive cells. **

The task is to quantify this density by extracting moments from FPMD simulations. By analyzing the second moments of these vector correlations, they derive a covariance matrix for the Gibbs measure. Symmetry considerations reduce the required data significantly:

**/µ(a, n; a′, n′) = E[vn(a) · vn′ (a′)] − E[vn(a)] · E[vn'(a′)], where (n, n') correspond to first or second nearest-neighboring pairs. **

Conjectures and Implications

The paper concludes with three main statements:

1. The Gibbs measure is fully encoded in the statistics of the pointed Voronoi cell geometries.

2. For crystalline silicon, it is supported by the data from the four largest Voronoi facets stable up to the melting line.

3. The Gibbs measure is singular and supported by a sub-manifold of configuration space, which becomes non-singular when restricted to this manifold with a density function derived from moments.

The results allow for generating realistic thermalized atomic configurations using Monte Carlo algorithms, which can then be fed into electronic structure simulations to compute realistic electronic properties at finite temperatures. The analysis suggests that the Gibbs measure is well approximated by a multivariate normal distribution, and the fluctuations related to second nearest-neighboring atoms are one order of magnitude smaller than those for first nearest-neighboring atoms. This technique provides a rigorous device for finding unique real-space markers for crystalline phases with complex phase diagrams.

Materials and Methods

The simulations utilize the second-generation Car-Parrinello method (CP2K/Quickstep) with a cubic cell of 1000 silicon atoms, running 10 FPMD simulations ranging from 300 K to 3000 K. The Voronoi tessellation is performed using the Voro++ library to compute Voronoi cells and their normal vectors.

As a fastidious and diligent researcher, I have analyzed this groundbreaking work on quantifying Gibbs measures in disordered crystals using Voronoi tessellations. The core insight is that the entire configuration space of a thermally disordered solid phase is uniquely encoded by the statistics (specifically, the four largest facets) of its Voronoi cells.

Here are the specific improvements and capabilities for AI systems derived from this research:

Improved AI System Capabilities

The primary capability unlocked is moving from simulating atomic configurations to directly quantifying and sampling the underlying thermodynamic equilibrium state (the Gibbs measure) of a disordered material. This allows for the creation of thermodynamically accurate models rather than purely kinetic or static approximations.

1. Enhanced Materials Property Prediction (Phase-Dependent Accuracy)

Traditional AI/ML models often rely on training data from specific, idealized states (e.g., 0K crystal structures). This paper enables the creation of surrogate models that inherently account for thermal disorder and phase transitions up to the melting line.

2. Real-Space Fingerprinting for Phase Identification

The system can be equipped with a real-space fingerprinting module capable of distinguishing between different crystalline phases (e.g., solid vs. liquid) based on the stability of specific Voronoi cell facets, even when those phases are structurally similar or exhibit high thermal fluctuations.

3. High-Fidelity Configuration Generation

The system can generate atomic configurations that are not just physically plausible but statistically representative of the true Gibbs measure at a given temperature, crucial for accurate electronic structure calculations.

4. Direct Quantization of Configurational Entropy

The system can directly calculate the configurational entropy (or related measures) by analyzing the moments of the Voronoi cell normal vectors, providing a rigorous thermodynamic quantity that is often difficult to compute in standard MD simulations.

Specific Improvements and Applications

Here are the concrete AI system improvements:

Improvement Area Specific Technical Implementation Application in AI System

:---:---:---

1. Gibbs Measure Sampling (Statistical Physics Layer) Integrate a Voronoi Vector Analysis module into the sampling routine. Instead of relying solely on atomic positions, the system samples the space of Voronoi cell normal vectors, weighted by their joint probability density function derived from FPMD moments (Eq. 7). Generates physically accurate disordered configurations for electronic structure simulations (e.g., DFT inputs), drastically improving predictions for transport coefficients and band structures at finite temperatures.

2. Phase Transition Detection Develop a classifier trained on the statistics of Voronoi cell facet areas and counts across temperature ranges (as shown in Fig. 3). Automatically identifies the precise temperature boundaries of solid-liquid phase transitions in complex materials based on characteristic changes in Voronoi geometry, offering a more robust alternative to traditional thermodynamic methods.

3. Crystalline Phase Marker Extraction Implement an algorithm that extracts the subset of Voronoi vectors corresponding to the four largest facets (Statement 2) for any given configuration snapshot. Provides a unique, orientation-invariant descriptor (the set of four dominant vectors) that serves as a robust real-space marker to classify and identify crystalline phases in experimental data (e.g., from neutron scattering or microscopy).

4. Surrogate Model Training (Reduced Complexity) Train neural networks on the reduced parameter space defined by the moments of the Gibbs measure, rather than high-dimensional atomic coordinates. Creates computationally cheap, yet thermodynamically rigorous, surrogate models for calculating macroscopic electronic properties (like conductivity or dielectric response) that are valid across a wide range of relevant temperatures.

5. Machine Learning Descriptor Development Use Voronoi cell normals as input features for molecular environment descriptors in Graph Neural Networks (GNNs). Enables the AI system to learn chemical environments based on geometric stability markers rather than just atomic adjacency, leading to better predictions for potential energy surfaces and reaction pathways in disordered systems.

In summary, this research moves AI from merely pattern recognition of static structures to a sophisticated tool capable of modeling the complex, thermally driven equilibrium state of disordered condensed matter.

Quantifying the configuration space and the Gibbs measure of thermally disordered condensed matter systems has been a long standing problem. The challenge is to avoid the Gibbs paradox, which forbids any ordering or labeling of the atoms. Our key observation is that the lattice of a thermally disordered condensed matter system, in either solid, liquid or gas phase, can be fully reconstructed from the Voronoi cells of the atoms alone, even if these Voronoi cells are disassembled and randomly scrambled. In the example of the crystalline phase of silicon, the statistics of the Voronoi cells reveals the existence of four, and only four, large facets that are present with probability one for all temperatures up to the solid-liquid melting line. These four largest facets, which separate nearest-neighboring atoms, can be also be used to reconstruct the lattice of the crystal. Hence, their collection supplies the optimal representation of the configuration of the crystal. We conjecture that the existence of Voronoi facets that, despite their large thermal fluctuations, survive with probability one up to the melting temperature, is the fundamental signature of the crystalline solid phase and therefore key to quantifying the Gibbs measure over the entire solid phase.