Statistical Inference for Disordered Sphere Packings

Sphere packings are essential to the development of physical models for powders, composite materials, and the atomic structure of the liquid state. There is a strong scientific need to be able to assess the fit of packing models to data, but this is complicated by the lack of formal probabilistic models for packings. Without formal models, simulation algorithms and collections of physical objects must be used as models. Identification of common aspects of different realizations of the same packing process requires the use of new descriptive statistics, many of which have yet to be developed. Model assessment will require the use of large samples of independent and identically distributed realizations, rather than the large single stationary realizations found in conventional spatial statistics. The development of procedures for model assessment will resemble the development of thermodynamic models, and will be based on much exploration and experimentation rather than on extensions of established statistical methods.

💡 Research Summary

The paper addresses a fundamental gap in the quantitative analysis of disordered sphere packings, which are central to models of powders, composite materials, and the atomic structure of liquids. While such packings are routinely generated by simulation algorithms or produced experimentally, there is no established probabilistic framework that treats a packing as a realization of a stochastic process. Consequently, traditional spatial‑statistics methods—designed for a single large, stationary realization—are ill‑suited for assessing how well a proposed packing model reproduces real data.

The authors argue that meaningful model assessment must be based on large collections of independent and identically distributed (i.i.d.) realizations of the same packing procedure. By repeating the same algorithmic or experimental protocol under strictly controlled conditions (identical random seeds, compression rates, boundary conditions, etc.), one can obtain a sample of packings that are statistically independent. This shift from a single massive sample to many independent samples changes the data‑collection paradigm and demands new descriptive statistics capable of capturing the common structure across realizations.

To that end, the paper proposes three families of novel descriptive statistics:

1. Contact‑network topology – Treating the set of touching spheres as a graph, the authors suggest measuring degree distributions, clustering coefficients, cycle lengths, and network diameters. These graph‑theoretic descriptors quantify global connectivity, which is directly linked to mechanical strength, transport properties, and deformation behavior of the packed material.

2. Multiscale density fluctuations – By sliding observation windows of varying sizes through a packing and recording the number of spheres in each window, one can construct volume‑distribution functions and scale‑dependent spatial autocorrelation functions. This approach captures heterogeneity and clustering that are invisible to traditional pair‑correlation or structure‑factor analyses.

3. Thermodynamic‑style metrics – The authors advocate estimating quantities analogous to energy, entropy, and free energy from the packing process (e.g., work done during compression, contact‑energy approximations, entropy of the contact network). These metrics provide a bridge between statistical description and the underlying physics, allowing model comparison in a manner reminiscent of free‑energy minimization in thermodynamic model fitting.

With these statistics in hand, model evaluation proceeds by generating the empirical distribution of each descriptor for every candidate packing model (e.g., random seeding, controlled compression, particle‑shape variations). The empirical distributions are then compared to those obtained from real data using divergence measures such as Kullback‑Leibler divergence, Wasserstein distance, or Bayesian posterior predictive checks. Rather than a simple hypothesis test, the process is an exploratory search for the model that minimizes a composite “statistical energy” reflecting the overall discrepancy across all chosen descriptors.

The paper also discusses the limitations of existing stochastic geometry theory for this problem. Disordered packings exhibit long‑range dependencies, non‑Markovian behavior, and complex interaction networks that are not captured by classic Poisson or Gibbs point processes. Consequently, the authors view the development of new probabilistic models—perhaps extensions of interacting Poisson processes, non‑standard Gibbs fields, or even deep generative models—as a long‑term research agenda. In the interim, they recommend a pragmatic, data‑driven workflow: (i) generate large i.i.d. samples, (ii) compute the proposed multiscale and topological statistics, (iii) compare empirical distributions across models, and (iv) iterate the simulation or experimental protocol based on the observed discrepancies.

Overall, the paper outlines a methodological roadmap for statistical inference in disordered sphere packings. It emphasizes the necessity of independent replication, the creation of richer descriptive statistics that go beyond pairwise correlations, and an assessment philosophy akin to thermodynamic model fitting—relying heavily on empirical exploration rather than on direct extensions of classical statistical theory. This framework is positioned as a foundation for future work that will eventually integrate rigorous stochastic models, thereby enabling more reliable validation of packing algorithms and more accurate physical predictions for materials that depend on disordered microstructures.