Data Coarse Graining

• Data Coarse Graining is a systematic process of reducing fine-scale, high-dimensional data into simplified representations while preserving essential observables.
• Techniques include stochastic kernel mappings, equilibrium Hamiltonian reductions, and generative latent-variable models to capture both static and dynamic properties.
• Applications span molecular dynamics, continuum mechanics, and network analysis, enabling efficient simulations and predictive modeling across diverse scientific fields.

Data coarse graining is the systematic reduction of a fine-scale, high-dimensional, or information-rich description to a lower-resolution representation intended to preserve specified quantities of interest. In the literature, this includes stochastic-kernel maps between probability measures and observables, equilibrium reductions of microscopic Hamiltonians to coarse variables, non-equilibrium stochastic models with memory, multiresolution abstractions of graphs and point clouds, and latent-variable generative models for molecular and materials systems (Gudder, 2021, Larini et al., 2010, Wang et al., 2021). Across these settings, coarse graining is not a single algorithm but a class of transformations whose adequacy is judged by which structures survive the reduction: ensemble averages, correlations, effective energies, transition laws, self-similar geometry, or predictive observables.

1. Formal definitions and conceptual scope

A mathematically explicit definition appears in the theory of probability measures and observables, where coarse-graining is defined through a stochastic kernel $v : \Omega_1 \times \mathcal{F}_2 \to [0,1]$. The induced affine map on measures is

$[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$

This formalism extends from measures to observables and instruments. In the finite case, stochastic kernels reduce to stochastic matrices, and coarse-graining coincides with post-processing. Discretization is a special case characterized by $0$-$1$ stochastic kernels, so deterministic lumping is distinguished from probabilistic smearing. The same work also notes that, unlike probability measures, two observables need not coexist (Gudder, 2021).

In equilibrium statistical mechanics, coarse graining is formulated as the construction of a reduced Hamiltonian or potential that reproduces selected expectation values. The generalized mean field theory writes

$H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$

with conjugate fields $\lambda_i$ chosen so that the coarse-grained model matches prescribed averages or correlations. For reduction from an atomistic ensemble to coarse coordinates $R=M[r]$, the effective equilibrium potential is

$$V_{\text{CG}}(R) = -k_B T \ln \int \delta[R-M(r)] e^{-\beta u(r)}\,dr + \text{const}.$$

Within this framework, Reverse Monte Carlo, MS-CG, relative entropy methods, and classical density functional theory are presented as particular cases of a common statistical-mechanical construction (Larini et al., 2010).

A distinct conceptual shift appears in predictive coarse-graining, where the reduced variables are treated not as deterministic images of fine variables but as latent generators of fine-scale data. The joint model is

$\bar p(X,x) = p_{cf}(x|X)\,p_c(X),$

with $p_c$ the coarse density and $[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$0 a probabilistic coarse-to-fine map. This replaces a prescribed many-to-one fine-to-coarse map by a directed generative model, permits predictive posterior distributions for fine reconstructions and observables, and makes uncertainty due to information loss explicit. Hierarchical priors are used to promote sparse coarse models and thereby address model complexity and model selection (Schöberl et al., 2016).

A more categorical formalization arises for open Markov processes. There, coarse-graining is not merely state aggregation but a morphism between open systems, represented as a 2-morphism in a symmetric monoidal double category. The construction is compatible with “black-boxing,” which maps an open Markov process to the linear relation between input and output data that holds in steady states, including nonequilibrium steady states with nonzero probability flux (Baez et al., 2017). This suggests that, in some settings, coarse graining is best viewed as structure-preserving composition rather than only as projection.

2. Dynamics, memory, and effective stochastic evolution

A recurrent theme is that coarse graining may preserve static or equilibrium structure while significantly altering dynamics. In non-equilibrium systems, this motivates the use of a non-stationary generalized Langevin equation for a coarse variable $[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$1,

$[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$2

where the memory kernel depends on two times rather than only on $[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$3. In this formulation, the dependence of non-equilibrium processes on initial conditions is represented explicitly, the two-time memory kernel is inferred from two-time auto-correlation data of non-equilibrium trajectory-averaged observables, and the non-Markovian process is embedded in an extended stochastic framework. To exploit the equivalence between the nsGLE and the extended dynamics, the memory kernel is parameterized by a two-time exponential expansion and fitted with a data-driven hybrid optimization procedure (Wang et al., 2021).

For linear overdamped Langevin systems with quadratic energy, a hierarchy of reduced models clarifies the distinction between equilibrium and dynamical fidelity. The standard Markovian effective dynamics,

$[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$4

and the improved model,

$[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$5

both reproduce the correct coarse-grained equilibrium distribution, but their dynamical behavior differs. The standard model systematically underestimates autocovariance at positive lag and yields mean-squared displacement that is too fast, whereas the augmented model gives a better global match to dynamical statistics, particularly when time-scale separation is significant (Hudson et al., 2023). The same analysis shows that equilibrium correctness alone is not a sufficient criterion for reduced stochastic dynamics.

A related data-driven strategy for condensed matter systems replaces unresolved bath forces and noise by a random variable sampled from a learned conditional distribution. The reduced integrator advances the distinguished degrees of freedom while substituting the interaction with solvent or other unresolved components by resampled auxiliary variables conditioned on present and past states. On a harmonic particle, a bistable particle, and a dimer with two metastable configurations, this procedure reproduces not only equilibrium distributions but also time correlations, memory effects, and first-passage behavior (Razo et al., 2023).

For continuous-time Markov chains, coarse graining is defined by a map $[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$6 from microstates to macrostates. The projected law $[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$7 evolves with a time-dependent coarse generator, while an effective closed dynamics is obtained from a time-independent generator

$[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$8

where $[V(\mu)](A) = \int_{\Omega_1} v(x,A)\,\mu(dx).$9 is the stationary measure and $0$0. Without assuming explicit scale separation, sufficient conditions based on logarithmic-Sobolev inequalities yield quantitative relative-entropy and total-variation error bounds between the exact coarse dynamics and the effective dynamics (Hilder et al., 2022).

3. Probabilistic, variational, and generative frameworks

Several recent approaches cast coarse graining as probabilistic inference with latent coarse states. In a physics-constrained state-space formulation, the coarse process is described by a transition law $0$1, while a probabilistic emission law $0$2 performs coarse-to-fine reconstruction. For random walkers, the transition law is parameterized by a feature library and sparse Bayesian learning with automatic relevance determination, and the emission law uses a softmax map to enforce conservation of mass. Stochastic variational inference identifies latent coarse states and model parameters, avoids overfitting, and quantifies predictive uncertainty caused by information loss (Felsberger et al., 2018).

A closely related small-data framework unifies dimensionality reduction and model-order reduction through an encoder-decoder architecture. The encoder $0$3 maps a high-dimensional microstructure to a lower-dimensional latent representation using sparse linear-Gaussian models on hand-crafted physical feature functions; the coarse-grained model $0$4 is typically deterministic and physics-based; and the decoder $0$5 lifts the coarse output back to the fine output space. Bayesian training is carried out with stochastic variational inference, and the variational objective is also used to suggest adaptive refinements of the reduced model (Grigo et al., 2019).

Manifold-learning formulations replace direct norms by geometric comparison of coarse and fine responses. In Grassmannian EGO, high-dimensional field data are projected by singular value decomposition onto the Grassmann manifold $0$6, and discrepancies are measured by the Grassmannian geodesic distance

$0$7

A Gaussian-process surrogate is trained on weighted sums of manifold distances and other discrepancy measures, and Efficient Global Optimization identifies coarse-grained continuum parameters that reproduce atomistic behavior. The method is demonstrated by calibrating the shear transformation zone theory of plasticity and coarse-graining parameters for amorphous solids (Kontolati et al., 2021).

A different development removes trajectory data altogether. In a data-free, energy-based molecular framework, an invertible map

$0$8

splits atomistic coordinates into slow collective variables $0$9 and fast variables $1$0. The target is the transformed Boltzmann density, and the model density is factorized as

$1$1

Training minimizes a reverse Kullback–Leibler divergence using only the interatomic potential, with tempering to stabilize optimization and encourage exploration of distinct modes. The resulting model generates unbiased one-shot equilibrium all-atom samples and was validated on a double-well potential, a Gaussian mixture, and alanine dipeptide (Stupp et al., 29 Apr 2025).

4. Geometry, renormalization, and hierarchical abstractions

In complex networks, coarse graining can be defined spectrally through the Laplacian. A field-theoretic construction on an undirected graph yields a Hamiltonian involving the Laplacian $1$2, and the correlation matrix is obtained from the pseudo-inverse $1$3 or from $1$4. The coarse-graining rule iteratively merges the pair of nodes with maximal off-diagonal correlation $1$5, forming super-nodes and a reduced adjacency matrix. Applied to Erdős–Rényi and Barabási–Albert networks and to several empirical networks, the method preserves degree distribution, degree-degree correlation, and clustering across scales in most cases; the Power Grid is an explicit counterexample in which self-similarity is not preserved (Loures et al., 2023).

For point-cloud and manifold data, diffusion condensation defines a time-inhomogeneous diffusion process that repeatedly recomputes the geometry as the data are smoothed. Starting from a Gaussian affinity

$1$6

the algorithm applies a sequence of Markov operators $1$7 to the evolving data coordinates. This produces a deep cascade of intrinsic low pass filters, gradually eliminates local variability, and yields a continuously hierarchical clustering in which the persistence of groupings across condensation times serves as an indicator of significance (Brugnone et al., 2019).

Renormalization-group-motivated learning adapts coarse graining to non-spatial data by explicitly balancing proximity and information loss. For nearly Gaussian data, pairings are chosen to maximize within-pair correlation while minimizing the overall projection error. For nonlinear data, correlation is replaced by mutual information, leading to an information-bottleneck-like objective

$1$8

where $1$9 is the compressed representation of a selected pair and $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$0 denotes the remaining variables. The method was examined on random Gaussian data, the Ising model, and glass systems, where it was used to recover collective or mesoscopic structure (Landy et al., 2023).

Tensor-network coarse graining provides a real-space RG setting in which reduced tensors directly encode universal field-theoretic data. For the 2D classical Ising model, HOTRG and GILT are used to coarse-grain vacuum and defect tensors, and the linearized tensor renormalization group extracts scaling dimensions from the eigenvalues of the linearized map. Coarse-grained defect tensors correspond to conformal states, OPE coefficients can be extracted from tensor fusion, and GILT+HOTRG yields accurate two- and four-point functions under specific conditions. The same study emphasizes limitations due to defect smearing and finite bond dimension, and finds that the minimal canonical form improves the stability of the RG flow (Guo et al., 2023).

5. Representative scientific applications

In dislocation plasticity, coarse graining is used to derive voxel-size-consistent internal energy densities from discrete dislocation simulations. A reference microscopic energy density $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$1 is computed on a fine grid, a mesoscopic continuum energy $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$2 is computed on the target grid, and the lost energy is defined as

$H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$3

The missing energy is then fitted as a function of coarse dislocation density fields such as $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$4, $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$5, and $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$6, with coefficients that depend explicitly on the voxel size $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$7. This resolves the voxel-size-dependent energy calculation problem and provides a systematic route to constitutive closure in continuum dislocation dynamics (Song et al., 2020).

In molecular coarse-grained force-field learning, symmetry-aware architectures change the data requirements of the reduction itself. For coarse-grained water, networks with equivariant convolutional operations were reported to produce functional models using datasets as small as a single frame of reference data, whereas networks without these operations could not. In the reported experiments, Allegro trained on 1 frame achieved force RMSE $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$8 for TIP3P and $H = H_{\text{ref}} + \sum_{i=1}^K \lambda_i o_i,$9 for SPC/E, while DeePMD trained on 100 frames achieved $\lambda_i$0 and $\lambda_i$1, respectively; DeePMD with 10 frames was unstable in production simulations. The stated explanation is that equivariant networks do not need to learn rotational equivariance by brute force from the data, because symmetry is built into the architecture (Loose et al., 2023).

In granular dynamics, adaptive coarse graining for DEM simulations combines fine and coarse regions in a single computation. The method embeds spatially confined fine-scale subregions inside a larger coarse-grained simulation and couples them in both directions by volumetric passing of boundary conditions. Stress, velocity, and mass-flow information are transferred across interfaces, multiple levels of coarse graining can be nested, and the same approach is proposed for coupled CFD-DEM simulations with adaptive CFD mesh resolution (Queteschiner et al., 2017).

These examples show that coarse graining is not restricted to dimensionality reduction in an abstract sense. It can act as constitutive closure, symmetry-aware model compression, multiresolution domain decomposition, or explicit resolution matching between microstructure, continuum fields, and solver architecture. This suggests that the practical meaning of “data coarse graining” is highly application-dependent, even when the underlying operation is always a controlled loss of information.

6. Error control, misconceptions, and current directions

A common misconception is that reproducing equilibrium distributions or structural observables is sufficient for a good coarse-grained model. Multiple studies explicitly reject that equivalence. In non-equilibrium generalized Langevin modeling, the whole construction is driven by the need to preserve dynamics after the loss of degrees of freedom, and in linear stochastic systems the standard Markovian reduction is shown to give systematically biased autocovariances and mean-squared displacements even when the equilibrium distribution is correct and even when the system exhibits large time-scale separation (Wang et al., 2021, Hudson et al., 2023).

Another misconception is that information loss is always detrimental to prediction. A solvable model of ridge-regularized linear regression under data coarse graining shows a nonmonotonic dependence of prediction risk on the degree of coarse graining. A “high-pass” scheme that filters out less relevant, lower-signal features can improve generalization, whereas a “low-pass” scheme that integrates out more relevant, higher-signal features is purely detrimental. With optimal regularization, this nonmonotonicity remains and is presented as distinct from double descent (Nguyen et al., 18 Sep 2025). The result does not claim that arbitrary compression helps; it claims that the structure of the discarded variables matters.

Quantitative control of the reduction depends strongly on the coarse-graining map. For continuous-time Markov chains, the error between projected and effective dynamics is bounded in relative entropy under logarithmic-Sobolev assumptions on the conditional stationary measures, and the theory identifies fast mixing within level sets as the key property of a good map (Hilder et al., 2022). In tensor-network coarse graining, defect smearing, edge effects, and finite bond dimension constrain the reliable extraction of scaling dimensions and correlation functions (Guo et al., 2023). In dislocation energetics, the relevant correction must be parameterized explicitly by voxel size, because otherwise the energy remains mesh- or voxel-dependent (Song et al., 2020).

Current directions include replacing data dependence by energy-based objectives, exploiting symmetry to reduce training requirements, and refining reduced models adaptively rather than fixing a single resolution in advance. Data-free flow-based coarse graining targets the all-atom Boltzmann distribution without sampled trajectories (Stupp et al., 29 Apr 2025); ARD- and ELBO-based frameworks use variational criteria to identify salient features and regions for refinement (Grigo et al., 2019); adaptive DEM combines several resolutions in one simulation to retain local detail where particle-size effects matter (Queteschiner et al., 2017). A plausible implication is that future coarse-graining pipelines will be increasingly hybrid: partially probabilistic, partially physics-constrained, and explicitly designed around what information is permitted to be lost.