Structural Coarse-Graining

• Structural coarse-graining is a reduction procedure that replaces detailed, fine-scale models with effective variables while retaining the underlying design principles that govern system behavior.
• It utilizes variational, fluctuation-based, and many-body approaches to derive reduced descriptions that capture essential dynamics and structural correlations.
• Applications span molecular simulations, network diffusion, and adaptive multiresolution methods, enabling faithful reproduction of properties from microscopic to macroscopic scales.

Searching arXiv for recent and foundational papers on structural coarse-graining to ground the article in cited work. Structural coarse-graining denotes a family of reduction procedures in which microscopic, atomistic, or fine-scale descriptions are replaced by reduced variables, effective units, or multiresolution representations while attempting to retain the structure that controls the target behavior. In the literature, that retained structure is not unique: it may be a large-deviation variational principle, a force-flux constitutive law, a many-body potential of mean force, a sequence-dependent elastic ground state, a defect-induced electronic perturbation, a graph-diffusion geometry, or the mesoscale scale of structural inhomogeneities. Taken together, these works suggest that structural coarse-graining is distinguished less by simple dimensionality reduction than by the preservation of a problem-specific organizing structure under projection, averaging, homogenization, or aggregation (Duong et al., 2014, Montefusco et al., 2020, Scherer et al., 2017, Queteschiner et al., 2017, Nicholas et al., 2020, Xue et al., 2024).

1. Scope and defining characteristics

A recurring theme is that the “right” coarse variables are not determined solely by convenience. In stochastic-process formulations, the reduced variables are selected by a many-to-one map $\Pi:\mathcal X\to\mathcal Y$, and the central question is how the projected variables behave in a singular limit. In nonequilibrium statistical mechanics, the coarse variables $X(y)$ are assumed to resolve the slow dynamics, and the structural output is a constitutive law in terms of an entropy $S(x)$ and a dissipation potential $\Psi^*(x,\xi)$. In molecular coarse-graining, the exact coarse interaction is the many-body potential of mean force, but practical models project it onto restricted bases such as pair, three-body, or basin-specific interactions. In adaptive multiscale methods, structural coarse-graining may preserve fine spatial detail only in selected subregions, while in graph and network settings it may aggregate nodes according to diffusion-defined communication or clique structure rather than geometric proximity (Duong et al., 2014, Montefusco et al., 2020, Scherer et al., 2017, Queteschiner et al., 2017, Fernandez-Iriondo et al., 9 Feb 2026, Xue et al., 2024).

Domain	Reduced object	Structure retained
Large deviations	$\hat\rho=\Pi_\#\rho$	Rate functional and zero-cost evolution
Nonequilibrium Markov processes	coarse state $x$	$S(x)$ and $\Psi^*(x,\xi)$
Molecular CG force fields	beads, COM sites, basin surfaces	PMF, RDFs, angular structure, cross-correlations
Polymer and DNA models	blobs, coarse segments	compressibility, excess free energy, sequence-dependent geometry
Adaptive multiscale simulation	embedded fine subdomains or representative nodes	local particle/electronic detail near critical regions
Network reduction	supernodes and weighted super-links	contagion-relevant dense structure or diffusion geometry

This breadth also clarifies a common misconception. Structural coarse-graining is not synonymous with pair-potential fitting, low-dimensional embedding, or state-variable elimination alone. The reduced description may preserve explicit many-body effects, local variational structure, equilibrium conditional measures on fibers, or hierarchy across scales, and several papers make explicit that a reduced model need not be postulated independently but can be induced from a fluctuation or homogenization principle (Duong et al., 2014, Dairyko et al., 2017, Bereau et al., 2018).

2. Variational and fluctuation-based formulations

One major line of work casts structural coarse-graining in a variational language. For empirical measures of i.i.d. Markov processes, the large-deviation principle

$$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$

is the central object, and the same rate functional identifies both fluctuations and typical evolution through $I^\varepsilon(\rho)=0$ (Duong et al., 2014). Coarse-graining then proceeds by restricting the dual variational representation $X(y)$0 to coarse observables $X(y)$1, with $X(y)$2, and by seeking a reduced dual object $X(y)$3 for $X(y)$4. The limiting reduced functional

$X(y)$5

then encodes both the fluctuation cost of the coarse variables and the deterministic reduced law via $X(y)$6 (Duong et al., 2014). The difficult step is the local-equilibrium or closure property

$X(y)$7

which the paper identifies as the heart of rigorous coarse-graining.

A closely related but more constitutive formulation appears in nonequilibrium statistical mechanics. There the macroscopic law is written as

$X(y)$8

and the structural claim is that the dissipation potential $X(y)$9 can be reconstructed from the nonlinear cumulant-generating expression

$S(x)$0

This generalizes Green–Kubo from the quadratic diffusion case to general Markov processes, especially jump processes, and makes the constitutive structure itself the output of coarse-graining rather than a prior assumption (Montefusco et al., 2020). In the reaction $S(x)$1, the resulting dissipation potential is nonquadratic,

$S(x)$2

which the paper uses to show that a Green–Kubo-like diffusion scheme cannot recover the correct entropy, friction sampling rule, and most probable evolution simultaneously (Montefusco et al., 2020).

A third variational strand quantifies how good a chosen coarse variable is. For overdamped and Langevin dynamics, a coarse-graining map $S(x)$3 or phase-space map $S(x)$4 induces conditional measures on fibers $S(x)$5 or $S(x)$6, effective coefficients by conditional averaging, and error estimates between the exact projected law and a closed effective Markov process on reduced space (Dairyko et al., 2017). The paper states that the quality of the chosen coarse-graining is measured by certain functional inequalities encoding the scale separation of the Gibbs measure, specifically fiberwise Poincaré, Talagrand, and logarithmic Sobolev inequalities, together with constants that quantify variation of drift and diffusion along fibers (Dairyko et al., 2017). This suggests that structural coarse-graining is good when the unresolved variables equilibrate strongly on level sets and the coarse coefficients are nearly constant there.

3. Many-body representability and coarse-grained force fields

In molecular systems, structural coarse-graining is repeatedly framed as a representability problem: the exact reduced interaction is a many-body free-energy object, but practical models project it onto restricted functional forms. One paper states this directly for equilibrium coarse-graining: the exact CG interaction is a unique many-body potential of mean force, while the practical force field depends on the representation chosen for it (Scherer et al., 2017). The same issue appears in a cluster-expansion treatment, where the exact effective Hamiltonian

$S(x)$7

is decomposed into a hierarchy

$S(x)$8

with pair, three-body, and higher-body terms (Tsourtis et al., 2016). In that hierarchy, pair PMFs are low-order approximations to an inherently many-body coarse-grained free energy.

This representability issue becomes concrete in water and methanol. When a short-range Stillinger–Weber three-body term is added and fitted simultaneously with a pair potential, the water pair potential becomes much more attractive while the three-body term compensates repulsively, even though the final $S(x)$9 improves (Scherer et al., 2017). The paper therefore introduces a residual-force strategy,

$\Psi^*(x,\xi)$0

motivated by Gram–Schmidt orthogonalization, so that the three-body fit targets only what remains after pair-level structure has been explained (Scherer et al., 2017). Water is the positive case, where three-body interactions are essential to reproduce structural properties; methanol is the contrasting case, where two-body interactions are sufficient to reproduce the main features of the atomistic system (Scherer et al., 2017).

The methane–ethane comparison in the cluster-expansion work sharpens the same point from another angle. For methane, pair PMF-based coarse-graining works very well, particularly at high $\Psi^*(x,\xi)$1 and low $\Psi^*(x,\xi)$2, because methane is nearly spherical. For ethane, a one-site COM model is substantially less accurate because the coarse map itself enforces a spherical approximation on an anisotropic molecule (Tsourtis et al., 2016). The paper concludes that three-body effective CG potentials give a small improvement in the liquid regime, but that significantly better results require even higher-order terms (Tsourtis et al., 2016).

A different response to the representability problem is to abandon the single-surface assumption. In the multisurface approach, one approximates the many-body potential of mean force by multiple basin-specific surfaces $\Psi^*(x,\xi)$3, with forces

$\Psi^*(x,\xi)$4

and state-dependent weights determined by structural proximity to conformational basins (Bereau et al., 2018). The paper argues that conventional additive force fields impose approximately factorized structure $\Psi^*(x,\xi)$5, whereas a multisurface model improves cross-correlations and, because it captures free-energy barriers more accurately, naturally achieves consistent long-time dynamics for systems governed by barrier-crossing events (Bereau et al., 2018).

4. Analytic renormalization in polymers, DNA, and molecular liquids

Another major strand of structural coarse-graining seeks analytic or semi-analytic renormalization across resolution levels. For polymer melts, an integral-equation-based framework derives coarse-grained interactions directly from liquid-state theory. Chains may be represented as one soft sphere, a small number of blobs, or many blobs, with

$\Psi^*(x,\xi)$6

and the coarse potential obtained from Ornstein–Zernike and PRISM relations rather than iterative fitting (McCarty et al., 2014). The paper states that the resulting model preserves structural correlations appropriate to the chosen mapping, the thermodynamic state, and the excess free energy across resolutions. In the soft-particle limit, the effective interaction becomes a free energy in the coarse-grained coordinates, and the compressibility is invariant under coarse-graining (McCarty et al., 2014).

For double-stranded DNA, the coarse-grained variables are again structural transformations rather than scalar elastic constants. Starting from rigid base-pair steps described by SE(3) junctions, the coarse model retains every $\Psi^*(x,\xi)$7-th base-pair frame and forms composite segment transforms $\Psi^*(x,\xi)$8 (Skoruppa et al., 2024). Static coarse-graining is exact at the group level, while fluctuation coarse-graining uses an approximate linear map built from transported rotational and translational generators. The output is a lower-resolution chain that preserves sequence-dependent geometry and Gaussian elasticity over roughly $\Psi^*(x,\xi)$9–$\hat\rho=\Pi_\#\rho$0 bp. The paper reports that rotational composite variables remain very nearly Gaussian up to $\hat\rho=\Pi_\#\rho$1, with variance errors under about $\hat\rho=\Pi_\#\rho$2, whereas translational variables, especially rise, are less accurate because the Gaussian coarse model cannot capture the left-skewness generated by bending fluctuations (Skoruppa et al., 2024). This suggests that structural coarse-graining can preserve intrinsic helicity and anisotropic elastic couplings at surprisingly coarse resolution, but not every translational statistic equally well.

A related but conceptually distinct example is the center-of-mass coarse-graining of ortho-terphenyl. There the aim is not to build a transferable force field but to expose the slow intermolecular structure relevant for density scaling. The exact coarse-grained interaction is the PMF

$\hat\rho=\Pi_\#\rho$3

approximated by force matching on center-of-mass coordinates (Jin et al., 2024). The resulting single-site OTP interactions are purely repulsive over the shown state points, the COM RDFs are reproduced very well, and the density-scaling exponent obtained from coarse-grained excess entropy is $\hat\rho=\Pi_\#\rho$4, compared with $\hat\rho=\Pi_\#\rho$5 from temporal coarse-graining (Jin et al., 2024). The paper argues that intramolecular vibrations and local roughness obscure hidden scale invariance at full atomistic resolution, whereas structural coarse-graining removes those fast modes and exposes a one-dimensional density-scaling picture.

5. Learned and transferable mappings

Structural coarse-graining is not always based on a fixed, hand-designed map. In an autoencoder-based formulation, the coarse coordinates are latent variables constrained to remain interpretable as weighted Cartesian averages,

$\hat\rho=\Pi_\#\rho$6

with an additional asymptotic requirement that each atom contribute to at most one coarse-grained variable (Wang et al., 2018). The mapping is learned jointly with a coarse-grained potential by minimizing a reconstruction loss,

$\hat\rho=\Pi_\#\rho$7

and an instantaneous force-matching term,

$\hat\rho=\Pi_\#\rho$8

combined as

$\hat\rho=\Pi_\#\rho$9

On the tested systems, the method recovers chemically intuitive or at least plausible bead definitions for OTP, methane, ethane, and aniline, while also showing that the learned map depends on the expressive power of the coarse potential class itself (Wang et al., 2018).

A different extension concerns transferability across chemical space. In a structure-based coarse-graining study of 3,441 C$x$0O$x$1 isomers, local chemical environments are represented by aSLATM descriptors, clustered, and summarized by 19 representative molecules whose environments overlap with more than 92% of the assigned aSLATM vectors in the full space (Kanekal et al., 2022). An extended-ensemble force-matching fit over 703 atomistic liquid and mixture state points, corresponding to 2,476 mapped ensembles, then yields a transferable model with 14 bead types and 105 pairwise nonbonded interactions (Kanekal et al., 2022). The paper reports that, across pure-liquid systems, the mean JSD is about 0.0024 for the extended-ensemble model versus 0.0038 for state-point-specific fits, and interprets this as a mean-force regularizer that smooths out force and structural correlations overly specific to a single state point (Kanekal et al., 2022). At the same time, the failure cases are structurally revealing: when one bead type merges environments such as carboxylic acids and esters, or when strongly directional interactions like hydrogen bonding and $x$2-stacking are collapsed into isotropic pair beads, transferability degrades (Kanekal et al., 2022).

These examples make two points explicit. First, structural coarse-graining can treat the map itself as an optimizable object. Second, transferability is not opposed to structure preservation in principle, but it depends strongly on whether the reduced representation respects genuine local-environment classes rather than only nominal chemical similarity (Wang et al., 2018, Kanekal et al., 2022).

6. Spatially adaptive and geometry-preserving multiresolution schemes

In several fields, structural coarse-graining is realized as adaptive retention of local detail rather than uniform variable elimination. For discrete element simulations, a concurrent multiresolution strategy embeds one or more fine-scale DEM subdomains inside a coarser DEM simulation of the full system, with two-way coupling by volumetric passing of boundary conditions (Queteschiner et al., 2017). Standard coarse-grained DEM keeps density and material parameters fixed while scaling particle radius by a factor $x$3,

$x$4

with stiffness coefficients scaling with $x$5 and damping coefficients with $x$6, but fails when the physics depends intrinsically on particle size, as in Beverloo discharge

$x$7

The adaptive method therefore resolves only critical regions, such as the silo outlet, and uses stress- and flux-based coupling between levels. In the reported silo discharge case, the fully coarse model with $x$8 underpredicts mass flow by more than 20%, whereas the two-level adaptive method gives $x$9, within $S(x)$0 of the $S(x)$1 reference, with a $S(x)$2 speedup (Queteschiner et al., 2017).

A related real-space idea appears in coarse-grained Kohn–Sham DFT. There the key move is first to eliminate explicit orbitals by expressing electron density, band energy, and entropy as spectral integrals evaluated by Gauss quadrature, and then to coarse-grain the defect-induced perturbation of local fields in space (Suryanarayana et al., 2012). The coarse reconstruction uses representative finite-difference nodes and a predictor-corrector decomposition,

$S(x)$3

with analogous formulas for pointwise band energy and entropy (Suryanarayana et al., 2012). Structurally, this means that both atomic geometry and electronic fields are retained at full resolution near defects and interpolated elsewhere, without changing the underlying Kohn–Sham theory. The paper reports, for sodium surface relaxation, that using representative nodes/total nodes $S(x)$4 and representative atoms/total atoms $S(x)$5 gives normalized errors of $S(x)$6 in surface energy and about $S(x)$7–$S(x)$8 in the first two layer displacements (Suryanarayana et al., 2012).

Concrete provides a homogenization-based variant. In Homogenization Coarse Graining of LDPM, one increases the actual size of the particles by a coarsening factor and calibrates the coarse parameters so that the macroscopic average response of a coarse RVE matches the fine one under multiple loading conditions (Lale et al., 2018). The coarse model preserves the same discrete mathematical structure as LDPM, including fracture and reinforcement interaction, rather than replacing it with a continuum law (Lale et al., 2018).

Geometry-preserving coarse-graining in chemistry uses yet another mechanism. Diverse tetrahedral $S(x)$9 frameworks are reduced to effective $\Psi^*(x,\xi)$0 and $\Psi^*(x,\xi)$1 sites, all structures are rescaled to a uniform minimum $\Psi^*(x,\xi)$2 distance, and local A-site environments are compared by SOAP kernels. The structural distance

$\Psi^*(x,\xi)$3

then defines a unified map for zeolites, ZIFs, ices, clathrates, and related materials (Nicholas et al., 2020). The method recovers known analogies such as the proximity of ZIF-8 and hydro-sodalite, distinguishes subtle geometry within shared topologies, and links map position to A-site heterogeneity and tetrahedral density (Nicholas et al., 2020). This suggests that structural coarse-graining can serve as a quantitative language for “nodes and linkers,” occupying a middle ground between atomistic chemistry and pure topology.

7. Networks, mesoscale structure, and persistent caveats

Outside molecular simulation, structural coarse-graining often identifies the effective building blocks directly from mesoscale organization. In colloidal physics, one dissertation proposes that the mesoscopic length-scale of structural inhomogeneities is the characteristic length-scale of the effective building blocks, while the degrees of freedom of the primary particles are integrated out (Zaccone, 2010). For sheared aggregating suspensions, the central coarse variable is an effective cluster volume fraction,

$\Psi^*(x,\xi)$4

which then enters a viscosity law

$\Psi^*(x,\xi)$5

(Zaccone, 2010). For arrested states, the same logic yields an inter-cluster modulus $\Psi^*(x,\xi)$6 governed by cluster diameter $\Psi^*(x,\xi)$7 and cluster volume fraction $\Psi^*(x,\xi)$8, rather than by the primary particle scale (Zaccone, 2010). Here structural coarse-graining is explicitly about identifying the stress-bearing scale in disordered matter.

In graph-based contagion reduction, the effective units are dense subgraphs. Iterative Structural Coarse-Graining repeatedly merges maximal cliques into weighted super-nodes and super-links, producing $\Psi^*(x,\xi)$9-clique coarse-grained networks

$$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$0

with exact preservation conditions for SIR contagion when

$$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$1

Under these conditions, the paper derives exact preservation of macroscopic outbreak size and contagion configuration probabilities between the original and coarse-grained networks (Xue et al., 2024). The structural premise is that a maximal clique is a contagion-coherent unit because, above threshold, internal infection becomes effectively certain.

A distinct network application uses structural diffusion geometry to reduce functional-connectivity estimation. Starting from a structural connectome with Laplacian $$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$2, the method computes a heat kernel

$$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$3

and groups nodes hierarchically according to diffusion-defined communication (Fernandez-Iriondo et al., 9 Feb 2026). Functional time series are then averaged within supernodes,

$$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$4

so that the reduced dimension $$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$5 becomes compatible with the available number of time points $$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$6, enabling Marchenko–Pastur filtering of noise-dominated modes (Fernandez-Iriondo et al., 9 Feb 2026). In the reported fMRI setting, $$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$7 and $$\mathrm{Prob}\bigl(\rho^{n,\varepsilon}\big|_{[0,T]}\approx \rho^\varepsilon\big|_{[0,T]}\bigr) \sim \exp\bigl[-n\,I^\varepsilon(\rho^\varepsilon)\bigr]$$8, so structural coarse-graining is used explicitly to move from a statistically underpowered regime to one in which reliable covariance estimation becomes possible (Fernandez-Iriondo et al., 9 Feb 2026).

These broader applications also sharpen the main caveats. Structural coarse-graining is representation-dependent; it may preserve one kind of structure while discarding another. Several papers state directly that closure is not automatic, that projected dynamics need not remain Markovian, that reduced variables must often be chosen in advance, and that effective interactions are frequently state-dependent or only approximate (Duong et al., 2014, Scherer et al., 2017, Jin et al., 2024, Dairyko et al., 2017). A plausible implication is that structural coarse-graining is best understood not as a single method but as a criterion for reduction: the reduced model is successful when the structure that organizes fluctuations, transport, mechanics, or topology at the target scale survives the projection in a quantitatively controlled form.