Maximally Learnable Coarse Graining

• Maximally learnable coarse graining is a framework that evaluates coarse representations based on statistical identifiability, data efficiency, and physical interpretability.
• It employs methods like spectral matching, uncertainty-aware sampling, and optimized mapping selection to preserve long-range predictive information and slow kinetics.
• The approach integrates thermodynamic consistency with representational fidelity, ensuring that both equilibrium properties and dynamic behaviors are accurately modeled.

Maximally learnable coarse graining denotes a research perspective in which a coarse representation is judged not only by thermodynamic consistency or dimensional reduction, but by whether it renders the reduced equilibrium or dynamical description statistically identifiable, data-efficient, physically interpretable, and compatible with the inductive biases of practical learning architectures. Taken together, recent work suggests that this perspective spans several adjacent criteria: preserving long-range predictive information, concentrating metastable multimodality in low-dimensional variables while making unresolved fluctuations conditionally simple, retaining slow spectral content, reducing force-label noise, and designing mappings whose symmetries and geometric signatures remain compatible with modern machine-learning force fields (Gökmen et al., 2021, Stupp et al., 29 Apr 2025, Nüske et al., 2019, Brunken et al., 24 Mar 2025).

1. Scope and conceptual status

The phrase is not used uniformly across the literature. Instead, it functions as an overview over several research programs that ask what makes a coarse description easy to infer, simulate, and generalize. In molecular coarse-graining, one recurrent distinction is between mapping selection and fixed-map learnability. Several works explicitly improve learning for a chosen mapping without claiming to solve the mapping problem itself, while others propose criteria that bear directly on variable selection.

At equilibrium, a broad statistical-mechanical backbone is provided by a generalized mean-field variational framework in which inverse problems and reduction problems are expressed through observable matching, free-energy stationarity, reduced-distribution matching, and relative-entropy interpretations. In that framework, reverse Monte Carlo, covariance-based coarse-graining, multiscale coarse-graining, relative-entropy coarse-graining, and classical density functional theory appear as particular cases (Larini et al., 2010). This suggests that maximally learnable coarse graining is not a single algorithmic recipe, but a criterion for choosing observables, mappings, and model classes so that equilibrium information can be reproduced with minimal ambiguity.

A broader analog appears outside molecular simulation. In decoherent-histories quantum mechanics, realistic quasiclassical realms are characterized as strongly decoherent, adaptive branch dependent coarse grainings of quasiclassical variables that are “maximally refined consistent with all these properties,” namely decoherence, classicality, narrative structure, and record formation in an environment (Gell-Mann et al., 2013). Although the setting is different, the underlying principle is closely aligned: refinement is desirable only to the extent that stable prediction and robust distinguishability are preserved.

2. Formal criteria for learnable coarse variables

Different communities formalize learnability through different objectives. The following criteria recur most clearly.

Framework	Core criterion	Learnability implication
RSMI-based real-space coarse graining	maximize $I_\Lambda(\mathcal H:\mathcal E)$	preserve long-range predictive information (Gökmen et al., 2021)
Predictive coarse-graining	minimize $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$	latent coarse variables should generate fine data well (Schöberl et al., 2016)
Spectral matching	$\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$ and match leading eigenvalue equations	preserve slow, predictable kinetics (Nüske et al., 2019)
Flow-based energy-based coarse-graining	$p_\phi(z)$ multimodal, $p_\phi(X\mid z)$ unimodal	isolate metastable structure while making unresolved variables easy to sample (Stupp et al., 29 Apr 2025)
RG-motivated pairwise compression	minimize projection error or use $I(X;Y)-\beta I(X^c;Y)$	preserve reconstructable or environment-relevant structure (Landy et al., 2023)

In the RSMI formulation, a local variable $\mathcal H$ is optimal when it retains as much as possible of the information that a visible block $\mathcal V$ shares with a distant environment $\mathcal E$, yielding the variational principle $$\Lambda^\star=\arg\max_\Lambda I_\Lambda(\mathcal H:\mathcal E)$$ (Gökmen et al., 2021). In critical systems, the formal solutions are tied to relevant operators, so learnability is linked to preserving long-distance predictive structure rather than merely compressing local data.

Predictive coarse-graining replaces a deterministic fine-to-coarse restriction by a probabilistic coarse-to-fine map $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$0 and a coarse prior $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$1, giving the induced fine marginal $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$2. Learning then minimizes $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$3, while posterior predictive distributions quantify uncertainty from both information loss and finite data (Schöberl et al., 2016). A plausible implication is that a maximally learnable coarse variable is one that remains strongly predictive of fine observables under this latent generative model.

Energy-based flow models pursue a related but distinct criterion. A good coarse representation is one in which slow variables $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$4 retain the multimodal marginal structure associated with metastability, while the complementary variables $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$5 have a unimodal conditional distribution $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$6 that is easy to sample (Stupp et al., 29 Apr 2025). In the demonstrated implementation, the map is fixed rather than learned, but the criterion still states a concrete version of learnability: the unresolved conditional should be statistically simple once the coarse state is known.

Spectral matching introduces a kinetic criterion. A mapping is favorable when the dominant slow eigenfunctions of the fine dynamics are approximately representable as functions of the coarse variables, $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$7, so that a reduced generator can match the leading eigenvalue equations of the full system (Nüske et al., 2019). This suggests that maximally learnable variables are those that retain the most persistent and therefore most predictable dynamical structure.

3. Fixed-map interventions: data, labels, and sampling policy

A substantial part of the literature shows that learnability can be improved dramatically even when the coarse map is fixed. The key interventions are changes in training distribution, force aggregation, uncertainty-guided data acquisition, and representation learning.

Enhanced sampling for force matching is the clearest data-side example. For a fixed coarse representation $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$8, one may bias atomistic sampling along coarse variables, recompute forces with respect to the unbiased atomistic potential, and train on the biased configurations without reweighting, provided the bias depends only on the coarse variables. The central invariance statement is that the conditional mean force $\mathrm{KL}(p_f(x)\,\|\,\bar p_f(x\mid\theta))$9 is unchanged under such a bias, while coverage of rare and transition regions improves markedly (Chen et al., 13 Oct 2025). In the Müller–Brown and capped alanine benchmarks, this yields lower error, lower variance, and better recovery of metastable structure under finite simulation budgets.

Label quality can also be optimized. For a fixed configurational map $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$0, the atomistic-to-CG force map is generally not unique. Valid force maps satisfy a compatibility condition with the coordinate map and an orthogonality condition with respect to constraints, and the paper on statistically optimal force aggregation shows that one can choose among them by minimizing $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$1. This reduces mapped-force variance, lowers stochastic-gradient variance, and improved efficiency by approximately a factor of 3 over basic forces in the protein benchmarks (Krämer et al., 2023). In this sense, maximally learnable coarse graining includes the supervision operator, not only the retained coordinates.

Uncertainty-aware active learning introduces a sampling-policy dimension. A sparse Gaussian-process coarse model can estimate local PMF uncertainty and trigger new atomistic labeling only when the uncertainty exceeds a threshold. In pentane-to-octane transfer, adapted models required far fewer octane frames than direct training while matching or improving force accuracy and, in the two-species case, improving population-ratio error (Duschatko et al., 2022). This suggests that learnability depends not only on the map and model, but on whether the training loop can identify which coarse configurations still require atomistic information.

Representation learning matters as well. CGSchNet replaces manually specified distances, angles, and torsions by learned graph features derived from continuous-filter convolutions, while preserving force-matching thermodynamic consistency through an energy-based scalar-output architecture (Husic et al., 2020). The work does not optimize the coarse map itself, but it shows that a fixed representation becomes more learnable when the model class is symmetry aware, graph based, and transferable across molecular systems.

4. Mapping design for machine-learning potentials

Recent work increasingly argues that mapping should be chosen with downstream learnability in mind. This is most explicit in studies where the mapping is tuned or audited relative to the representational assumptions of equivariant machine-learning potentials.

A central result is that mapping still matters even with expressive E(3)-equivariant models such as MACE and NequIP. In liquid hexane, amino acids, and polyalanine, mappings can create unphysical bond permutations when bonded and nonbonded length scales overlap, and neglecting species or stereochemistry introduces unphysical symmetries (Görlich et al., 8 Dec 2025). The practical message is stringent: a mapping is not maximally learnable merely because an exact PMF exists for it; it must also preserve the distinctions that a geometry-based equivariant model can actually resolve.

A constructive version of this idea appears in graph-based tunable coarse-graining. There, the atom-to-bead grouping is determined by spectral priorities on chemical graphs, combined as $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$2, and the hyperparameters are tuned to reduce the mean magnitude of projected CG forces, used as a proxy for force-target noise (Brunken et al., 24 Mar 2025). The reported Pearson correlations between mean CG force magnitude and training losses are $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$3, $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$4, and $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$5, supporting the view that lower-noise coarse forces make the downstream MACE learning problem easier. This is one of the clearest operational implementations of “choose the map to make the force-learning task more learnable.”

Autoencoder-based mapping discovery exposes the limits of reconstruction-only criteria. Variational-autoencoder-inspired models can learn discrete atom-to-bead assignments through Gumbel-Softmax variables and reconstruct atomistic coordinates from latent CG variables, but the resulting mappings are highly sensitive to random initialization, can break molecular connectivity, and fail to respect rotational invariance when Cartesian coordinates are used directly (Nasikas et al., 2022). The need for a post hoc connectivity-preserving acceptability criterion, and the observation that reconstruction quality degrades into averaged, overlapping structures on diverse datasets, show that maximal learnability cannot be equated with compression plus backmapping accuracy alone.

5. Kinetics, thermodynamic structure, and nonequilibrium learnability

A coarse representation can be easy to learn in equilibrium and still fail dynamically. Several frameworks therefore define learnability through retention of slow kinetics or by constraining the reduced dynamics to remain thermodynamically structured.

Spectral matching targets the low-lying spectrum of the generator rather than equilibrium forces alone. The objective fits a parametric coarse generator so that the leading eigenvalue equations are matched in weak form, and the method can either learn effective potentials or correct the dynamics induced by force matching through a position-dependent diffusion field (Nüske et al., 2019). In projection benchmarks and alanine dipeptide, this restores implied timescales and metastable structure more faithfully than thermodynamics-only training. The implication is that a maximally learnable mapping should allow the dominant slow eigenfunctions to remain observable in the coarse variables.

For nonequilibrium systems, a more radical answer is to augment the coarse state and restrict the hypothesis class. Data-driven particle dynamics in a metriplectic framework learns coarse-grained models of the form

$\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$6

with degeneracy conditions $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$7 and $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$8, latent entropy variables, and self-supervised identification of entropic state variables (Hernandez et al., 18 Aug 2025). The framework preserves discrete notions of the first and second laws of thermodynamics, conservation of momentum, and fluctuation-dissipation balance. This suggests that learnability can increase when the reduced state is expanded just enough to recover an approximately Markovian, thermodynamically closed dynamics.

A complementary limitation comes from stochastic thermodynamics. For a coarse-grained internal variable $\psi_i(x)\approx \tilde\psi_i^\xi(\xi(x))$9, the learning efficiency $p_\phi(z)$0 is bounded not only by the Clausius-type inequality $p_\phi(z)$1, but more tightly by

$p_\phi(z)$2

derived from a lower bound on the coarse-grained entropy production rate via the Cauchy–Schwarz inequality (Li et al., 2023). This gives maximally learnable coarse graining a literal thermodynamic content: the efficiency with which a retained variable can learn about its environment is itself bounded by the coarse-grained dissipative structure.

6. Tradeoffs, misconceptions, and open problems

Several misconceptions are addressed implicitly by the current literature. The first is that ever more expressive machine-learning potentials remove the mapping problem. They do not. Mapping can still introduce unphysical bond permutations, erase stereochemistry, collapse species distinctions, or render topology unrecoverable from local geometry, all of which make the PMF representation ambiguous to the architecture (Görlich et al., 8 Dec 2025). A second misconception is that reconstruction accuracy alone identifies good coarse variables. Autoencoder-based mapping studies show that reconstruction can remain acceptable while connectivity, invariance, and robustness are poor (Nasikas et al., 2022).

A third misconception is that improving force loss is sufficient. Enhanced sampling and active learning both show that transition regions and rare states are disproportionately important: force accuracy can improve monotonically while PMF-sensitive observables or basin populations remain underconstrained, and uncertainty estimates can be more informative than passive data accumulation about where atomistic effort is still needed (Chen et al., 13 Oct 2025, Duschatko et al., 2022). This suggests that learnability is fundamentally distribution dependent.

Open problems remain substantial. The most persistent one is mapping selection itself. Many influential methods assume the coarse map is fixed and only improve sampling, supervision, uncertainty quantification, or representation learning afterward. The flow-based energy-based framework is especially explicit: it proposes a strong criterion for a good coarse representation—multimodal $p_\phi(z)$3, unimodal $p_\phi(z)$4—but in the demonstrated experiments the bijection $p_\phi(z)$5 is fixed rather than learned, and learning $p_\phi(z)$6 is identified as the most important future direction (Stupp et al., 29 Apr 2025). Related unresolved questions include choosing the latent dimension, identifying collective variables in high-dimensional biomolecular systems, maintaining long-horizon stability in transferable ML coarse-grained force fields, and reconciling equilibrium consistency with kinetic and nonequilibrium fidelity.

Taken together, these works suggest that maximally learnable coarse graining is best viewed as a multi-criterion design problem. A favorable coarse description must retain the variables that matter for long-range prediction, metastability, or slow kinetics; suppress irrelevant fast fluctuations or force noise; remain compatible with architectural symmetries and locality assumptions; support uncertainty-aware acquisition of new labels; and, in nonequilibrium settings, often augment the state enough to restore a structured reduced dynamics. In that sense, the topic is less a single variational principle than a convergence of information-theoretic, statistical, thermodynamic, and representational criteria for deciding which coarse variables can actually be learned well.