Pith. sign in

REVIEW 3 major objections 4 minor 39 references

Equivariant learning of a transferable three-dimensional classical density functional

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Equilibrium density fields alone can train a transferable three-dimensional liquid free-energy functional.

desk verdict A well-built paper that convincingly demonstrates a transferable 3D learned excess free-energy functional; the main open issue is the finite receptive field, which deserves an ablation but does not sink the central claim. read the letter →

arxiv 2608.13506 v1 pith:IHZ5F5AL submitted 2026-08-13 cond-mat.stat-mech cond-mat.softcs.LGphysics.chem-phphysics.comp-ph

Bingqing Cheng This is my paper · ORCID
classification cond-mat.stat-mechcond-mat.softcs.LGphysics.chem-phphysics.comp-ph
keywords classicaldensityfunctionaltheorymachine-learnedequivariantneuralnetworkfree-energychemical-potentialbalanceLennard-Jonesfluidsolvationforceadsorptionisotherm
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that the excess free-energy functional of a three-dimensional liquid can be learned directly from equilibrium one-body density fields, without free-energy or chemical-potential labels, by enforcing local chemical-potential balance. It demonstrates this for the truncated-and-shifted Lennard-Jones fluid and shows that one learned functional transfers across temperature, system size, and statistical ensemble. From that same functional, quantities that never entered the loss emerge: structure factors, the compressibility-route equation of state, liquid-vapor coexistence, interfacial broadening, the non-monotonic force between two colloids, and adsorption in a gyroid pore. The point is that equilibrium density data become a reusable thermodynamic generator rather than a fitted mapping for a single observable.

What carries the argument

The central object is the local excess free-energy map $a_{\mathrm{exc}}(\chi_g, T)$, shared across all grid points, giving $F_{\mathrm{exc}}[\rho,T] = \Delta V \sum_g \rho_g a_{\mathrm{exc}}(\chi_g,T)$. Each local environment $\chi_g$ is encoded by Cartesian moments $\mathbf{A}_\ell$ symmetrized over the 48 operations of the cubic point group $O_h$ to form invariant features $B_K^{(\nu)}$. Training uses automatic differentiation to obtain $c^{(1)}_g = -\frac{\beta}{\Delta V}\partial F_{\mathrm{exc}}/\partial \rho_g$ and enforces that $\mu^{\mathrm{loc}}_g = \ln(\Lambda^3\rho_g)+\beta V_{\mathrm{ext},g}-c^{(1)}_g$ is spatially constant, with the unknown canonical chemical potential eliminated analytically as its spatial mean.

What would settle it

Simulate a dense truncated-and-shifted Lennard-Jones fluid in an external potential modulated on a length scale between $1.5\sigma$ and $2.5\sigma$ (for example, a sinusoidal wall with period $2\sigma$), then compare the density from fixed-$N$ cDFT minimization with grand-canonical Monte Carlo or molecular dynamics.

Watch

Extended reading notes

Core claim

The central claim is that a shared, symmetry-adapted local map from density environment and temperature to excess free energy per particle, summed over the grid, is sufficient to represent $F_{\mathrm{exc}}[\rho,T]$ for a three-dimensional fluid. Trained by minimizing the spatial variance of the local chemical potential, with the unknown constant chemical potential removed analytically as a spatial mean, the functional reproduces equilibrium structure and thermodynamics at held-out temperatures and in larger cells. Because the one-body and two-body direct correlation functions are derivatives of the same scalar, response functions, pressure, coexistence, and interfacial profiles remain mutually consistent. The resulting equilibrium densities also predict solvent-mediated forces and adsorption in geometries entirely absent from training.

Load-bearing premise

The excess free energy at a point depends only on the density within $1.5\sigma$ (three grid spacings), while the pair potential extends to $2.5\sigma$, so correlations between $1.5\sigma$ and $2.5\sigma$ must be captured indirectly.

Editorial extensions

If this is right

  • At held-out temperatures, the functional inverts density to external field and external field to density, with errors small enough for quantitative use.
  • The same functional gives static structure factors and compressibility-route equations of state, including van der Waals loops at subcritical temperatures.
  • Slab minimization yields liquid-vapor coexistence densities and interfacial broadening, with a mean-field extrapolated critical point near the direct-simulation estimate.
  • The solvent-mediated force between two colloids is recovered from the minimized density alone, including the attraction, repulsive maximum, and decay to zero.
  • Adsorption in a gyroid pore follows from matching bulk and confined chemical potentials, with only one global additive gauge per temperature.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Because the functional is an extensive scalar, the same training loss should extend to mixtures by treating each species density as an input channel and coupling channels through CACE products; no new loss term is needed beyond per-species local balance.
  • Outside the paper, one could use the learned functional's derivative as the adiabatic driving force in dynamical density functional theory; the equilibrium claim would then be tested by whether time-dependent density evolution matches Brownian dynamics.
  • The $1.5\sigma$ receptive field is the natural stress point: a systematic test with external potentials modulated between $1.5\sigma$ and $2.5\sigma$ would show where the local approximation must be extended with nonlocal features.
  • For ionic or polar fluids, a long-ranged electrostatic branch would be needed; the local functional could be augmented by reciprocal-space features with Coulomb kernels, analogous to latent Ewald summation.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This paper proposes Equi-cDFT, a machine-learned excess free-energy functional for a three-dimensional classical fluid. The functional is expressed as a grid sum of a shared local free energy per particle, with local environments encoded by cubic-symmetry-adapted CACE invariants and temperature as an input. It is trained on canonical MD density fields using local chemical-potential balance, eliminating the unknown constant chemical potential analytically, and the same scalar is then differentiated to obtain one- and two-body direct correlation functions. Benchmarks include held-out forward (ρ→V_ext) and inverse (V_ext→ρ) reconstruction at three temperatures, transfer to larger cells, transfer from canonical to grand-canonical chemical-potential differences, structure factors, compressibility-route equation of state, liquid-vapor coexistence and interface profiles, solvent-mediated colloid-bridging forces, and gyroid-pore adsorption isotherms. The central claim is that equilibrium density data alone can be converted into a transferable thermodynamic generator connecting liquid structure to response, phase behavior, and collective phenomena.

Significance. The work is significant. If the claims hold, it is the first learned classical density functional operating directly on fully three-dimensional density fields with variational consistency and explicit equivariance, and it demonstrates emergent prediction of properties not used in training. The training signal is a legitimate inverse-problem formulation of the Euler-Lagrange equation rather than fitting target observables, and the extensive held-out tests (including S(k), EOS, coexistence, solvation forces, and adsorption) are appropriate and demanding. The paper reports public code and data, and the limitation statements about fixed discretization and near-criticality are candid. The main unresolved issues concern the adequacy of the finite local receptive field and the discrete cubic symmetry for an isotropic liquid, which the benchmarks do not directly isolate.

major comments (3)
  1. [§II.A, Eq. (2); Methods A] The local receptive field is three grid spacings, i.e. 1.5σ at the training grid spacing, which is shorter than the LJTS pair potential cutoff of 2.5σ in Eq. (16). By Eq. (10), c^(2)(r_g,r_g') is exactly zero for separations greater than 1.5σ, so the attractive tail between 1.5σ and 2.5σ is not represented by any explicit pair kernel. The paper explicitly calls the finite receptive field a 'modeling approximation' but provides no ablation or quantitative test of its adequacy. Since transferability and emergent thermodynamics are the central claims, I request an ablation (for example retraining with qcut = 4 or 5 grid spacings, or a spectral analysis of the omitted band) or an explicit argument why the omitted band is redundant given the local density environment. Without this, the conclusion that the learned functional is a general thermodynamic generator is stronger than demonstrated.
  2. [§V.C and §V.D, Fig. 2c] The training loss and all reported metrics mask out voxels with ρ < 10^-3σ^-3 (Methods C, D1, D3), leaving the local free-energy map unconstrained in exactly the dilute regime that determines the low-density vapor branch, coexistence plateau densities, and interface tails. The liquid-vapor coexistence in Fig. 2c uses fixed-N slab solutions whose vapor phase is near or below this threshold at the lowest temperatures. Please report the sensitivity of coexistence densities and interface profiles to the mask threshold, or otherwise demonstrate that the low-density extrapolation is controlled.
  3. [§II.A, Eqs. (4)-(5); Figs. 3-4] The symmetry statement is limited to the discrete cubic point group O_h, while the physical fluid is isotropic. Since a transferable functional should be approximately invariant under continuous rotations, the absence of any non-O_h rotated test is a gap: every benchmark external field and application geometry appears aligned with the grid axes (colloid separation along x in Fig. 3; gyroid axes in Fig. 4), and the randomized training fields use axis-aligned one- and two-dimensional components or isotropic three-dimensional Gaussians. Please add a rotated-field test, for example applying a held-out V_ext rotated by an angle not in O_h and comparing the predicted density to MD, or report the magnitude of cubic-lattice anisotropy in the learned functional, such as the orientation dependence of the colloid force.
minor comments (4)
  1. [§III.B, Fig. 3, §V.F] The words 'colliods' and 'colliod' should be 'colloids' in several places, including the Fig. 3 caption and Section V.F.
  2. [§V.B] The thermostat name is garbled by accent encoding as 'Nos´ e–Hoover'; it should read 'Nosé–Hoover'.
  3. [Fig. 3c] The abbreviation 'HF' is used in the Fig. 3c legend without being defined in the text; please define it as the Feynman–Hellmann evaluation.
  4. [Methods G] The reference 'Fig. 4a– Fig. 4c' mixes range notation inconsistently; use 'Fig. 4a–c' for uniformity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the learned functional is constrained by the cDFT Euler-Lagrange balance, and the claimed predictions are derivatives or minimizers of the resulting scalar rather than refitted training targets.

full rationale

The paper's training signal (Eq. 8) is the local chemical-potential balance condition, which is the exact Euler-Lagrange equation of the excess functional; fitting c^(1) this way is a legitimate inverse problem and does not put the benchmark observables into the loss. The quantities presented as predictions are obtained post hoc from the learned scalar: S(k) from the Hessian c^(2) and the Ornstein-Zernike relation (Eqs. 10-11), the equation of state from compressibility integration (Eqs. 12-13), coexistence and interfacial profiles from fixed-N minimization, the colloid force from the Feynman-Hellmann integral over the minimized density (Eq. 14), and adsorption from matching independently computed bulk and confined chemical-potential branches. None of these targets is a fitted label; the only fitted constants are the network parameters and the physically forced b(T)N gauge of Eq. (9), which the paper explicitly calibrates per temperature and which cancels in fixed-N and c^(2) predictions. The finite receptive field (Methods A, cutoff 1.5 sigma) is a stated modeling approximation, not a circular reduction, and the held-out and larger-cell benchmarks directly test it. Self-citations to CACE (Ref. 27) and latent Ewald summation (Ref. 34) are methodological tools with published, independent derivations and are not used as evidence for the central transferability claim. No circular step can be exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a local, cubic-symmetry-adapted representation of F_exc trained from canonical MD densities. No new physical entities are postulated. The main fitted quantities are canonical chemical-potential gauges and the mean-field critical extension, both disclosed. The locality and cubic-symmetry assumptions carry the largest burden.

free parameters (3)
  • Per-temperature additive chemical-potential gauge b(T) = Not reported; one constant per temperature
    Canonical training leaves the gauge freedom F_exc -> F_exc + b(T)N (Eq. 9). For absolute chemical potentials, one additive constant per temperature is fitted, used for GCMC transfer (Fig. 5d) and gyroid calibration (Methods G). It does not affect fixed-N densities, c(2) or coexistence.
  • Critical-point continuation coefficients (A, B, T_c, rho_c) = T_c=1.09, rho_c=0.31 sigma^-3
    Coexistence width squared and diameter from five highest-temperature slab solutions are fitted to linear forms to estimate the critical point. The paper states this is a mean-field continuation, not an independently solved state.
  • Per-field gauge constant for forward external-field reconstruction = Not reported; one per test field
    In Eq. (17), a least-squares constant equal to the active-voxel mean of reference minus prediction is added per held-out field. This aligns the unknown canonical chemical potential in evaluation only and is not a functional parameter.
assumptions (5)
  • domain assumption Locality of the excess functional: F_exc = Delta V sum_g rho_g a_exc(chi_g,T) with finite cutoff of three grid spacings
    Eq. (2) and Methods A. The shared local map is the transfer mechanism. If fluid correlations beyond 1.5 sigma were essential, the approximation would fail. It is tested by larger-cell and 3D benchmarks but is not derived.
  • domain assumption Cubic point-group symmetry, not continuous rotational symmetry
    The features in Eqs. (3)-(5) are symmetrized over the 48 operations of O_h. Arbitrary rotations of a physical density that are not grid rotations are not symmetry operations of the learned functional.
  • standard math Equilibrium MD density fields satisfy the local chemical-potential balance condition
    Eqs. (7)-(8) are the Euler-Lagrange condition of cDFT for equilibrium densities. This is standard statistical mechanics, not an additional physical assumption.
  • standard math Ornstein-Zernike relation and compressibility route connect c(2) to S(k) and the equation of state
    Eqs. (11)-(13) are standard liquid-state theory used to convert the learned second derivative into structure and thermodynamics.
  • domain assumption Training corpus is representative of the liquid states of interest
    10,957 fields over T=0.625-1.8 and whole-box densities 0.02-0.91 with randomized Gaussian external fields. Generalization to colloid and gyroid geometries depends on the diversity of local environments in this corpus.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Equivariant learning of a transferable three-dimensional classical density functional." pith.science (2026). https://pith.science/paper/IHZ5F5AL

@misc{pith2026260813506,
  author       = {Pith},
  title        = {Pith review of: Equivariant learning of a transferable three-dimensional classical density functional},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IHZ5F5AL}},
  note         = {Machine review of arXiv:2608.13506}
}
read the original abstract

Liquids exhibit collective behavior that depends sensitively on thermodynamic conditions, interfaces and confinement, yet predicting each new state commonly requires a separate atomistic simulation. Classical density functional theory offers a reusable variational description, but its central excess free-energy functional is generally unknown, and learned approximations have largely remained restricted to planar or lower-dimensional settings. Here we show that this functional can be learned directly from fully three-dimensional equilibrium density fields while preserving spatial symmetry and variational consistency, without free-energy or chemical-potential labels. A single learned functional transfers across temperatures, system sizes and statistical ensembles, and recovers structure factors, the equation of state, liquid--vapor coexistence and interfacial broadening, none of which are used as training targets. Applied to complex three-dimensional geometries, it predicts the non-monotonic force associated with formation and rupture of a solvent-depleted bridge between colloids and adsorption in an interconnected gyroid pore. These results demonstrate that equilibrium density data can be converted into a transferable thermodynamic generator connecting microscopic liquid structure to response, phase behavior and collective phenomena.

Figures

Figures reproduced from arXiv: 2608.13506 by the authors.

Figure 1
Figure 1. Equi-cDFT architecture. Local neighborhoods of the three-dimensional density field are encoded using cubic￾symmetry-adapted features and processed by a shared neural readout. Summing the local contributions produces an extensive excess free-energy functional, whose derivative gives the one-body direct correlation field. Here g labels a density-grid point, χg is its finite spherical environment, R ∈ Oh denotes a rota… view at source ↗
Figure 2
Figure 2. Equation of state, phase behavior, and interfaces from the learned free-energy functional. a Static structure factors of homogeneous fluids. Solid curves and shaded standard errors are direct MD estimates; dashed curves with open circles follow from twice differentiating the learned functional and applying the Ornstein–Zernike relation. b Pressure– density isotherms. Dashed curves are cDFT compressibility-route pred… view at source ↗
Figure 3
Figure 3. Solvent-mediated interactions between two fixed colloids. a Schematic of two colloids at center-to-center separation D immersed in a one-component Lennard–Jones fluid. Thin blue contours mark the colloid–fluid attractive regions; the colloids have no direct mutual interaction. Red arrows denote the mean force ⟨F⟩D, with negative values corresponding to attraction and positive values to repulsion. b Central 1σ-thick … view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Gyroid-pore structure and adsorption at T = 1.2. a,b Equilibrium MD and cDFT density fields at the matched isosurface levels, for whole-box mean density ⟨ρ⟩σ 3 = 0.29. c Grid-point density parity, colored by the minimum periodic distance dwall to the nominal pore wall …
Figure 5
Figure 5. Figure 5: Combined benchmark of the learned excess free-energy functional. a Gauge-aligned external fields inferred from all 1,592 held-out NVT equilibrium density fields; one spatially constant gauge is aligned for each canonical field. b Fixed-particle-number equilibrium densi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 35 canonical work pages

  1. [1]

    The held-out test set contains 1,592 complete fields inL= 8σcells on 16 3 grids with spacing 0.5σ, spanningN= 8–464

    Held-out forward and inverse tests We first test the functional in both directions on NVT fields atT= 0.7, 1.1 and 1.5, temperatures omitted en- tirely from training and validation. The held-out test set contains 1,592 complete fields inL= 8σcells on 16 3 grids with spacing 0.5σ, spanningN= 8–464. For the forward test, every held-out density field is supp...

    work page

  2. [2]

    All 30 fields at each temperature are evaluated, giving 90 fields in total

    Transfer to larger canonical systems To test locality and extensivity across system size, the model trained onL= 8σfields is applied toL= 12σ boxes discretized on 24 3 grids at the same three held- out temperatures. All 30 fields at each temperature are evaluated, giving 90 fields in total. They spanN= 32– 1568 and whole-box mean densitiesρσ 3 = 0.02–0.91...

    work page

  3. [3]

    TheL= 8σ, 16 3 GCMC test set contains 22 complete fields at each ofT= 1.0, 1.2 and 1.4, giving 66 fields in total

    Grand-canonical chemical-potential differences Finally, we assess transfer from canonical training to grand-canonical chemical-potential differences. TheL= 8σ, 16 3 GCMC test set contains 22 complete fields at each ofT= 1.0, 1.2 and 1.4, giving 66 fields in total. The imposed reservoir chemical potentials spanµ/ϵ=−4 to 2. For fielda, the raw prediction is...

    work page

  4. [4]

    Hohenberg and W

    P. Hohenberg and W. Kohn, Inhomogeneous electron gas, Physical Review136, B864 (1964)

    1964

  5. [5]

    Kohn and L

    W. Kohn and L. J. Sham, Self-consistent equations in- cluding exchange and correlation effects, Physical Review 140, A1133 (1965)

    1965

  6. [6]

    Evans, The nature of the liquid-vapour interface and other topics in the statistical mechanics of non-uniform, classical fluids, Advances in Physics28, 143 (1979)

    R. Evans, The nature of the liquid-vapour interface and other topics in the statistical mechanics of non-uniform, classical fluids, Advances in Physics28, 143 (1979)

    work page 1979

  7. [7]

    Rosenfeld, Free-energy model for the inhomogeneous hard-sphere fluid mixture and density-functional theory of freezing, Physical Review Letters63, 980 (1989)

    Y. Rosenfeld, Free-energy model for the inhomogeneous hard-sphere fluid mixture and density-functional theory of freezing, Physical Review Letters63, 980 (1989)

    work page 1989

  8. [8]

    Roth, Fundamental measure theory for hard-sphere mixtures: a review, Journal of Physics: Condensed Mat- ter22, 063102 (2010)

    R. Roth, Fundamental measure theory for hard-sphere mixtures: a review, Journal of Physics: Condensed Mat- ter22, 063102 (2010)

    work page 2010

Show all 39 references
  1. [9]

    Evans, M

    R. Evans, M. Oettel, R. Roth, and G. Kahl, New devel- opments in classical density functional theory, Journal of Physics: Condensed Matter28, 240401 (2016)

    2016

  2. [10]

    Simon and M

    A. Simon and M. Oettel, Machine learning approaches to classical density functional theory, inArtificial Intel- ligence and Intelligent Matter, Machine Intelligence for Materials Science, edited by M. te Vrugt (Springer Na- ture Switzerland, Cham, 2026) pp. 83–113

    2026

  3. [11]

    Lin and M

    S.-C. Lin and M. Oettel, A classical density functional from machine learning and a convolutional neural net- work, SciPost Physics6, 025 (2019)

    2019

  4. [12]

    P. Cats, S. Kuipers, S. de Wind, R. van Damme, G. M. Coli, M. Dijkstra, and R. van Roij, Machine-learning free-energy functionals using density profiles from sim- ulations, APL Materials9, 031109 (2021)

    2021

  5. [13]

    M. M. Kelley, J. Quinton, K. Fazel, N. Karimitari, C. Sutton, and R. Sundararaman, Bridging electronic and classical density-functional theory using universal machine-learned functional approximations, The Journal of Chemical Physics161, 144101 (2024)

    2024

  6. [14]

    A. T. Bui and S. J. Cox, Learning classical density func- tionals for ionic fluids, Physical Review Letters134, 148001 (2025)

    2025

  7. [15]

    A. T. Bui and S. J. Cox, A unified machine-learning framework for ab initio multiscale modeling of liquids, Proceedings of the National Academy of Sciences123, e2610049123 (2026)

    2026

  8. [16]

    Samm¨ uller and M

    F. Samm¨ uller and M. Schmidt, Neural density function- als: Local learning and pair-correlation matching, Phys- ical Review E110, L032601 (2024)

    2024

  9. [17]

    Dijkman, M

    J. Dijkman, M. Dijkstra, R. van Roij, M. Welling, J.-W. van de Meent, and B. Ensing, Learning neural free-energy functionals with pair-correlation matching, Physical Re- view Letters134, 056103 (2025)

    2025

  10. [18]

    K. Ram, J. Dijkman, R. van Roij, J.-W. van de Meent, B. Ensing, M. Welling, and D. Cremers, Learned free- energy functionals from pair-correlation matching for dy- namical density functional theory, Physical Review E 112, 045314 (2025)

    2025

  11. [19]

    R. Pan, X. Fang, K. Azizzadenesheli, M. Liu-Schiaffini, M. Gu, and J. Wu, Neural operators for forward and inverse potential-density mappings in classical density functional theory, The Journal of Chemical Physics163, 164120 (2025)

    2025

  12. [20]

    J. Yang, R. Pan, J. Sun, and J. Wu, High-dimensional operator learning for molecular density functional theory, Journal of Chemical Theory and Computation21, 5905 (2025)

    2025

  13. [21]

    Samm¨ uller, S

    F. Samm¨ uller, S. Hermann, D. de las Heras, and M. Schmidt, Neural functional theory for inhomogeneous fluids: Fundamentals and applications, Proceedings of the National Academy of Sciences120, e2312484120 (2023)

    2023

  14. [22]

    Samm¨ uller, M

    F. Samm¨ uller, M. Schmidt, and R. Evans, Neural density functional theory of liquid-gas phase coexistence, Physi- cal Review X15, 011013 (2025)

    2025

  15. [23]

    S. M. Kampa, F. Samm¨ uller, M. Schmidt, and R. Evans, Metadensity functional theory for classical fluids: Ex- tracting the pair potential, Physical Review Letters134, 107301 (2025)

    2025

  16. [24]

    Robitschko, F

    S. Robitschko, F. Samm¨ uller, M. Schmidt, and R. Evans, Learning the bulk and interfacial physics of liquid-liquid phase separation with neural density functionals, The Journal of Chemical Physics163, 161101 (2025)

    2025

  17. [25]

    Samm¨ uller and M

    F. Samm¨ uller and M. Schmidt, Determining the chem- ical potential via universal density functional learning, Physical Review Letters136, 068202 (2026)

    2026

  18. [26]

    Glitsch, J

    F. Glitsch, J. Weimar, and M. Oettel, Neural density functional theory in higher dimensions with convolutional layers, Physical Review E111, 055305 (2025)

    2025

  19. [27]

    S. M. Kampa, M. Schmidt, and F. Samm¨ uller, Spher- ical metadensity functional learning for inhomogeneous classical fluids, arXiv preprint arXiv:2606.14370 (2026)

    2026

  20. [28]

    M. Thol, G. Rutkai, R. Span, J. Vrabec, and R. Lustig, Equation of state for the Lennard-Jones truncated and shifted model fluid, International Journal of Thermo- physics36, 25 (2015)

    2015

  21. [29]

    Vrabec, G

    J. Vrabec, G. K. Kedia, G. Fuchs, and H. Hasse, Com- prehensive study of the vapour–liquid coexistence of the truncated and shifted Lennard-Jones fluid including pla- nar and spherical interface properties, Molecular Physics 104, 1509 (2006)

    2006

  22. [30]

    Cheng, Cartesian atomic cluster expansion for ma- chine learning interatomic potentials, npj Computational Materials10, 157 (2024)

    B. Cheng, Cartesian atomic cluster expansion for ma- chine learning interatomic potentials, npj Computational Materials10, 157 (2024)

    2024

  23. [31]

    Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Physical Review B 99, 014104 (2019)

    R. Drautz, Atomic cluster expansion for accurate and transferable interatomic potentials, Physical Review B 99, 014104 (2019)

    2019

  24. [32]

    Evans and U

    R. Evans and U. Marini Bettolo Marconi, Phase equilib- ria and solvation forces for fluids confined between par- allel walls, The Journal of Chemical Physics86, 7138 (1987)

    1987

  25. [33]

    R. P. Feynman, Forces in molecules, Physical Review56, 340 (1939)

    1939

  26. [34]

    C. Qiao, X. Yu, X. Song, T. Zhao, X. Xu, S. Zhao, and K. E. Gubbins, Enhancing gas solubility in nanopores: A combined study using classical density functional theory and machine learning, Langmuir36, 8527 (2020). 14

    2020

  27. [35]

    E. d. A. Soares, A. G. Barreto Jr., and F. W. Tavares, Classical density functional theory reveals structural in- formation of H2 and CH4 fluids adsorbed in MOF-5, Fluid Phase Equilibria574, 113887 (2023)

    2023

  28. [36]

    Stierle, G

    R. Stierle, G. Bauer, N. Thiele, B. Bursik, P. Rehner, and J. Gross, Classical density functional theory in three dimensions with GPU-accelerated automatic differenti- ation: Computational performance analysis using the example of adsorption in covalent-organic frameworks, Che...

    2024

  29. [37]

    Cheng, Latent ewald summation for machine learning of long-range interactions, npj Computational Materials 11, 80 (2025)

    B. Cheng, Latent ewald summation for machine learning of long-range interactions, npj Computational Materials 11, 80 (2025)

    2025

  30. [38]

    J. A. Keith, V. Vassilev-Galindo, B. Cheng, S. Chmiela, M. Gastegger, K.-R. M¨ uller, and A. Tkatchenko, Com- bining machine learning and computational chemistry for predictive insights into chemical systems, Chemical re- views121, 9816 (2021)

    2021

  31. [39]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Kopf, E. Yang, Z. DeVito, M. Rai- son, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala, Pytorch: An imperative style, high-p...

    2019

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.