{"id":"7ac635f5-f7fe-4531-a410-72cb154073f7","arxiv_id":"2608.13506","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Equi-cDFT learns a symmetry-preserving classical density functional from 3D equilibrium densities and uses it to predict thermodynamics, coexistence, solvation forces and adsorption in new geometries.","lead":"A new machine-learning method learns the excess free-energy functional of a three-dimensional liquid directly from equilibrium density data, without free-energy or chemical-potential labels. The learned functional transfers across temperatures, system sizes and ensembles, and predicts structure, phase coexistence, solvation forces and adsorption in geometries never seen in training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The finite receptive field (1.5 sigma) is shorter than the LJTS potential cutoff (2.5 sigma), leaving the transferability claim resting on an unablated local-range approximation.","rationale":"The central claim is that a functional learned from equilibrium one-body densities is a transferable thermodynamic generator. For that to be true, the learned F_exc must faithfully represent the physics of the LJTS fluid. The model's only channel for that physics is the local map a_exc(chi_g,T) with chi_g a 1.5 sigma stencil. Since F_exc determines c^(1) and c^(2) by differentiation, this imposes a hard range: c^(2)(r,r') = 0 for |r-r'| > 1.5 sigma, well inside the 2.5 sigma potential cutoff. The paper's benchmarks are extensive and, as reported, support transfer: held-out temperatures, larger cells, GCMC chemical potentials, structure factors, EOS, coexistence, interfacial broadening, colloid bridging, and gyroid adsorption all agree with reference data, and the implementation and datasets are public. These are real independent supports, and the reader's conditional verdict is fair. The residual risk is that all these tasks are consistent with the short-ranged kernel, so the approximation's failure mode, if any, has not been isolated. The paper itself flags the receptive field as an approximation and notes remaining sensitivities near criticality and at fixed resolution. A cutoff ablation is the direct, decisive check: if predictions are insensitive to extending the stencil to 2.5 sigma, the concern is retired; if not, the functional's transferability is narrower than claimed. Because the paper already discloses this limitation and the benchmarks are strong, I do not move the verdict; CONDITIONAL remains appropriate.","tokens_in":15632,"tokens_out":17219,"duration_ms":195853,"concrete_test":"Retrain the identical Equi-cDFT architecture on the same training split with stencil cutoff increased from 3 to 5 grid spacings (2.5 sigma, matching the LJTS cutoff), with 4 grid spacings as an intermediate point. Compare the headline emergent predictions---compressibility-route EOS, liquid-vapor coexistence and interface profiles, the colloid-bridging force curve, and the gyroid adsorption isotherm---against the published q_cut = 3 results using the existing MD/GCMC reference data and block-error bars. If all benchmarks shift by less than the reference uncertainties, the short receptive field is not load-bearing; if any shifts systematically (for example, coexistence densities or the repulsive force maximum), the functional's range must be enlarged before the transferability claim is accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (2) and Methods A approximate F_exc[rho,T] as a sum of local contributions rho_g a_exc(chi_g,T), where chi_g is the density in a spherical stencil of radius three grid spacings = 1.5 sigma at the training resolution. Because c^(1)_g (Eq. 6) and c^(2)_{g,g'} (Eq. 10) are derivatives of this sum, the learned direct correlation kernel is exactly zero for |r_g - r_{g'}| > 1.5 sigma. The LJTS potential (Eq. 16) has a 2.5 sigma cutoff, so the attractive tail between 1.5 sigma and 2.5 sigma is not represented by any explicit pair kernel; it can only be folded into shorter-range density features. All headline transfer tests (structure factors, EOS, coexistence, interface profiles, colloid bridging, gyroid adsorption) pass with this 1.5 sigma kernel, but no experiment isolates the contribution of the excluded 1.5-2.5 sigma band. The central claim that equilibrium density data yield a transferable thermodynamic generator therefore rests on the unproven adequacy of this locality/range assumption, especially for states where long-wavelength or wall-induced correlations in that band matter. The paper itself labels the receptive field a modeling approximation and lists fixed discretization and near-criticality as remaining sensitivities, but does not ablate the cutoff.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":15908,"tokens_out":9249,"duration_ms":103423,"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":[{"comment":"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.","section":"§II.A, Eq. (2); Methods A"},{"comment":"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.","section":"§V.C and §V.D, Fig. 2c"},{"comment":"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.","section":"§II.A, Eqs. (4)-(5); Figs. 3-4"}],"minor_comments":[{"comment":"The words 'colliods' and 'colliod' should be 'colloids' in several places, including the Fig. 3 caption and Section V.F.","section":"§III.B, Fig. 3, §V.F"},{"comment":"The thermostat name is garbled by accent encoding as 'Nos´ e–Hoover'; it should read 'Nosé–Hoover'.","section":"§V.B"},{"comment":"The abbreviation 'HF' is used in the Fig. 3c legend without being defined in the text; please define it as the Feynman–Hellmann evaluation.","section":"Fig. 3c"},{"comment":"The reference 'Fig. 4a– Fig. 4c' mixes range notation inconsistently; use 'Fig. 4a–c' for uniformity.","section":"Methods G"}],"recommendation":"major_revision","confidential_remarks":"This is a strong, well-executed study with a clear central advance and unusually extensive validation. My recommendation of major revision is driven by three unablated modeling choices (finite receptive field, low-density mask, cubic-only symmetry) that bear directly on the transferability claim. I do not see a fundamental flaw, and I would expect the requested ablations to be feasible within the manuscript's scope; if they confirm robustness, the paper would be suitable for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read this one with a genuinely favorable eye. It is the first fully three-dimensional learned excess free-energy functional for a classical fluid that I know of, trained directly from canonical MD density fields without free-energy or chemical-potential labels. The cubic equivariance is handled properly with CACE-style invariants, and because F_exc is a single scalar, both c^(1) and c^(2) come from automatic differentiation of one object, so reciprocity and variational consistency are built in rather than enforced. The benchmark suite is serious: held-out forward and inverse tests, transfer to larger cells and to grand-canonical chemical-potential differences, structure factors, compressibility-route equation of state, liquid-vapor coexistence, interfacial profiles, colloid-bridging forces, and gyroid adsorption. The key point is that none of those are fitted targets. That is a strong package.\n\nThe soft spot is the one the stress-test note identifies: the receptive field radius is three grid spacings, i.e. 1.5 sigma, while the LJTS pair potential is cut at 2.5 sigma. The learned c^(2) is therefore exactly zero beyond 1.5 sigma, and the effect of the 1.5-2.5 sigma band must be folded into shorter-range density features. The paper calls the finite range a modeling approximation and the transfer benchmarks do pass, which is real evidence in its favor. But there is no ablation showing what happens with a larger stencil, and a reviewer should ask for one. This is not fatal, but it is a missing control. The other limitations, such as the mean-field continuation of the critical point, fixed grid resolution, and the per-field gauge constants, are stated honestly and are proportionately minor.\n\nOverall, the central claim is provisionally supported: equilibrium density data can indeed be converted into a transferable thermodynamic generator, at least for the states and geometries tested. The paper is clearly written, the code and data are public, and the reasoning is transparent. This should go to peer review rather than be desk-rejected, with a request for a cutoff ablation and a sharper statement of the regime where the local approximation is expected to hold. I would cite it and bring it to a reading group.","headline":"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.","tokens_in":16415,"tokens_out":2835,"would_cite":true,"duration_ms":32547,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Equilibrium density fields alone can train a transferable three-dimensional liquid free-energy functional.","keywords":["classical density functional theory","machine-learned functional","equivariant neural network","free-energy functional","chemical-potential balance","Lennard-Jones fluid","solvation force","adsorption isotherm"],"falsifier":"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.","tokens_in":15407,"feed_emoji":"💧","tokens_out":5172,"duration_ms":53108,"temperature":0.7,"pith_summary":"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.","feed_headline":"One learned functional reproduces 3D liquid structure, phases, and forces","feed_subtitle":"Trained only on equilibrium densities, it recovers S(k), the equation of state, coexistence, and solvation forces.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Establishes the classical density functional variational principle that the learned functional is designed to satisfy.","marker":"[3]"},{"why":"Introduced neural functional theory and the idea of learning the one-body direct correlation functional.","marker":"[18]"},{"why":"Extended neural density functionals to temperature dependence and liquid-gas phase coexistence benchmarks.","marker":"[19]"},{"why":"Showed how canonical molecular dynamics data can be used by inferring chemical potentials as latent variables.","marker":"[22]"},{"why":"Provides the reference equation of state for the truncated-and-shifted Lennard-Jones fluid used to validate the learned functional.","marker":"[25]"},{"why":"Supplies direct-simulation liquid-vapor coexistence and critical-point data for comparison.","marker":"[26]"},{"why":"The Cartesian atomic cluster expansion that is adapted to the density grid to build cubic-symmetry-adapted local features.","marker":"[27]"}],"fun_headline_variants":["No-label learning yields transferable 3D density functional","Density-only training gives transferable liquid thermodynamic generator","Symmetry-aware functional learns liquid phases from 3D densities","One learned functional predicts 3D liquid structure, phases, and forces","Equivariant DFT trained on densities transfers across conditions and geometries"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["No-label learning yields transferable 3D density functional","Density-only training gives transferable liquid thermodynamic generator","Symmetry-aware functional learns liquid phases from 3D densities","One learned functional predicts 3D liquid structure, phases, and forces","Equivariant DFT trained on densities transfers across conditions and geometries"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000746,"raw_usage":{"total_tokens":3287,"prompt_tokens":870,"completion_tokens":2417,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":2332}},"tokens_in":486,"tokens_out":2417,"duration_ms":17656,"temperature":1.0,"reasoning_tokens":2332,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:27:22.089673+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"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","cited_arxiv_id":null,"evidence_quote":"Establishes the classical density functional variational principle that the learned functional is designed to satisfy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduced neural functional theory and the idea of learning the one-body direct correlation functional."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Extended neural density functionals to temperature dependence and liquid-gas phase coexistence benchmarks."},{"cited_title":"Samm¨ uller, M","cited_arxiv_id":null,"evidence_quote":"Showed how canonical molecular dynamics data can be used by inferring chemical potentials as latent variables."},{"cited_title":"Samm¨ uller and M","cited_arxiv_id":null,"evidence_quote":"Provides the reference equation of state for the truncated-and-shifted Lennard-Jones fluid used to validate the learned functional."},{"cited_title":"Glitsch, J","cited_arxiv_id":null,"evidence_quote":"Supplies direct-simulation liquid-vapor coexistence and critical-point data for comparison."}],"review_version":1}