A unified machine-learning framework for ab initio multiscale modeling of liquids

Anna T. Bui; Stephen J. Cox

REVIEW 2 major objections 4 minor 3 cited by

Machine-learned interatomic potentials can train neural classical density functionals, giving a single first-principles route from the Schrödinger equation to liquid structure and thermodynamics across scales.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-13 21:28 UTC pith:K4JJT6A2

load-bearing objection Solid first demonstration that MLIP density profiles can train neural cDFT for real polyatomic fluids, delivering usable ab initio binodals and confined phase diagrams with quantitative MD validation. the 2 major comments →

arxiv 2603.20493 v2 pith:K4JJT6A2 submitted 2026-03-20 physics.chem-ph cond-mat.mtrl-scicond-mat.othercond-mat.stat-mechphysics.comp-ph

A unified machine-learning framework for ab initio multiscale modeling of liquids

Anna T. Bui , Stephen J. Cox This is my paper

classification physics.chem-ph cond-mat.mtrl-scicond-mat.othercond-mat.stat-mechphysics.comp-ph

keywords machine-learned interatomic potentialsneural classical density functional theoryab initio multiscale modelingliquid-vapor coexistenceconfined waterFisher-Widom lineWidom linesupercritical carbon dioxide

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper shows that quantum-derived machine-learned interatomic potentials can be used to generate the inhomogeneous density profiles needed to train neural classical density functional theory. The resulting ab initio neural cDFT yields both equilibrium density profiles and free energies for homogeneous and inhomogeneous fluids far more cheaply than direct molecular simulation, while remaining faithful to the underlying electronic-structure Hamiltonian. Applied to water and carbon dioxide with several exchange-correlation functionals, the method recovers bulk equations of state and liquid–vapor binodals, predicts how graphene confinement shifts water’s coexistence curve, and locates the Fisher–Widom and Widom lines in supercritical carbon dioxide. A sympathetic reader cares because the same variational object—the excess free-energy functional—now links microscopic interactions to mesoscopic and macroscopic thermodynamics without intermediate empirical models or ensemble-specific free-energy calculations.

Core claim

Training a neural representation of the one-body direct correlation functional on planar density profiles generated by MLIPs produces an ab initio neural cDFT that accurately reproduces bulk thermodynamics, liquid–vapor phase diagrams, confinement-modified coexistence of water, and the Fisher–Widom and Widom lines of supercritical CO2, thereby establishing a unified first-principles multiscale framework for fluids.

What carries the argument

Ab initio neural cDFT: the one-body direct correlation functional c^(1)[ρ;T] is learned by treating the Euler–Lagrange equation itself as the loss on canonical MLIP density profiles (with chemical potential as a latent variable); functional line integration then recovers the excess free energy, so structure and thermodynamics of any external potential follow from a single self-consistent minimization.

Load-bearing premise

Both the machine-learned potentials and the neural functional assume that only nearby density values matter, so long-range electrostatic correlations can be omitted without changing the reported asymptotic decay and Fisher–Widom line.

What would settle it

Recompute the Fisher–Widom line for the same PBE-D3 carbon dioxide with an MLIP that explicitly includes long-range electrostatics (or with hyper-DFT mean-field corrections) and check whether the line still intersects the binodal near 0.9 Tc and 1.93 ρc.

Watch this falsifier — get emailed when new claim-graph text bears on it.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

Solid first demonstration that MLIP density profiles can train neural cDFT for real polyatomic fluids, delivering usable ab initio binodals and confined phase diagrams with quantitative MD validation.

read the letter

This paper closes a practical loop that people have been talking about for a couple of years: take MLIPs trained on quantum data, generate inhomogeneous density profiles, and train a neural representation of c^(1) that then gives you bulk EOS, liquid-vapor binodals, confined coexistence, and even Fisher-Widom poles by solving the Euler-Lagrange equation. That is new for water and CO2 with multiple xc functionals; earlier neural cDFT work stayed with LJ, hard spheres, or TIP4P/2005.

What they do well is the validation. Figures 2–4 and the SI show direct, quantitative agreement with MD for pressures, S(k), coexistence densities, confined profiles, and effective pressures across SCAN, RPBE-D3, PBE-D3, BLYP-D3, SCAN-rVV10, and the usual empirical models. Thermodynamic consistency between the structural and free-energy routes for disjoining pressure is checked. Critical points match prior simulation literature. The confined-water phase diagram and the scCO2 crossover lines are obtained by functional minimization, not by fitting, and the computational gain (minutes vs. free-energy campaigns) is real.

The softest spot is the one they already flag: both the MLIPs and the neural functional are local, so the asymptotic decay that defines the Fisher-Widom line could be an artifact. Everything else in the pipeline (EOS, binodals, confinement) does not rest on that particular feature. Geometry is still planar, data/code are promised but not yet public, and long-range electrostatics are left for hyper-DFT extensions. None of these undercut the central claim.

This is for people who care about first-principles nanofluidics, supercritical fluids, or mesoscale thermodynamics without heroic free-energy sampling. The math is standard cDFT plus a clean residual loss; the citation pattern is honest. I would send it to referees and I would cite the confined-water and scCO2 results once the data land.

Referee Report

2 major / 4 minor

Summary. The manuscript introduces ab initio neural cDFT by training neural representations of the one-body direct correlation functional c^(1) on planar inhomogeneous density profiles generated from MLIPs (DeepMD, HD-NNP, MACE) that themselves approximate DFT energies/forces for water (SCAN, RPBE-D3) and CO2 (PBE-D3, BLYP-D3, SCAN-rVV10), plus empirical references. The learned functional is then used via the Euler–Lagrange equation and functional line integration to obtain bulk equations of state, structure factors S(k), liquid–vapor binodals, confined density profiles and effective pressures, confinement-modified coexistence for SCAN water between graphene sheets, and the Fisher–Widom and Widom lines of supercritical PBE-D3 CO2. Direct comparisons to MD/GCMC are shown throughout Figs. 2–4 and S3–S5.

Significance. If the results hold, the work supplies a practical, first-principles multiscale route that treats micro-, meso- and macroscale fluid thermodynamics within a single grand-canonical framework, bypassing the cost and ensemble limitations of direct MLIP simulation for open or large-scale inhomogeneous systems. Strengths include quantitative MD validation across multiple xc functionals, explicit thermodynamic-consistency checks between structural and free-energy routes for the disjoining pressure (Fig. S5), and the ability to map entire confined phase diagrams and asymptotic crossover lines in minutes. The locality caveat for the Fisher–Widom line is already flagged by the authors; the remainder of the pipeline does not rest on that single asymptotic feature.

major comments (2)

Discussion (and abstract claim that the method “captures o the Fisher–Widom and Widom lines”): the authors correctly note that the reported FW line (Fig. 4) may be an artifact of the strictly local MLIP description. Because this feature is presented as a key demonstration of the framework’s reach, the abstract and concluding claims should be tempered to “Widom lines and a candidate Fisher–Widom line under the locality assumption,” or a short additional calculation with an existing long-range MLIP (or a mean-field electrostatic correction via hyper-DFT) should be added to quantify the sensitivity.
Methods / training protocol: the chemical potentials of the canonical training trajectories are treated as latent variables optimized inside the Euler–Lagrange residual loss. While the subsequent bulk EOS and coexistence densities match independent MD, it remains unclear how unique or transferable the learned {μ_k} set is when the same network is applied to qualitatively different external potentials (e.g., the graphene walls of Fig. 3). A brief hold-out test that freezes the network and re-optimizes only μ for a new V_ext family would strengthen the claim that the functional itself, rather than the latent variables, carries the physics.

minor comments (4)

Fig. 2B (right) and Fig. 4B: the non-monotonic S(k→0) and the two Widom-line definitions are clear, but the color scale for κ_T^{-1} is unlabeled in the heat map; a color bar would help.
Eq. (3) and Methods: the intramolecular partition function “ is set to unity without comment; a one-sentence justification (or reference to the hyper-DFT treatment of molecular degrees of freedom) would avoid confusion for readers unfamiliar with the earlier papers.
Supporting Information S7: the hyper-DFT hydrogen profiles are a nice illustration, yet they appear only in the SI; a brief mention in the main-text Discussion would better advertise the route to multi-site observables.
Typographical: “ab initioneural” appears repeatedly without a space or hyphen; consistent “ab initio neural cDFT” would improve readability.

Circularity Check

0 steps flagged

No significant circularity: neural cDFT is trained on independent MLIP MD density profiles via the Euler-Lagrange residual; subsequent binodals, confined coexistence, and supercritical lines are obtained by solving the learned functional and are cross-checked against held-out MD.

full rationale

The derivation chain is: QM energies/forces train MLIPs; MLIPs generate planar inhomogeneous density profiles under random V_ext (and some coexistence) via canonical MD; those profiles supply the training targets for a neural representation of c^(1) by treating the Euler-Lagrange equation itself as the loss (with chemical potentials as latent variables). Once trained, the functional is evaluated by self-consistent solution of the EL equation (or by functional line integration for free energies, automatic differentiation + OZ for S(k) and poles). Bulk EOS, liquid-vapor binodals, confined density profiles and effective pressures, and the Fisher-Widom/Widom lines are all obtained this way and are quantitatively compared to independent MD (or GCMC) trajectories that were not used as the loss targets for those particular state points. Inclusion of a modest number of coexistence profiles in the training set does not force the entire binodal by construction; the network still interpolates across temperatures and recovers van der Waals loops and critical points that match separate direct-coexistence runs. Self-citations to the authors' earlier neural-cDFT and hyper-DFT papers supply the learning architecture and the grand-potential formalism, but those prior results are not uniqueness theorems that forbid alternatives, nor do they encode the numerical values of the present phase diagrams or crossover lines. The only soft point (locality of both MLIP and neural cDFT) is already flagged by the authors themselves for the Fisher-Widom line and does not collapse any reported prediction into a fitted parameter. Hence the central multiscale claim remains non-circular.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The framework rests on the exact variational principle of classical DFT, the uniqueness of the excess free-energy functional for a given intermolecular potential, the locality ansatz of both MLIPs and neural cDFT, and standard numerical choices (network size, window width, random external-potential ensemble). No new physical entities are postulated; free parameters are the usual ML hyperparameters and the number of training trajectories.

free parameters (4)

neural-network architecture (3 layers: 128-64-32, softplus)
Chosen by the authors; controls capacity of the learned c^(1) functional.
local density window size (1 nm)
Hyperparameter of the local learning strategy; determines the range of density information fed to the network.
number and form of random external potentials (500–1000 trajectories)
Determines the training distribution; amplitudes and phases drawn from fixed Gaussians.
MLIP training loss weights p_E, p_f and architecture (DeepMD/HD-NNP/MACE)
Inherited from prior MLIP papers but re-used here; affect the quality of the density profiles that train neural cDFT.

axioms (4)

domain assumption Classical DFT variational principle: equilibrium density minimizes the grand-potential functional (Eq. 1)
Standard statistical mechanics; invoked throughout to justify solving the Euler-Lagrange equation.
domain assumption For a given intermolecular potential and temperature the excess free-energy functional F_ex is unique
Fundamental theorem of cDFT; allows a single neural representation per fluid/xc functional.
ad hoc to paper c^(1)(r;[ρ];T) depends only on the density in a local neighborhood of r
Locality ansatz of neural cDFT (and of the underlying MLIPs); stated as an assumption that enables the sliding-window architecture.
ad hoc to paper Canonical MD density profiles under random V_ext plus latent chemical potentials are sufficient to learn the grand-canonical functional
Recent technical development (Ref. 51) adopted here because GCMC is impractical with MLIPs.

pith-pipeline@v1.1.0-grok45 · 28789 in / 2819 out tokens · 33098 ms · 2026-07-13T21:28:24.561666+00:00 · methodology

0 comments

read the original abstract

Understanding and predicting the behavior of liquid matter across length scales, using only the microscopic interactions encoded in the Schr\"odinger equation, remains a central challenge in the physical sciences. Achieving this goal requires not only an accurate and efficient description of intermolecular forces but also a consistent framework that bridges the micro-, meso-, and macroscales. Here, by combining machine-learned interatomic potentials (MLIPs) with neural classical density functional theory (neural cDFT), we present such a framework. The underlying idea is simple: MLIPs trained on quantum-mechanical energies and forces are used to generate inhomogeneous microscopic density profiles, which in turn serve as the training data for neural cDFT. The resulting ab initio neural cDFT is not only significantly more computationally efficient than molecular simulations, but also provides a conceptually transparent route to the thermodynamics of both homogeneous and inhomogeneous systems. We demonstrate the approach for both water and carbon dioxide using several exchange-correlation functionals. Beyond accurately reproducing bulk equations of state and liquid-vapor phase diagrams, ab initio neural cDFT predicts, from first principles, how confinement modifies liquid-vapor coexistence in water. It also captures complex behavior in supercritical carbon dioxide such as the Fisher-Widom and Widom lines. Ab initio neural cDFT establishes a general first-principles route to multiscale modeling of fluids within a single unified conceptual framework.

discussion (0)

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

Entropy power functional theory for Brownian many-body dynamics
cond-mat.soft 2026-07 reject novelty 7.0

The paper proposes an entropy-based power functional theory for Brownian dynamics, but its ideal dissipation functional omits the J²/2ρ term, leaving the one-body current undetermined in uniform systems.
Entropy density functional universality: Correlation, response, and entropic Ornstein-Zernike structure
cond-mat.soft 2026-07 accept novelty 6.0

A unique universal excess entropy functional of density and global pair-distance distribution determines classical equilibrium many-body physics via joint grand-potential minimization and coupled entropic Ornstein-Zer...
Entropy density functional theory for inhomogeneous fluids
cond-mat.soft 2026-06 unverdicted novelty 6.0

An exact variational density functional theory for inhomogeneous fluids using metadensity minimization and a universal excess entropy functional that bypasses the two-body density.