[Paper Review] Machine Learning approaches to classical density functional theory

A. Simon, Martin Oettel|arXiv (Cornell University)|Jun 11, 2024

History and advancements in chemistryChemistry3 citations

TL;DR

This paper explores machine learning (ML) techniques to construct and improve classical density functional theory (cDFT) free energy functionals from simulation data, bypassing traditional physics-based derivations. It demonstrates that ML models—particularly neural networks—can accurately learn functional maps (e.g., density to excess free energy or external potential) with high precision, even in 1D systems where exact solutions exist, and outlines pathways for extending these methods to complex 3D systems and nonequilibrium dynamics.

ABSTRACT

In this chapter, we discuss recent advances and new opportunities through methods of machine learning for the field of classical density functional theory, dealing with the equilibrium properties of thermal nano- and micro-particle systems having classical interactions. Machine learning methods offer the great potential to construct and/or improve the free energy functional (the central object of density functional theory) from simulation data and thus they complement traditional physics- or intuition-based approaches to the free energy construction. We also give an outlook to machine learning efforts in related fields, such as liquid state theory, electron density functional theory and power functional theory as a functionally formulated approach to classical nonequilibrium systems.

Motivation & Objective

• Address the challenge of constructing accurate free energy functionals in classical density functional theory (cDFT), which remains difficult for realistic many-body systems with complex interactions.
• Overcome the limitations of traditional, intuition- or physics-based functional construction by leveraging simulation data as 'ground truth' for machine learning.
• Develop interpretable and differentiable ML models that map particle density to key cDFT quantities such as the excess free energy functional or external potential.
• Extend ML-based cDFT to nonequilibrium systems by formulating dynamic functionals for steady-state current-driven systems.
• Explore the conceptual and methodological parallels between ML-enhanced cDFT and related fields such as electron DFT and power functional theory.

Proposed method

• Use supervised machine learning to directly parameterize the excess free energy functional $\mathcal{F}_{\text{ex}}[\rho]$ from simulation data of particle densities and corresponding free energy values.
• Train neural networks to learn the functional derivative $-\beta \delta \mathcal{F}_{\text{ex}} / \delta \rho$ via $c_1[\rho]$ representation, enabling differentiable and invertible mappings.
• Apply $c_2$ matching techniques to learn the direct correlation function $c_2$ from radial distribution functions, enabling functional construction from liquid structure data.
• Train convolutional neural networks to map local density and velocity profiles $\rho(z'), v_z(z')$ to the internal force density $\mathbf{F}_{\text{int}}(z)$, representing $\delta R^{\text{int}}_{\text{ex}} / \delta \mathbf{J}$ in nonequilibrium systems.
• Utilize Bayesian and uncertainty-aware ML frameworks to quantify and propagate errors in learned functionals, improving reliability and interpretability.
• Leverage high-throughput Brownian dynamics simulations to generate ground-truth data for training ML models in steady-state, current-carrying systems.

Experimental results

Research questions

• RQ1 Can machine learning accurately reconstruct the exact free energy functional in 1D systems where the analytical solution is known?
• RQ2 To what extent can ML models learn the functional dependence of $\mathcal{F}_{\text{ex}}[\rho]$ or its derivatives from simulation data without prior physical assumptions?
• RQ3 How well can ML models generalize across different external potentials or system parameters in equilibrium cDFT?
• RQ4 Can ML representations of nonequilibrium functionals (e.g., $\delta R^{\text{int}}_{\text{ex}} / \delta \mathbf{J}$) reproduce physical behavior independent of the external driving force?
• RQ5 What are the prospects for extending ML-based cDFT to 3D systems with complex particle shapes, internal degrees of freedom, or anisotropic interactions?

Key findings

• ML models successfully reconstruct the exact free energy functional in 1D hard rod systems with quantitative accuracy, validating the feasibility of learning cDFT functionals from data.
• Neural networks trained on $c_1[\rho]$ data accurately recover the functional derivative $-\beta \delta \mathcal{F}_{\text{ex}} / \delta \rho$, enabling differentiable and invertible representations.
• Convolutional networks trained on local density and velocity profiles accurately predict the internal force density $\mathbf{F}_{\text{int}}(z)$, demonstrating robustness across varying external forces.
• The ML-based representation of $\mathbf{F}_{\text{int}}$ as $f_{z,\text{int}}^\star[\rho(z'), v_z(z')]$ captures the functional dependence on $\rho$ and $\mathbf{J}$, confirming its validity in steady-state nonequilibrium systems.
• Uncertainty-aware ML techniques, such as those in Malpica et al. (2023), enable quantification of prediction errors in learned functionals, improving reliability for future refinement.
• The conceptual framework shows strong parallels to electron DFT and power functional theory, suggesting broad applicability of ML-enhanced functional formulations to both equilibrium and nonequilibrium classical many-body problems.