High-Dimensional Operator Learning for Molecular Density Functional Theory

Jinni Yang,Runtong Pan,Jikai Sun,Jianzhong Wu

2024-11-06

Abstract:Classical density functional theory (cDFT) provides a systematic approach to predict the structure and thermodynamic properties of chemical systems through the single-molecule density profiles. Whereas the statistical-mechanical framework is theoretically rigorous, its practical applications are often constrained by challenges in formulating a reliable free-energy functional and the complexity of solving multidimensional integro-differential equations. In this work, we established an optimized operator learning method that effectively separates the high-dimensional molecular density profile into two lower-dimensional components, thereby exponentially reducing the vast input space. The convoluted operator learning network demonstrates exceptional learning capabilities, accurately mapping the relationship between the density profile of a carbon dioxide system to its one-body direct correlation function using an atomistic polarizable model. The neural operator model can be generalized to more complex systems, offering high-precision cDFT calculations at low computational cost.

Chemical Physics

What problem does this paper attempt to address?

This paper attempts to address two major issues in the practical application of classical density functional theory (cDFT):

1. **Construction of a reliable free energy functional**: Although the theoretical framework of cDFT is rigorous, constructing a reliable free energy functional in practical applications is very challenging. Existing methods often rely on approximations, which may not be accurate enough in complex systems.

2. **Solving multidimensional integral-differential equations**: cDFT requires solving high-dimensional integral-differential equations, which is computationally very complex and time-consuming, especially when dealing with multidimensional molecular density profiles.

To tackle these problems, the authors propose an optimized operator learning method (Convoluted Operator Learning Network, COLN), which can decompose high-dimensional molecular density profiles into two low-dimensional components, thereby significantly reducing the dimensionality of the input space. Using this method, the authors successfully mapped the relationship between the density profile of the carbon dioxide system and its first-order direct correlation function, demonstrating high accuracy and low computational cost in complex systems.