V2Rho-FNO: Fourier Neural Operator for Electronic Density Prediction
Yingdi Jin #, Xinming Qin #, Ruichen Liu, Jie Liu, Zhenyu Li, Jinlong Yang · 2026
Read V2Rho-FNO: Fourier Neural Operator for Electronic Density Prediction by Yingdi Jin #, Xinming Qin #, Ruichen Liu, Jie Liu, Zhenyu Li, Jinlong Yang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-15669.txt.
Introduction
Jin and co-authors predict electron density with a Fourier Neural Operator. Density functional theory is exact enough for wide use but slow at scale. Their model learns a direct map from outside potentials to electron density. It skips atomic orbitals, basis sets, and handcrafted descriptors. Frequency-domain operator learning captures global interactions and long-range links. Training spanned many elements and molecular shapes. The model then hit unseen molecules with no retraining. The authors frame it as a fast route for screening work.
Catalog record
| CREATOR | Yingdi Jin #, Xinming Qin #, Ruichen Liu, Jie Liu, Zhenyu Li, Jinlong Yang |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 45306 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.15669
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
V2Rho-FNO: Fourier Neural Operator for Electronic Density Prediction Yingdi Jin #, Xinming Qin #, Ruichen Liu, Jie Liu, Zhenyu Li, Jinlong Yang Abstract Density functional theory (DFT) is a cornerstone of modern computational chemistry and materials science; nevertheless, its computational expense limits its deployment in large-scale and high-throughput applications that rely on explicit electronic-structure calculations. While machine learning methods have accelerated energy calculations for specific molecular categories 4 ; 20 ; 23 , achieving transferable electronic density predictions across diverse chemical spaces remains a significant challenge 3 ; 9 ; 6 . This study introduces a universal framework based on Fourier Neural Operators (FNOs), which directly learns mappings from external potentials to electronic density distributions 16 ; 12 . After training on datasets containing various elements and molecular configurations, our model achieves zero-shot generalization to completely unknown molecular systems, accurately predicting their electronic densities. This transferability stems from FNOs’ inherent capability to capture global electronic interactions without explicit atomic orbitals or basis sets 16 ; 12 . The proposed method opens new avenues for rapid and accurate electronic structure calculations, providing a powerful tool for exploring chemical space. Density functional theory (DFT) is a cornerstone of modern computational chemistry and materials science; nevertheless, its computational expense limits its deployment in large-scale and high-throughput applications that rely on explicit electronic-structure calculations. While machine learning methods have accelerated energy calculations for specific molecular categories 4 ; 20 ; 23 , achieving transferable electronic density predictions across diverse chemical spaces remains a significant challenge 3 ; 9 ; 6 . This study introduces a universal framework based on Fourier Neural Operators (FNOs), which directly learns mappings from external potentials to electronic density distributions 16 ; 12 . After training on datasets containing various elements and molecular configurations, our model achieves zero-shot generalization to completely unknown molecular systems, accurately predicting their electronic densities. This transferability stems from FNOs’ inherent capability to capture global electronic interactions without explicit atomic orbitals or basis sets 16 ; 12 . The proposed method opens new avenues for rapid and accurate electronic structure calculations, providing a powerful tool for exploring chemical space. † † email: [email protected] † † affiliation: State Key Laboratory of Precision and Intelligent Chemistry, University of Science and Technology of China, Hefei, Anhui 230026, China † † affiliation: Hefei National Research Center for Physical Sciences at the Microscale, University of Science and Technology of China, Hefei, Anhui 230026, China † † affiliation: Hefei National Research Center for Physical Sciences at the Microscale, University of Science and Technology of China, Hefei, Anhui 230026, China † † email: [email protected] † † affiliation: Hefei National Laboratory, University of Science and Technology of China, Hefei 230088, China † † email: [email protected] † † affiliation: Hefei National Laboratory, University of Science and Technology of China, Hefei 230088, China † † affiliation: State Key Laboratory of Precision and Intelligent Chemistry, University of Science and Technology of China, Hefei, Anhui 230026, China † † email: [email protected] † † affiliation: Hefei National Laboratory, University of Science and Technology of China, Hefei 230088, China † † affiliation: State Key Laboratory of Precision and Intelligent Chemistry, University of Science and Technology of China, Hefei, Anhui 230026, China 1 Introduction Density functional theory (DFT) has become the cornerstone of modern electronic structure calculations due to its favorable balance between accuracy and computational efficiency 8 ; 11 . Central to DFT is the Hohenberg–Kohn (HK) theorem, which establishes that the ground-state electron density uniquely determines all properties of an interacting many-electron system and, in particular, is uniquely determined by the external potential up to an additive constant 8 . In practical Kohn–Sham formulations, the electron density is obtained through a self-consistent solution of effective single-particle equations 11 . Despite its success, the associated self-consistent field (SCF) procedure remains computationally demanding for large systems, high-throughput screening, and time-dependent simulations 18 ; 24 . Recent years have witnessed rapid progress in machine learning approaches aimed at accelerating or partially bypassing electronic structure calculations 4 ; 20 . A particularly active direction focuses on learning electron densities directly, either to replace SCF iterations or to enable fast downstream property evaluation 21 ; 3 ; 9 ; 6 . These approaches span kernel methods, deep neural networks, and equivariant graph neural networks, achieving impressive accuracy across molecules, liquids, and solids 23 ; 9 . In most existing formulations, however, density prediction is posed as a mapping from discrete atomic descriptors—such as element types and nuclear coordinates—to continuous density fields. As a consequence, substantial architectural effort is required to encode physical symmetries, long-range interactions, and resolution dependence through handcrafted embeddings, message-passing schemes, or equivariant constraints. A conceptually more fundamental alternative is to learn the HK map itself, namely the functional mapping from the external potential to the ground-state electron density. Several studies have explored this idea in reduced settings, including one-dimensional model systems and molecular datasets, demonstrating that the HK map can be approximated by machine learning models with encouraging accuracy 3 ; 1 . Nevertheless, extending this approach to fully three-dimensional systems remains challenging, particularly when accurate densities must be resolved on fine spatial grids. Moreover, most existing models are intrinsically tied to a fixed spatial discretization, limiting their ability to generalize across resolutions and system representations. In parallel, the development of neural operators has introduced a new paradigm for learning mappings between function spaces 12 . Unlike conventional neural networks, which learn finite-dimensional input–output relations, neural operators are designed to approximate operators acting between infinite-dimensional function spaces. A key consequence of this formulation is that, once trained, a neural operator can be evaluated on inputs represented on discretizations different from those used during training, provided that the underlying functions lie within the operator’s domain of approximation 12 ; 5 . Among these methods, the Fourier neural operator (FNO) has proven particularly effective for learning nonlocal operators by combining global spectral convolutions with local nonlinear transformations 16 . While neural operators have been widely applied to explicit partial differential equations, their potential for learning implicit physical maps in quantum many-body problems has only recently begun to be explored 10 . In this work, we propose V2Rho-FNO , a three-dimensional Fourier neural operator that directly learns the external-potential-to-density mapping for electronic systems. Unlike structure-based density models that rely on discrete atomic descriptors, our approach takes the nuclear electrostatic potential as the sole input and predicts the ground-state electron density on uniform spatial grids. This formulation is directly motivated by the Hohenberg–Kohn theorem 8 and treats both input and output as continuous physical fields, placing the problem naturally within the neural operator framework 12 ; 16 . From the perspective of operator learning, V2Rho-FNO aims to approximate a universal functional relationship between external potentials and ground-state densities, rather than a mapping tied to specific molecular structures. In neural operator theory, generalization is governed not by similarity between individual training and test samples, but by coverage of the input function space on which the operator is defined 12 ; 14 . Consequently, once sufficient and diverse training data are provided to span the relevant space of external potentials, a trained operator is expected to generalize to previously unseen molecular systems without requiring structural similarity to