ELECTRA: A Cartesian Network for 3D Charge Density Prediction with Floating Orbitals
Jonas Elsborg Luca Thiede, [1ex] Alán Aspuru-Guzik Tejs Vegge Arghya Bhowmik · 2025
Read ELECTRA: A Cartesian Network for 3D Charge Density Prediction with Floating Orbitals by Jonas Elsborg Luca Thiede, [1ex] Alán Aspuru-Guzik Tejs Vegge Arghya Bhowmik (2025) 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-2503-08305.txt.
Introduction
Jonas Elsborg and co-authors built a model in 2025 that predicts where electric charge sits inside a molecule. Charge maps guide work on new drugs and materials, but computing them from scratch is slow. Their model leans on an old idea: letting the cloud pieces that describe charge float free in space, rather than pinning them to atoms.
Finding the best spots for those pieces once took deep expert skill, which held the idea back. Their network learns the spots and the weights from data, while keeping the map's symmetry as the molecule turns. On standard tests, the model balances speed and accuracy better than rivals. Starting physics solvers from its maps cut their work by about half on new molecules.
Catalog record
| CREATOR | Jonas Elsborg Luca Thiede, [1ex] Alán Aspuru-Guzik Tejs Vegge Arghya Bhowmik |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 50 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-08305
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
ELECTRA: A Cartesian Network for 3D Charge Density Prediction with Floating Orbitals Jonas Elsborg Luca Thiede, [1ex] Alán Aspuru-Guzik Tejs Vegge Arghya Bhowmik Abstract We present the Electronic Tensor Reconstruction Algorithm (ELECTRA) - an equivariant model for predicting electronic charge densities using floating orbitals. Floating orbitals are a long-standing concept in the quantum chemistry community that promises more compact and accurate representations by placing orbitals freely in space, as opposed to centering all orbitals at the position of atoms. Finding the ideal placement of these orbitals requires extensive domain knowledge, though, which thus far has prevented widespread adoption. We solve this in a data-driven manner by training a Cartesian tensor network to predict the orbital positions along with orbital coefficients. This is made possible through a symmetry-breaking mechanism that is used to learn position displacements with lower symmetry than the input molecule while preserving the rotation equivariance of the charge density itself. Inspired by recent successes of Gaussian Splatting in representing densities in space, we are using Gaussian orbitals and predicting their weights and covariance matrices. Our method achieves a state-of-the-art balance between computational efficiency and predictive accuracy on established benchmarks. 1 Introduction High-accuracy simulations for the design of materials and molecules at the atomic scale are most often done with density functional theory (DFT) based simulations ( Kohn & Sham 1965 ) , as DFT provides a good balance between cost and accuracy for quantum mechanical simulations of matter ( Marzari et al. 2021 ) . However, the O( n^{3} ) scaling of DFT still limits the system sizes and time scales that can be simulated. Linear scaling ML surrogates, such as neural network potentials trained with a large number of DFT simulations, can alleviate this problem by learning a direct mapping between atomic structure and corresponding energy, forces, and other properties with accuracy similar to those from DFT simulations ( Friederich et al. 2021 ) . This approach, although first envisioned three decades ago ( Blank et al. 1995 ) , has become successful and popular in recent years based on multiple seminal developments ( Behler 2021 ; Deringer et al. 2021 ; Unke et al. 2021c ) . An alternative data-efficient ML-accelerated physics simulation approach can be taken where the underlying fundamental variable of the DFT simulations, the electron density, is predicted directly from atomic structures, without self-consistent field (SCF) iterations. Following the Hohenberg-Kohn theorem ( Hohenberg & Kohn 1964 ) , all ground-state properties can be calculated once this ground-state electron density is known ( Grisafi et al. 2022 ; Bogojeski et al. 2020 ) . In recent years, researchers have addressed this task in multiple ways, differentiated by the representation of density data, the molecular representation, and the ML architecture itself ( Grisafi et al. 2018 ; Chandrasekaran et al. 2019 ; Jørgensen & Bhowmik 2022 ; Rackers et al. 2023 ) . The target electron density is commonly predicted on real space grids ( Chandrasekaran et al. 2019 ; Jørgensen & Bhowmik 2020 ; Li et al. 2024 ) or as an expansion of atom-centered basis functions ( Grisafi et al. 2018 ; Unke et al. 2021a ; Cuevas-Zuviría & Pacios 2021 ; Rackers et al. 2023 ) which usually take the form \Phi_{l,m}(\boldsymbol{r})=R_{l}(r)Y_{lm}\left(\theta,\phi\right),\quad l=0,...,L (1) R_{l}(r) represents the radial dependence relative to a center, while Y_{lm}(\theta,\phi) captures the angular dependencies. Larger quantum numbers l correspond to higher-frequency components. The accuracy of the represented density is dependent on the "quality" of the basis set or the grid density. Two key attributes define the quality of a basis set: The number of basis functions per angular quantum number l and the maximum angular momentum quantum number L included in the expansion. Different systems and properties necessitate varying levels of basis set complexity. There is no universal basis set that provides both, high accuracy and optimal computational efficiency for all types of systems. Instead, the selection of an appropriate basis set depends on the specific requirements of the system under investigation and requires deep domain expertise. For example, accurate descriptions of systems involving highly polarizable molecules or those with diffuse electron distributions far away from atom centers may require augmented basis sets like aug-cc-pVTZ ( Kendall et al. 1992 ) , which include functions with high angular momentum and diffuse components that have long-tailed radial functions designed for modeling long-range dependencies. For smaller systems or those dominated by core-electron interactions, these basis sets lead to unnecessarily large compute costs. In particular, basis functions with higher angular quantum numbers L incur significant costs. A more compact representation of densities can be achieved by putting extra basis functions at locations of presumed interest, particularly in areas far away from atoms, with rapidly varying densities. These basis functions are called "floating" orbitals, and their utility is well-established in electronic structure theory ( Tao & Pan 1992 ; Tao 1993 ; Tasi & Császár 2007 ) . They date back to the floating spherical Gaussian orbital (FSGO) model ( Frost 1968 ) . When chosen wisely, floating orbitals can lead to significant improvements in calculation speed and accuracy ( Lorincz & Nagy 2024 ) by reducing the need for diffuse and high angular momentum basis functions. Well-placed floating orbitals can represent densities more efficiently, using lower maximal angular quantum numbers L . ELECTRA is the first model to predict floating orbital positions without human input. However, the optimal locations of floating orbitals are often hard to determine Zheng et al. 2021 , and picking good locations therefore requires deep electronic structure domain expertise ( Lorincz & Nagy 2024 ) . Our core contribution is a data-driven solution to this problem. We are training a model that, given a molecular graph, accurately reconstructs ground truth charge densities by predicting the 3D position of floating orbitals as well as the coefficients and parameters that define them. Since charge densities are rotation invariant, we use a rotation equivariant neural network as the backbone of our model. However, a naive implementation of equivariant neural networks is destined to fail, since good placements of floating orbitals can have lower symmetry than the input molecular graph, as we will discuss in later sections. We address this problem by developing a symmetry-breaking mechanism that retains rotational equivariance. We call the resulting model the Electronic Tensor Reconstruction Algorithm (ELECTRA). We test ELECTRA on the widely used QM9 charge density dataset ( Jørgensen & Bhowmik 2022 ) and achieve results that are competitive with state-of-the-art while being consistently faster. 2 Related work 2.1 Charge density prediction Prior work on machine learning prediction of charge density (CD) generally falls into two main approaches, inspired by earlier non-ML methods. Orbital-based methods are rooted in linear combinations of atom-centered orbitals (LCAO), which take the form \rho(\boldsymbol{r})=\sum_{i}^{N}\sum_{j}^{N_{b}^{i}}\sum_{m=-l_{i,j}}^{l_{i,j}}c_{i,j,m}\Phi_{\alpha_{i,j},l_{i,j},m,\boldsymbol{r}_{i}}(\boldsymbol{r}), (2) where the first sum runs over all atoms, and the other two sum index into all basis functions per atom. \Phi usually takes a form as in 1 . In the ML community, methods based on this construction typically predict coefficients c_{i,j,m} extracted from ground truth DFT calculations ( Fabrizio et al. 2019 ; Qiao et al. 2022 ; Rackers et al. 2023 ; Cheng & Peng 2024 ; del Rio et al. 2023 ;