Global Plane Waves From Local Gaussians: Periodic Charge Densities in a Blink
Jonas Elsborg, Felix Ærtebjerg, Luca Thiede, Alán Aspuru-Guzik, Tejs Vegge, Arghya Bhowmik · 2026
Read Global Plane Waves From Local Gaussians: Periodic Charge Densities in a Blink by Jonas Elsborg, Felix Ærtebjerg, Luca Thiede, Alán Aspuru-Guzik, Tejs Vegge, Arghya Bhowmik (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-2601-19966.txt.
Introduction
Jonas Elsborg and five colleagues predict the electron charge density of crystals with a fast neural model. Charge density is a fundamental property that determines how a material behaves. Computing it with traditional quantum methods is slow. Their model, ELECTRAFI, builds the density from anisotropic Gaussians placed in real space.
The trick is mathematical: Gaussians have closed-form Fourier transforms. Using the Poisson summation formula, the model evaluates plane-wave coefficients analytically. That makes the whole model differentiable end to end and fast. Accurate, quick charge densities would speed materials screening for batteries and catalysts. The work shows how classical harmonic analysis can sharpen modern machine learning. The trick is neat and fast. It turns a hard sum into a short one. That speed is the whole point of the work.
Catalog record
| CREATOR | Jonas Elsborg, Felix Ærtebjerg, Luca Thiede, Alán Aspuru-Guzik, Tejs Vegge, Arghya Bhowmik |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 65 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2601-19966
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Global Plane Waves From Local Gaussians: Periodic Charge Densities in a Blink Jonas Elsborg, Felix Ærtebjerg, Luca Thiede, Alán Aspuru-Guzik, Tejs Vegge, Arghya Bhowmik Abstract We introduce ELECTRAFI, a fast, end-to-end differentiable model for predicting periodic charge densities in crystalline materials. ELECTRAFI constructs anisotropic Gaussians in real space and exploits their closed-form Fourier transforms to analytically evaluate plane-wave coefficients via the Poisson summation formula. This formulation delegates non-local and periodic behavior to analytic transforms, enabling reconstruction of the full periodic charge density with a single inverse FFT. By avoiding explicit real-space grid probing, periodic image summation, and spherical harmonic expansions, ELECTRAFI matches or exceeds state-of-the-art accuracy across periodic benchmarks while being up to 633 × 633\times faster than the strongest competing method, reconstructing crystal charge densities in a fraction of a second. When used to initialize DFT calculations, ELECTRAFI reduces total DFT compute cost by up to ∼ \sim 20 %, whereas slower charge density models negate savings due to high inference times. Our results show that accuracy and inference cost jointly determine end-to-end DFT speedups, and motivate our focus on efficiency. † † affiliationnotice: Equal contribution 1 Introduction Kohn–Sham density-functional theory (DFT) is the most widely used tool for electronic structure calculations across physics, chemistry, and materials science ( 26 ; 38 ) . By reformulating the many-electron problem in terms of the electron density, DFT achieves a favorable balance between accuracy and computational cost and underpins a mature ecosystem of widely used simulation codes ( 41 ; 22 ; 28 ) . Even so, the cubic scaling of canonical DFT remains prohibitive for large systems, long molecular-dynamics runs, and high-throughput screening. As a result, machine-learning methods are increasingly developed to accelerate DFT calculations, following two complementary strategies. The most common approach is to train ML models, typically interatomic potentials, to directly predict DFT observables such as energies, forces, or band gaps, and thus bypass the costly self-consistent field (SCF) cycle, resulting in large speedups, but losing any guarantee of reproducing a chosen DFT functional ( 11 ; 2 ) . An alternative approach is to learn the electron density (or a high-quality initial guess) with ML, which can be used to initialize the SCF procedure and reduce the number of iterations required for convergence to the true self-consistent DFT result ( 4 ) . Direct ML guesses for the charge density are especially natural for plane-wave basis DFT implementations, since these accept a real-space charge density on a grid as the SCF starting point ( 41 ; 22 ) . These methods are predominantly used for periodic materials. Recent density prediction models demonstrate strong accuracy for crystalline systems ( 31 ; 39 ) , but they rely on computationally expensive dense real-space grid evaluation, which can negate the benefits of SCF acceleration when accounting for model inference time. A gap, therefore, remains for a model that naturally captures the long-range dependencies of periodic structures while still providing fast and accurate inference. 2 Related work & motivation In contrast to MLIPs, density-based approaches or slight variations of their architectures can predict spatially resolved quantities, which are of great practical interest to understand the chemistry of a system and make targeted interventions. For example, Fukui functions ( 53 ) f(\mathbf{r})=\left(\frac{\partial\rho(\mathbf{r})}{\partial N}\right)_{v} are fundamental to predicting the most likely regions of a molecular system to be involved in a reaction. Other examples include the Quantum Theory of Atoms in Molecules (QTAIM), which allows the location and classification of bonds by calculating bond critical points of \rho(\mathbf{r}) , again important to elucidate reaction mechanisms. The density can also be used directly in density-based generative models for drug design ( 62 ; 73 ) , or as descriptors of electrochemical systems ( 44 ; 9 ; 65 ) and electrocatalysts ( 76 ; 37 ) . Densities can also be used to build cheap surrogates for observables like ionic diffusion ( 32 ) . However, the most obvious benefit is the potential for a significant reduction in wall time requirements associated with routine DFT calculations, since DFT is one of the largest global consumers of supercomputing time in chemistry and materials science ( 21 ; 25 ) . This demand is unlikely to diminish, particularly for complex interfaces and nanoscale systems ( 27 ) . Both task-specific machine learning potentials and emerging foundation models for materials discovery are data-hungry and rely on DFT (or beyond-DFT) calculations for pretraining and continual improvement ( 42 ; 58 ) . Using ML-predicted charge densities as high-quality initial guesses for self-consistent DFT calculations offers a direct mechanism to reduce the number of electronic iterations required for convergence, thereby lowering the aggregate CPU/GPU hours consumed by large-scale electronic structure campaigns. Machine learning models for charge density prediction mirror two paradigms from electronic structure theory, which in the literature have been referred to as orbital models and probe models . Orbital models. These approaches expand the density in atom-centered basis functions, \rho(\boldsymbol{r})=\sum_{i=1}^{N}\sum_{j=1}^{N_{b}^{i}}\sum_{m=-l_{i,j}}^{l_{i,j}}c_{i,j,m}\,\Phi_{\alpha_{i,j},\,l_{i,j},\,m,\mathbf{r_{i}}}(\boldsymbol{r}), (1) where i indexes atoms, j enumerates basis functions on atom i , and m runs over magnetic sublevels. The coefficients \{c_{i,j,m}\} weigh the basis functions \Phi_{l,m}(\boldsymbol{r})=R_{l}(r)Y_{lm}\left(\theta,\phi\right) , which are products of radial basis functions R_{l}(r) and a spherical harmonic Y_{lm}(\theta,\phi) . Learning to predict the density can thus be framed as the prediction of the coefficients \{c_{i,j,m}\} and, in some cases, refining parts of the radial function ( 13 ; 59 ; 61 ; 8 ; 10 ; 14 ; 17 ) . However, a fixed atom-centered basis generally limits expressivity, often forcing large expansions to capture inter-atomic features. Bond-centered augmentations increase accuracy ( 17 ) , but learned, per-atom displacements provide a more general solution ( 12 ) : simple 3D Gaussians placed at predicted offsets deliver high accuracy while eliminating high-order spherical harmonics. This leads to much faster inference and avoids introducing a strong model dependence on the choice of basis sets and their associated coefficient representation. Current orbital-based models perform well on small molecules and organic chemistries, but in solids, the non-local features of the charge density require large orbital expansions with many diffuse and high-angular-momentum functions, which undermines the efficiency of the LCAO ansatz for materials ( 45 ) . As a result, high accuracy has not been demonstrated in general for inorganic crystals and metals ( 8 ; 34 ; 17 ) . Probe models. Another class of models treat points of the numerical grid ( 6 ) as nodes of a big graph neural network and allows each point to receive messages from an atomic graph encoder ( 23 ; 31 ; 39 ; 57 ; 46 ) . This yields high functional flexibility and results in significant reductions in the number of SCF steps when utilized as initial guesses in regular DFT calculations ( 39 ) . However, sending messages to all grid points requires many GNN queries and, therefore, higher inference time compared to orbital approaches ( 17 ) . Common to most modern density prediction models is that they build on graph neural network models and/or interatomic potentials, which we describe briefly below. Molecular modeling with graph neural networks Graph Neural Networks (GNNs) provide a powerful framework for processing data that can be structured as graphs. Formally,