Faculty of Physics - University of Bucharest

UNIVERSITY OF BUCHAREST
FACULTY OF PHYSICS

Guest
2026-06-11 23:58

HOME CONFERENCES SEARCH LOGIN NEW USER IMAGES

Conference: Bucharest University Faculty of Physics 2026 Meeting

Section: Theoretical and Computational Physics, Applied Mathematics

Title:

Heuristic Charge Densities for Faster Materials Simulations: AI-Powered Self-Consistent DFT with mcv2v-cGAN

Authors:

Calin-Andrei PANTIS-SIMUT (1,2), Nicolae FILIPOIU (2,3), Andrei MANOLESCU (4), George Alexandru NEMNES (1,2,5)

^*
Affiliation:

1) Faculty of Physics, Doctoral School of Physics, University of Bucharest, Atomistilor 405, Magurele, 077125, Ilfov, Romania.

2) Horia Hulubei National Institute for Physics and Nuclear Engineering, Atomistilor 405, Magurele, 077125, Ilfov, Romania.

3) National Institute of Materials Physics, Atomistilor 405, Magurele, 077125, Ilfov, Romania.

4) Department of Engineering, Reykjavik University, Menntavegur 1, Reykjavik, IS-102, Iceland.

5) Research Institute of the University of Bucharest (ICUB), 90 Panduri Street, Bucharest, 050663, Romania.

E-mail

calin.pantis@nipne.ro

Keywords:

density functional theory, SC charge density prediction, conditional generative-adversarial network, reduction of the SCF cycle, predicting molecular dynamics

Abstract:

Density Functional Theory (DFT) is the cornerstone of computational materials science and quantum chemistry, applied across diverse systems from molecules to surfaces and bulk materials. However, solving the Kohn-Sham equations requires an iterative self-consistent field (SCF) loop starting from an initial guess. For large-scale simulations spanning hundreds of atoms or high-throughput screenings, this iterative loop becomes computationally prohibitive, even when utilizing linear-scaling DFT methods. To overcome this bottleneck, we introduce the mcv2v-cGAN model, an equivariant conditional generative adversarial network built upon a previous 2D framework for interacting charge densities [1], performing a multi-channel voxel-to-voxel mapping. The model maps easily accessible structural information, specifically the superposition of atomic densities (SAD) and the neutral atom potential (VNA), directly to the final self-consistent (SC) charge density, while structurally preserving strict charge conservation. By deploying this predicted density as a highly optimized heuristic starting guess within the SIESTA software package, the number of required SCF iterations for the QM9 dataset is reduced by 40%. This significantly outperforms current state-of-the-art methods, which achieve reductions of 26.7% using graph neural networks (GNNs) [2] and 35.5% using ResNet [3]. We demonstrate the robustness and scale-free nature of this model across diverse, complex datasets: the QM9 database, non-fullerene acceptor (NFA) organic molecules up to 106 atoms, and silicon supercells with vacancies containing 245 atoms (Si245). The model consistently yields highly accurate predictions for both self-consistent charge densities and derived physical observables. Furthermore, this approach significantly reduces total computational time for molecular dynamics (MD) simulations of organic molecules by 59.75%, where the MD trajectory is driven entirely by the predicted heuristic charge density without performing a single SCF iteration, thus paving the way for accelerated materials discovery.

References:

[1] Pantis-Simut, C.-A., Preda, A. T., Ion, L., Manolescu, A. & Alexandru Nemnes, G. Mapping confinement potentials and charge densities of interacting quantum systems using conditional generative adversarial networks. Machine Learning: Science and Technology 4, 025023 (2023).

[2] Koker, T., Quigley, K., Taw, E., Tibbetts, K. & Li, L. Higher-order equivariant neural networks for charge density prediction in materials. npj Computational Materials 10, 161 (2024).

[3] Li, C., Sharir, O., Yuan, S. & Chan, G. K.-L. Image super-resolution inspired electron density prediction. Nature Communications 16, 4811 (2025)

Acknowledgement:

This work was supported by a grant of the Romanian Ministry of Education and Research, under project number PN 23210204 and partly from IFA-CERN 07/2024 project and benefited from services and resources provided by EGI, with the dedicated support of CLOUDIFIN for hosting of the interactive platform.