Compact Gaussian basis sets for stochastic DFT calculations

Marcel David Fabian, Eran Rabani, Roi Baer · 2025

Read Compact Gaussian basis sets for stochastic DFT calculations by Marcel David Fabian, Eran Rabani, Roi Baer (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-2504-17115.txt.

Introduction

Marcel Fabian, Eran Rabani, and Roi Baer built compact basis sets for a randomized quantum lab science method in a 2025 paper. Stochastic density working theory replaces exact electronic form sums with statistical sampling, working on real-space grids. Standard basis sets borrowed from plain math work proved wasteful there. Their new single- and double-zeta Gaussian sets are built to resemble norm-conserving pseudo-orbitals for common elements while occupying minimal room in both real and momentum space.

That compactness suits the auxiliary grids the stochastic method uses, cutting cost without giving up accuracy. Tests report accuracy comparable to set up sets with better speed for grid-based stochastic runs. Tailoring the basis to the method, rather than inheriting one, is the lesson.

Catalog record

CREATORMarcel David Fabian, Eran Rabani, Roi Baer
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE45 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2504-17115

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Compact Gaussian basis sets for stochastic DFT calculations Marcel David Fabian, Eran Rabani, Roi Baer Abstract This work presents new Gaussian single- and double-zeta basis sets optimized for stochastic density functional theory (sDFT) using real-space auxiliary grids. Previous studies showed standard basis sets like STO-3G and 6-31G are sub-optimal for this approach. Our basis-set’s Gaussian-type orbitals (GTOs) resemble norm-conserving pseudo-orbitals for H, C, N, O, F, and Si, but minimize real-space and momentum-space support. These basis sets achieve accuracy comparable to established sets while offering improved efficiency for sDFT calculations with auxiliary grids. I Introduction The nearsightedness principle in electronic structure ( Kohn 1996 ) asserts that the local physical and chemical properties of molecules and materials are primarily determined by their immediate atomic environments. In non-metallic systems this principle is reflected in the exponential decay of the real space representation of the density matrix, and it underlies the development of linear-scaling electronic structure methods ( Goedecker 1999 ; Bowler and Miyazaki 2012 ; Ochsenfeld et al. 2007 ) . In these real-space approaches, computational efficiency can be obtained by employing localized, atom-centered basis functions to represent quantum mechanical operators and wave functions. For example, nonorthogonal generalized Wannier functions in ONETEP ( Prentice et al. 2020 ) , a combination of bsplines (“basis spline”) functions (finite range splines centered on atoms) and pseudoatomic orbitals in CONQUEST ( Nakata et al. 2020 ) , numerical pseudo atomic orbitals (NAOs) in SIESTA ( García et al. 2020 ) , and Gaussian-Type Orbitals (GTOs) in ORCA ( Neese et al. 2020 ) , CP2K’s Quickstep ( VandeVondele et al. 2005 ; Kühne et al. 2020 ) , Ergo ( Rudberg et al. 2018 ) and Dalton’s LSDalton ( Olsen et al. 2020 ) project. Localized basis sets are not the only way to achieve linear scaling: stochastic Density Functional Theory (sDFT) ( Baer et al. 2013 ; Baer et al. 2022 ) achieves this by statistical sampling with stochastic wave functions and noise reduction techniques ( Neuhauser et al. 2014 ; Chen et al. 2019a ; Li et al. 2019 ; Nguyen et al. 2021 ) . sDFT is based on representing the Kohn-Sham density by the expected value: n\left(\boldsymbol{r}\right)=2\text{E}\left[\xi\left(\boldsymbol{r}\right)^{2}\right], (1) where \xi\left(\boldsymbol{r}\right)=\left(1+e^{\beta\left(H_{ks}-\mu\right)}\right)^{-1/2}\chi\left(\boldsymbol{r}\right) (2) and \chi\left(\boldsymbol{r}\right) is a stochastic wave function. Operating on \chi is the square root of the Fermi-Dirac operator, where H_{KS} is the Kohn-Sham Hamiltonian, composed of the kinetic energy operator, the Hartree and exchange correlation potentials, and the sum, over all nuclei, of their local and non-local (norm-conserving) pseudopotentials ( Bachelet et al. 1982 ; Troullier and Martins 1991 ) . \beta is the inverse temperature (taken large for zero temperature calculations) and \mu is the chemical potential, adjusted to obtain the desired number of electrons in the system. The Fermi-Dirac operator is applied using a Chebyshev expansion ( Baer and Head-Gordon 1997 ; Goedecker and Colombo 1994 ; Kosloff 1988 ) of length proportional to \beta\Delta E where \Delta E is the spectral energy range of H_{KS} . The sDFT and DFT calculations are implemented in our in-house code, called Inbar ( Inbar is the Hebrew word for Amber , which is Electron in ancient Greek). In the real-space grid calculations, implemented in RS-Inbar, ( Baer et al. 2013 ; Arnon et al. 2017 ) \chi\left(\boldsymbol{r}_{g}\right) takes values of \pm 1 with equal probability on each grid point \boldsymbol{r}_{g} . When using a plane waves basis, implemented in PW-Inbar, ( Cytter et al. 2018 ; Chen et al. 2019b ; Hadad et al. 2024 ) \chi\left(\boldsymbol{r}\right)=\sum_{\boldsymbol{G}}\tilde{\chi}_{\boldsymbol{G}}e^{i\boldsymbol{G}\cdot\boldsymbol{r}} , \boldsymbol{G} are reciprocal lattice wave vectors and \tilde{\chi}_{\boldsymbol{G}} samples uniformly the unit circle in the complex plane. Finally, and relevant to the present work, in basis set calculations, implemented in BS-Inbar ( Fabian et al. 2019 ; Shpiro et al. 2022 ) , \chi\left(\boldsymbol{r}\right)=\sum_{m}\chi_{m}\psi_{m}\left(\boldsymbol{r}\right) , where \psi_{m}\left(\boldsymbol{r}\right) are contracted Gaussian-type orbitals (discussed in detail below) and the coefficients \chi_{m} take values of \pm 1 with equal probability. Even though sDFT achieves linear scaling without using localized atom-centered real-space basis functions, it can still greatly benefit from them in several ways. These representations significantly boost the speed of calculations ( Shpiro et al. 2022 ; Fabian et al. 2019 ) by providing a smaller energy range \Delta E of the Hamiltonian (relative to grids or plane waves), allowing for shorter Chebyshev expansions of the Fermi operator, and they enable economical application of observable operators, represented as sparse matrices. In addition, localized basis sets offer flexibility in choosing the local accuracy of representation according to the role the atoms play in the chemical process being explored. This flexibility could be beneficial when studying large systems where the atoms directly participating in the chemical reaction require higher precision representation than others. For example, hydrogen/oxygen atoms within water molecules inside the PSII photosynthetic reaction center require high-quality atom-centered basis functions. In contrast, a minimal or double zeta basis set may be sufficient for describing the same elements in a distant water molecule. Our previous publications on sDFT with localized basis functions used standard valence electron GTOs ( Fabian et al. 2019 ; Shpiro et al. 2022 ) . These functions, however, are not optimal when used within our code. One issue is that they do not efficiently combine with our auxiliary real-space grid, with which we represent the norm-conserving pseudopotentials and solve the Poisson equation using fast Fourier transform methods. Moreover, they have suboptimal accuracy when used with norm-conserving pseudopotentials. In this paper, we develop basis sets that are more suitable for the unique features of our approach. The Quickstep code, which, like ours, combines Gaussian functions, fast Fourier transforms, and pseudopotentials, has also benefited from developing specialized basis sets ( VandeVondele and Hutter 2007 ; Zijlstra et al. 2009 ) . These are designed to optimally integrate with Quickstep’s use of separable dual-space Gaussian pseudopotentials ( Goedecker et al. 1996 ) . However, as we demonstrate below, these latter basis sets are not optimal for our framework. Section II defines the fundamental quantities and notation for Gaussian-type orbitals and auxiliary grids used in our calculations. We then explain how some characteristics of the basis functions affect the speed of our code, leading to a definition of the goals of our basis set design. Finally, we explain how we constructed the basis functions. In section III , we demonstrate the efficacy of our basis functions. We compare their accuracy, robustness to grid coarsening and small support on the grid to those of Pople’s minimal basis set. By splitting off one primitive from each contraction, we create a double zeta basis set, which we scale and benchmark while comparing it to the double zeta basis set of ref. ( Zijlstra et al. 2009 ) , which, unlike any other basis set we examined, is compatible with pseudopotentials and reasonably efficient on our real-space auxiliary grid. The new single and double valence basis sets developed in this work offer comparable accuracy to standard Pople basis sets. However, they exhibit greater robustness to grid coarsening. This allows for the use of coarser auxiliary grids with minimal impact on accuracy, enhancing computational efficiency without sacrificing precision. II Method II.1 Definitions and general considerations II.1.1 The contracted Gaussian Type Orbitals

Read the complete text