Natural-Orbital-Based Neural Network Configuration Interaction
Louis Thirion, Yorick L. A. Schmerwitz, Max Kroesbergen, Gianluca Levi, Elvar Ö. Jónsson, Pavlo Bilous, Hannes Jónsson, Philipp Hansmann · 2025
Read Natural-Orbital-Based Neural Network Configuration Interaction by Louis Thirion, Yorick L. A. Schmerwitz, Max Kroesbergen, Gianluca Levi, Elvar Ö. Jónsson, Pavlo Bilous, Hannes Jónsson, Philipp Hansmann (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-2510-27665.txt.
Introduction
Louis Thirion and seven co-authors test a better orbital choice for neural-aided molecule calculations in 2025. Configuration interaction builds a molecule's state from a set of electron layouts, keeping only the most key ones. A neural net can pick those layouts fast. Past work used standard Hartree-Fock orbitals, which are tuned at a mean-field level that ignores much electron interplay.
Natural orbitals come from a density matrix built mid-run and reflect that interplay better. Even rough forms cut the count of layouts needed for a set energy accuracy. Tests cover water, ammonia, carbon monoxide, and propane. Fewer layouts for the same accuracy means faster, cheaper quantum chemistry on hard molecules.
Catalog record
| CREATOR | Louis Thirion, Yorick L. A. Schmerwitz, Max Kroesbergen, Gianluca Levi, Elvar Ö. Jónsson, Pavlo Bilous, Hannes Jónsson, Philipp Hansmann |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 27 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2510-27665
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Natural-Orbital-Based Neural Network Configuration Interaction Louis Thirion, Yorick L. A. Schmerwitz, Max Kroesbergen, Gianluca Levi, Elvar Ö. Jónsson, Pavlo Bilous, Hannes Jónsson, Philipp Hansmann Abstract Selective configuration interaction methods approximate correlated molecular ground- and excited states by considering only the most relevant Slater determinants in the expansion. While a recently proposed neural-network-assisted approach efficiently identifies such determinants, the procedure typically relies on canonical Hartree-Fock orbitals, which are optimized only at the mean-field level. Here we assess approximate natural orbitals – eigenfunctions of the one-particle density matrix computed from intermediate many-body eigenstates – as an alternative. Across our benchmarks for H 2 O \text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{O} , NH 3 \text{NH}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}} , CO , and C 3 H 8 \text{C}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} we see a consistent reduction in the required determinants for a given accuracy of the computed correlation energy compared to full configuration interaction calculations. Our results confirm that even approximate natural orbitals constitute a simple yet powerful strategy to enhance the efficiency of neural-network-assisted configuration interaction calculations. I Introduction Accurate computation of molecular ground-state energies is challenging because the formally exact full configuration interaction (CI) solution becomes intractable as the configuration space grows Szabo and Ostlund 2012 . Hartree-Fock offers a compact mean-field reference; the missing electron correlation is commonly partitioned into dynamic, often well captured by Møller–Plesset perturbation theory (MP n ) and coupled cluster (CC) methods Møller and Plesset 1934 ; Bartlett and Musiał 2007 , and static, which calls for multireference treatments such as complete active space self-consistent field (CASSCF) and complete active space second-order perturbation theory (CASPT2) Roos et al. 1980 ; Andersson et al. 1992 . CI represents the wave function in a basis of Slater determinants. Selective CI (SCI) methods retain only the most important determinants, yielding compact and systematically improvable expansions. Classical and modern realizations - configuration interaction using a perturbative selection made iteratively (CIPSI), heat-bath configuration interaction (HCI), semistochastic HCI, and adaptive sampling configuration interaction (ASCI) - differ in selection criteria and the treatment of the residual space, but all deliver high accuracy at reduced cost Huron et al. 1973 ; Holmes et al. 2016 ; Sharma et al. 2017 ; Li et al. 2018 ; Tubman et al. 2020 . Machine learning (ML) has recently introduced a new twist and has been used to accelerate determinant selection by predicting importance directly. Neural classifiers trained during the CI iterations, as in MLCI and active-learning CI, reduce the number of explicitly evaluated determinants while approaching full CI accuracy Coe 2018 ; Jeong et al. 2021 . Building on this idea, we recently introduced a neural-network‑assisted SCI (NNCI) framework that trains a classifier on-the-fly to identify relevant determinants during SCI iterations in condensed matter and quantum chemistry contexts Bilous et al. 2025 ; Schmerwitz et al. 2025 ; Thirion et al. 2024 . NNCI turns out to accurately recover correlation energies with a small fraction of determinants, and our prior results indicate that operating in larger single‑particle bases can be much more beneficial than trying to reach the full CI solution in a small basis Schmerwitz et al. 2025 . This observation motivates coupling SCI with orbital optimization on the single-particle level. A practical limitation of many SCI approaches, including NNCI, is the use of canonical Hartree-Fock molecular orbitals (MOs), which are optimized only at the mean‑field level and do not necessarily yield the most compact many‑body expansions. Natural orbitals (NOs) - eigenfunctions of the one‑particle reduced density matrix (1-RDM) of a correlated state- typically provide a more natural starting point for correlated calculations and accelerate CI convergence. Importantly, they can be approximated inexpensively from preliminary correlated calculations. The present work quantifies the impact of using NOs as the single‑particle basis for a many-body NNCI treatment. We generate approximate NOs from intermediate NNCI solutions and continue the selection in the rotated basis. Benchmarks for \text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{2}}}\text{O} , \text{NH}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}} , CO , and \text{C}{\vphantom{\text{X}}}_{\smash[t]{\text{3}}}\text{H}{\vphantom{\text{X}}}_{\smash[t]{\text{8}}} demonstrate that, for a fixed active space, NOs consistently and significantly reduce the number of determinants required to reach a target accuracy compared with Hartree-Fock orbitals. We provide practical guidance for constructing these orbitals and for integrating single‑particle basis optimization into ML‑accelerated SCI workflows. II Methods II.1 Electronic-structure setup The single-particle basis for the configuration interaction Hamiltonian is generated with GPAW Mortensen et al. 2024 . Hartree-Fock calculations are carried out in the plane-wave mode with a 1000 eV cutoff and a uniform real-space grid of 0.18 Å. Convergence is reached when the squared residual of the Hartree-Fock equations falls below 10^{-11}\,\text{eV}^{2} per valence electron. GPAW employs the projector-augmented wave (PAW) formalism Blöchl 1994 ; Blöchl et al. 2003 together with the frozen-core approximation, i.e. scalar relativistic atomic calculations provide the core electrons, which are not explicitly considered in the subsequent Hartree-Fock and NNCI steps. These settings yield the canonical Hartree-Fock occupied orbitals, while leaving the virtual space undefined. In principle, as many orbitals exist as plane waves in the basis. In practice, only \sim\!100 high-quality orbitals are tractable for many-body calculations. Occupied orbitals are initialized from predefined atomic basis functions, evaluated on the real-space grid, Fourier-transformed, orthonormalized, and then optimized by direct minimization Ivanov et al. 2021a ; Ivanov et al. 2021b . Virtual orbitals are generated analogously from the remaining atomic basis functions, but without optimization. The periodic cell is chosen sufficiently large to accommodate all basis functions up to their cutoff values in real space. For propane, numerical atomic orbital (NAO) basis sets Larsen et al. 2009 are used: four s and two p functions for hydrogen; four s , three p , and two d for carbon. For ammonia, water and carbon monoxide, mixed NAO/Gaussian bases are employed: one NAO s for hydrogen, and one NAO s plus one p for carbon, nitrogen, and oxygen. Gaussian primitives are taken from the aug-cc-pVTZ basis, comprising three s , three p , and two d functions for hydrogen, and three s , three p , three d , and two f functions for carbon, nitrogen, and oxygen. This results in a set of single-particle orbitals \{\varphi_{i}(\bm{r})\} for i\leq\lceil N_{\mathrm{el}}/2\rceil being the canonical Hartree-Fock orbitals and else the virtual orbitals. To construct the Hamiltonian for a subsequent many-body calculation, we compute the single- and two-particle integrals \displaystyle t_{ij} \displaystyle\equiv\int\text{d}\mathbf{r}\varphi^{*}_{i}(\mathbf{r})\left(-\frac{1}{2}\nabla^{2}+V^{\text{eff}}(\mathbf{r})\right)\varphi_{j}(\mathbf{r}) (1) \displaystyle U_{ijkl} \displaystyle\equiv\int\text{d}\mathbf{r}\,\text{d}\mathbf{r^{\prime}}\varphi^{*}_{i}(\mathbf{r})\varphi_{j}(\mathbf{r})\frac{1}{|\mathbf{r}-\mathbf{r}^{\prime}|}\varphi_{k}^{*}(\mathbf{r}^{\prime})\varphi_{l}(\mathbf{r}^{\prime}) (2) where V_{\text{eff}}(\mathbf{r}) is the self-consistent mean-field potential. The electronic Hamiltonian for the many-body system is composed as follows: H=H_{0}+H_{\text{int}}-\text{MF}[H_{\text{int}}], (3) with \displaystyle H_{0} \displaystyle=\;\sum_{i,j,\sigma}\;\;t_{ij}\;\;\;\;c_{i\sigma}^{\dagger}c^{\phantom{\dagger}}_{j\sigma}\,, (4) \displaystyle H_{\text{int}} \displaystyle=\,\sum_{\begin{subarray}{c}i,j,k,l\\ \sigma,\sigma^{\prime}\end{subarray}}\,U_{ijkl}\;c_{i\sigma}^{\dagger}c_{j\sigma}^{\phantom{\dagger}}c^{\dagger}_{k\sigma^{\prime}}c^{\phantom{\dagger}}_{l\sigma^{\prime}}\,, (5) using the one- and two-particle integrals and c_{i\sigma}^{\dagger} ( c_{i\sigma} ) being the creation (annihilation) operator for spin \sigma in molecular orbital i . \mathrm{MF}[H_{\text{int}}] denotes the mean-field decoupled interaction used to remove double counting. II.2 NNCI algorithm Full details are provided in Refs. Bilous et al. 2025 ; Schmerwitz et al. 2025 . Here we only provide a compact summary. The algorithm combines active learning with iterative determinant selection to construct an efficient CI basis for correlated electron systems. Starting from a small initial determinant set, typically the Hartree-Fock solution, the Hamiltonian is repeatedly applied to generate new candidate configurations, which we denote as the pool. A convolutional neural network classifier that is trained on information obtained from a smaller exact diagonalization on determinants sampled from the pool