A neural-network-based Python package for performing large-scale atomic CI using pCI and other high-performance atomic codes Journal: Computer Physics Communications

Pavlo Bilous, Charles Cheung, Marianna S. Safronova · 2025

Read A neural-network-based Python package for performing large-scale atomic CI using pCI and other high-performance atomic codes Journal: Computer Physics Communications by Pavlo Bilous, Charles Cheung, Marianna S. Safronova (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-2503-01379.txt.

Introduction

Pavlo Bilous, Charles Cheung, and Marianna Safronova present a 2025 software paper for atomic physics. Exact models of atoms need vast sums over electron states, which can overwhelm even supercomputers. A neural network trick recently cut that burden by picking the states that matter most.

The paper ships a Python package that brings the trick to working codes. It links with a leading atomic code and other high-performance programs in the field. Users can run large configuration interaction jobs with network-guided state picking from a friendly language. The authors frame the package as a bridge between the new method and daily research. References point to the two base papers that prove and document the approach.

Catalog record

CREATORPavlo Bilous, Charles Cheung, Marianna S. Safronova
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE33 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-01379

Browse the full library · About the Tower of Babel

Full text

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

A neural-network-based Python package for performing large-scale atomic CI using pCI and other high-performance atomic codes Journal: Computer Physics Communications Pavlo Bilous, Charles Cheung, Marianna S. Safronova Abstract Modern atomic physics applications in science and technology pose ever higher demands on the precision of computations of properties of atoms and ions. Especially challenging is the modeling of electronic correlations within the configuration interaction (CI) framework, which often requires expansions of the atomic state in huge bases of Slater determinants or configuration state functions. This can easily render the problem intractable even for highly efficient atomic codes running on distributed supercomputer systems. Recently, we have successfully addressed this problem using a neural-network (NN) approach [1]. In this work, we present our Python code for performing NN-supported large-scale atomic CI using pCI [2] and other high-performance atomic codes. PROGRAM SUMMARY Program Title: nn_for_pci CPC Library link to program files: (to be added by Technical Editor) Developer’s repository link: https://github.com/pavlobilous/nn_for_pci Code Ocean capsule: (to be added by Technical Editor) Licensing provisions: GPLv3 Programming language: Python Nature of problem: Exponential scaling of the basis size in the atomic CI approach Solution method: Iterative NN-based selection of the relevant basis elements out of a large CI basis 1 Introduction Advanced applications of computational atomic physics, such as the development of novel atomic clocks or studies of astrophysical spectra, require ever greater precision and, therefore, highly accurate modeling of electronic correlations. The high-precision treatment of electronic correlations can be performed within the configuration interaction (CI) framework [ 1 ] , in which the electronic many-body wave function is expanded in a given basis, usually Slater determinants or configuration state functions (CSFs). Unfortunately, in this approach, higher precision implies an exponential increase of the number of needed basis elements, quickly leading to a numerically intractable eigenvalue problem, even with highly parallelized codes running on advanced supercomputers. This requires the development of approximation methods, in particular, based on a priori selection of the most important basis states. In our recent work [ 2 ] , we presented a neural network (NN) approach to running large-scale atomic CI computations. Following the developments in Refs. [ 3 , 4 , 5 , 6 ] , we introduced important algorithmic improvements that made the method suitable for large computations of atomic spectra without considerable overhead for NN operations. Coupled with the recent release of the pCI atomic structure codes [ 7 ] , our NN-based approach proved to be a powerful tool to tackle the exponential growth of the CI basis, which is undoubtedly one of the most challenging problems in high-precision atomic computations. In this work, we aim to enable such NN-supported CI calculations as presented in Refs. [ 2 , 3 ] to the atomic physics community. We present our Python package implementing the NN-based selection algorithm, along with a demonstration example of a NN-supported CI computation. The computation consists of two-block iterations shown in Fig. 1 : (1) the Python code selects the relevant subbasis from the CI basis using the NN, and (2) an atomic code performs CI with the NN-selected subbasis. CI using an atomic code Basis sortout using NN Figure 1: NN-supported computation iteration. While the Python codes are in principle suitable for supporting any CI atomic code, we demonstrate the NN support for CI computations performed by the recently published pCI atomic code package [ 7 , 2 ] . The pCI codes are suitable for extremely large parallelized CI calculations [ 8 ] and have demonstrated exceptional accuracy for a wide variety of many-electron systems, including neutral atoms, highly charged ions, and negative ions [ 9 , 10 , 11 , 12 , 13 , 14 , 15 ] . We also present how the NN algorithm can be implemented with any other atomic code performing CI. The article is structured as follows. In Sec. 2 , we describe the NN-supported algorithm. Sec. 3 is devoted to the installation and description of the corresponding Python codes. In Sec. 4 , we provide an example of a direct CI computation with the pCI codes. Then in Sec. 5 , we demonstrate how the same example is solved using our implementation of the NN algorithm. In Sec. 6 , we demonstrate how the presented NN support can be coupled with any atomic code performing CI. The article is concluded with a summary in Sec. 7 . 2 NN-supported CI algorithm In the CI approach [ 1 ] , the wave function is represented as an expansion \ket{\Psi}=\sum_{\alpha}c_{\alpha}\ket{\alpha} in a fixed basis \mathcal{B}=\{\ket{\alpha}\} , where \ket{\alpha} are typically Slater determinants or CSFs. The unknown expansion coefficients c_{\alpha} are then obtained by forming the Hamiltonian with matrix elements H_{\alpha\beta}=\braket{\alpha|\hat{H}|\beta} , and then solving the eigenvalue problem on the basis \mathcal{B} . Obtaining high-precision results for complex atomic systems typically requires a huge basis that may render the CI computation intractable. In this case, our NN approach can be used to replace the CI computation by a number of smaller ones on an iteratively growing subbasis \mathcal{B}_{i}\subset\mathcal{B} managed by the NN. In general, the NN first trains on the data obtained in iteration i-1 , and then classifies the candidate basis elements from \mathcal{B}\setminus\mathcal{B}_{i-1} as “important” or “unimportant”. A basis element \ket{\alpha} is considered “important” if its weight in the CI expansion w_{\alpha}=|c_{\alpha}|^{2} exceeds a cut-off x_{i} , set by the user in each iteration. The basis states classified by the NN as important are included in the subbasis \mathcal{B}_{i} , on which a CI calculation with an atomic code is then performed. The latter yields the true CI expansion coefficients, which are subsequently used to give the NN feedback on its performance and retrain it in the next iteration. The basis elements which turned out to be unimportant are excluded from \mathcal{B}_{i} . The cutoff for the next iteration is chosen as x_{i+1}<x_{i} . These iterations are repeated until the corresponding energies obtained at the CI calculation step are converged or until the available computational resources are exhausted. Crucial improvements were introduced in Ref. [ 2 ] to the described procedure, which enabled the use of the NN-supported approach in large-scale CI for atomic spectra without considerable overhead for the NN operation. We highlight these here: • We treat basis states (Slater determinants or CSFs) not separately but in groups of relativistic configurations. The weight of a relativistic configuration \Gamma is the total weight of its determinants: w^{(k)}_{\Gamma}=\sum_{\ket{\alpha}\in\Gamma}w^{(k)}_{\alpha} , where k labels the energy levels. • Computations are performed not for one energy level, but for a number of energy levels at once. The “compound” weight of a relativistic configuration w_{\Gamma} used in the NN-supported algorithm is given as w_{\Gamma}=\max_{k}w^{(k)}_{\Gamma} . • We assume a smaller-scale (direct or NN-supported) computation to have been performed prior, which is used in the first iteration of the presented algorithm, in which the NN is not yet trained. 3 Installation and structure of the package The Python package presented in this work can be cloned directly from the GitHub repository https://github.com/pavlobilous/nn_for_pci . It employs NumPy [ 16 ] and Pandas [ 17 ] for efficient processing of data arrays, as well as TensorFlow [ 18 ] for leveraging the NN functionalities.

Read the complete text