A Machine Learning Model for Predicting QSGW Band Gaps Using the Partial Density of States in LDA

Shota Takano, Takao Kotani, Masao Obata, Hitoshi Fujii, Kazunori Sato, Harutaka Saito, Tatsuki Oda · 2025

Read A Machine Learning Model for Predicting QSGW Band Gaps Using the Partial Density of States in LDA by Shota Takano, Takao Kotani, Masao Obata, Hitoshi Fujii, Kazunori Sato, Harutaka Saito, Tatsuki Oda (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-2507-19189.txt.

Introduction

Shota Tankano and co-authors build, in a 2025 paper, a machine learning model that can predict a key property of crystals without the slow calculations that give exact values. The property is the band gap, the energy step electrons must clear in a solid. A cheap standard method, the local density approximation, gets this step wrong, while an accurate method called QSGW takes huge computing time.

The team ran QSGW on 1,516 materials from the Materials Project database. They then trained a model to map a cheaper fingerprint, the partial density of states from the cheap method, onto the accurate QSGW gap. The learned model significantly outperformed linear regression approaches. The authors present it as a first model of its kind, and a step toward fast screening of new materials by their electronic structure.

Catalog record

CREATORShota Takano, Takao Kotani, Masao Obata, Hitoshi Fujii, Kazunori Sato, Harutaka Saito, Tatsuki Oda
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE26 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2507-19189

Browse the full library · About the Tower of Babel

Full text

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

A Machine Learning Model for Predicting QSGW Band Gaps Using the Partial Density of States in LDA Shota Takano, Takao Kotani, Masao Obata, Hitoshi Fujii, Kazunori Sato, Harutaka Saito, Tatsuki Oda Abstract An accurate calculation of the band gaps for given crystal structures is highly desirable. However, conventional first-principles calculations based on density functional theory (DFT) within the local density approximation (LDA) fail to predict band gaps accurately. To address this issue, the quasiparticle self-consistent G ​ W GW (QSGW) method is often used, as it is one of the most reliable theoretical approaches for predicting band gaps. However, QSGW requires significant computational resources. To overcome this limitation, we propose to combine QSGW with our new machine learning model named DOSnp. DOSnp is an improvement of DOSnet [V. Fung, Nature Comm. 12, 88 (2021)], with the addition of pooling layers. In this study, we applied QSGW to 1,516 materials from the Materials Project [ https://materialsproject.org/ ] and used DOSnp to predict QSGW band gaps as a function of the partial density of states (PDOS) in LDA. Our results demonstrate that DOSnp significantly outperforms linear regression approaches with linearly independent descriptor generation [ https://github.com/Hitoshi-FUJII/LIDG ]. This model is a prototype for predicting the properties of materials based on PDOS. \usephysicsmodule braket I Introduction For material discovery, such as in solar cell applications, it is often necessary to calculate band gaps for given crystal structures. However, such calculations are challenging because standard first-principles methods, the density functional theory (DFT) in the local density approximation (LDA), or its derivatives, cannot predict band gaps. This is because the band gaps in DFT can not be interpreted as the band gaps of electronic excitations in the independent-particle picture [ 1 ] . DFT cannot do effective masses as well. The first-principles GW approximation originally started in the 1980s [ 2 , 3 , 4 , 5 ] , is a promising alternative to achieve accurate band-gap predictions. This method is based on Hedin’s many-body perturbation theory [ 6 ] . Among the various GW methods, the quasiparticle self-consistent GW (QSGW) method is particularly promising due to its self-consistency within the GW framework [ 1 , 7 ] . QSGW inherently incorporates the so-called U effect, which is usually handled in LDA+ U methods, making it applicable to complex materials such as NiO [ 8 , 9 ] , magnetic metals [ 9 ] , and 3d / 4f impurities [ 10 , 11 ] . Furthermore, QSGW is superior to hybrid functionals such as HSE06 [ 12 ] when the screened Coulomb interaction is place-dependent as in grain boudaries. In contrast, HSE06 requires an ad hoc treatment, dielectric-dependent HSE06 [ 13 ] . In this study, we use QSGW implemented in the ecalj package [ 14 , 15 , 16 ] . One significant advantage of ecalj is its ability to generate eigenfunctions/eigenvalues at any \mathbf{k} points in the Brillouin zone (BZ) without relying on interpolation techniques such as Wannier interpolation [ 17 ] . Furthermore, spin fluctuations and other effects can be incorporated on top of the QSGW results [ 18 , 19 ] . Despite such advantages in QSGW, QSGW requires significant computational resources. To address this challenge, we propose two strategies. The first strategy involves accelerating computations using GPUs [ 20 ] . Recently, Ness et al. performed QSGW calculations on the InAs/Al interface (containing 138 atoms in a supercell) [ 13 ] . The second strategy is to use machine learning. By training a model with band gaps calculated in QSGW, we can promptly predict QSGW band gaps for given crystal structures. In this article, we focus on the second strategy. In this approach, we first apply QSGW to many materials to generate a dataset. To reduce computational effort, we limit our calculations to crystal structures with a small number of atoms per primitive cell. In this study, we calculate 1,516 crystal structures obtained from the Materials Project (MP) [ 21 ] . We use our automated QSGW computational system, ecalj_auto , implemented in the package ecalj , as explained in Appendix VI.1 . To ensure efficient operation of ecalj_auto , the QSGW codes in ecalj have been optimized, particularly for effective memory management. Using this system, we generate a QSGW dataset, train a machine learning model on it, and then use the model to predict band gaps for given crystal structures. There are possible choices for the machine learning model. In this study, we developed a new model DOSnp that takes the partial density of states (PDOS) as input. In our case, we predict the QSGW band gaps using only PDOS in LDA, to avoid large computational efforts. In Ref. [ 22 ] , V. Fung et al. developed DOSnet, a model trained to predict molecular absorption energies from PDOS. Our model DOSnp, is a modification of DOSnet. The main difference lies in the pooling layer for feature vectors of atoms, as we will explain later. In contrast to models that use varieties of material descriptors as inputs [ 23 , 24 , 25 ] , DOSnp is advantageous in simplicity because we only use PDOS as input. PDOS can contain features of crystal structures. Let us explain why we do not utilize another machine learning model for crystal structures, the crystal graph convolutional neural network (CGCNN) [ 26 ] , proposed by Xie and Grossman. CGCNN requires the crystal graph (CG) [ 27 ] as input. This idea of using CG as input is more attractive from a point of view of simplicity than the methods with feature vectors [ 28 ] of Seko et al. However, we are concerned that representing crystal structures by CG may lose important information of crystal structures. Note that CGCNN is a procedure for performing machine learing via the spectrums at the atomic nodes as intermediate quantities, where the spectrums are generated from CG. Pay attention to the fact that the graph spectrum of CG is the density of states (DOS) of the eigenvalues for the adjacency matrix of CG. This is nothing but the DOS with an orbital per atom, where we set all the tight-binding parameters to be unity. Repeating convolutions in CGCNN resembles the recursion method in electronic structure calculations. Even if we can includes informations of atomic species during the convolution process, we guess that it just has the effect modifying the tight-binding parameters. We may improve CGCNN by incorporating multichannel spectrums at atomic nodes with geometrical features, such as atomic distances and three-body angles. However, we guess that such improvements will not provide better spectra distributions at atomic nodes than those obtained by PDOS in the first-principles calculations. In other words, PDOS can be the most meaningful set of spectra distributed at atomic nodes for characterizing the Kohn-Sham Hamiltonian in DFT. This is why PDOS has been widely discussed to explain physical properties. In Sec. II , we explain our method for generating datasets and assessing their quality. We then describe our machine learning model, DOSnp, as a modification of DOSnet. In Sec. III , we present our results demonstrating the superiority of DOSnp compared to a linear regression (LR) approach as a reference. Since the LR approach

Read the complete text