Saved in:
Bibliographic Details
Main Authors: Hou, Bowen, Xu, Xian, Wu, Jinyuan, Qiu, Diana Y.
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2507.05480
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866911044674256896
author Hou, Bowen
Xu, Xian
Wu, Jinyuan
Qiu, Diana Y.
author_facet Hou, Bowen
Xu, Xian
Wu, Jinyuan
Qiu, Diana Y.
contents Recently, radical progress in machine learning (ML) has revolutionized computational materials science, enabling unprecedentedly rapid materials discovery and property prediction, but the quantum many-body problem -- which is the key to understanding excited-state properties, ranging from transport to optics -- remains challenging due to the complexity of the nonlocal and energy-dependent interactions. Here, we propose a symmetry-aware, grid-free, transformer-based model, MBFormer, that is designed to learn the entire many-body hierarchy directly from mean-field inputs, exploiting the attention mechanism to accurately capture many-body correlations between mean-field states. As proof of principle, we demonstrate the capability of MBFormer in predicting results based on the GW plus Bethe Salpeter equation (GW-BSE) formalism, including quasiparticle energies, exciton energies, exciton oscillator strengths, and exciton wavefunction distribution. Our model is trained on a dataset of 721 two-dimensional materials from the C2DB database, achieving state-of-the-art performance with a low prediction mean absolute error (MAE) on the order of 0.1-0.2 eV for state-level quasiparticle and exciton energies across different materials. Moreover, we show explicitly that the attention mechanism plays a crucial role in capturing many-body correlations. Our framework provides an end-to-end platform from ground states to general many-body prediction in real materials, which could serve as a foundation model for computational materials science.
format Preprint
id arxiv_https___arxiv_org_abs_2507_05480
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle MBFormer: A General Transformer-based Learning Paradigm for Many-body Interactions in Real Materials
Hou, Bowen
Xu, Xian
Wu, Jinyuan
Qiu, Diana Y.
Materials Science
Computational Physics
Recently, radical progress in machine learning (ML) has revolutionized computational materials science, enabling unprecedentedly rapid materials discovery and property prediction, but the quantum many-body problem -- which is the key to understanding excited-state properties, ranging from transport to optics -- remains challenging due to the complexity of the nonlocal and energy-dependent interactions. Here, we propose a symmetry-aware, grid-free, transformer-based model, MBFormer, that is designed to learn the entire many-body hierarchy directly from mean-field inputs, exploiting the attention mechanism to accurately capture many-body correlations between mean-field states. As proof of principle, we demonstrate the capability of MBFormer in predicting results based on the GW plus Bethe Salpeter equation (GW-BSE) formalism, including quasiparticle energies, exciton energies, exciton oscillator strengths, and exciton wavefunction distribution. Our model is trained on a dataset of 721 two-dimensional materials from the C2DB database, achieving state-of-the-art performance with a low prediction mean absolute error (MAE) on the order of 0.1-0.2 eV for state-level quasiparticle and exciton energies across different materials. Moreover, we show explicitly that the attention mechanism plays a crucial role in capturing many-body correlations. Our framework provides an end-to-end platform from ground states to general many-body prediction in real materials, which could serve as a foundation model for computational materials science.
title MBFormer: A General Transformer-based Learning Paradigm for Many-body Interactions in Real Materials
topic Materials Science
Computational Physics
url https://arxiv.org/abs/2507.05480