arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2404.14601cond-mat.mtrl-sciphysics.comp-ph
收藏

个体Kohn-Sham态的无监督学习:可解释表示及其对多体效应下游预测的影响

Unsupervised Learning of Individual Kohn-Sham States: Interpretable Representations and Consequences for Downstream Predictions of Many-Body Effects

Bowen Hou, Jinyuan Wu, Diana Y. Qiu

AI总结:

针对电子结构问题,提出使用变分自编码器无监督学习高维DFT波函数,发现其位于低维流形,并利用该潜空间表示监督训练神经网络,实现了对多体效应(GW能带)的高精度下游预测。

摘要:

电子结构问题的表示学习是计算凝聚态与材料物理中机器学习面临的主要挑战。在量子力学第一性原理方法中,Kohn-Sham密度泛函理论(DFT)是理解电子结构的卓越工具,该方法计算的高维波函数是下游计算相关多体激发及相关物理可观测量构建模块。在此,我们使用变分自编码器(VAE)对高维DFT波函数进行无监督学习,并证明这些波函数位于潜空间内的低维流形中。我们的模型自主确定电子结构的最佳表示,避免了以往工作中由于手动特征工程和选择带来的限制。为了证明DFT波函数潜空间表示的实用性,我们将其用于神经网络(NN)的监督训练,以在GW形式体系内下游预测准粒子能带结构,该体系包含了超越DFT的多电子相关性。对于取自计算二维材料数据库(C2DB)的金属和半导体组合测试集,GW预测实现了0.11 eV的低误差,这表明潜空间表示捕获了原始数据中的关键物理信息。最后,我们探讨了VAE表示的可解释性,并表明模型成功的表示学习和下游预测源于VAE潜空间的平滑性,这也使得能够在潜空间中的任意点上生成波函数。我们的工作为研究电子结构和多体物理提供了一个新颖且通用的机器学习框架。

英文摘要:

Representation learning for the electronic structure problem is a major challenge of machine learning in computational condensed matter and materials physics. Within quantum mechanical first principles approaches, Kohn-Sham density functional theory (DFT) is the preeminent tool for understanding electronic structure, and the high-dimensional wavefunctions calculated in this approach serve as the building block for downstream calculations of correlated many-body excitations and related physical observables. Here, we use variational autoencoders (VAE) for the unsupervised learning of high-dimensional DFT wavefunctions and show that these wavefunctions lie in a low-dimensional manifold within the latent space. Our model autonomously determines the optimal representation of the electronic structure, avoiding limitations due to manual feature engineering and selection in prior work. To demonstrate the utility of the latent space representation of the DFT wavefunction, we use it for the supervised training of neural networks (NN) for downstream prediction of the quasiparticle bandstructures within the GW formalism, which includes many-electron correlations beyond DFT. The GW prediction achieves a low error of 0.11 eV for a combined test set of metals and semiconductors drawn from the Computational 2D Materials Database (C2DB), suggesting that latent space representation captures key physical information from the original data. Finally, we explore the interpretability of the VAE representation and show that the successful representation learning and downstream prediction by our model is derived from the smoothness of the VAE latent space, which also enables the generation of wavefunctions on arbitrary points in latent space. Our work provides a novel and general machine-learning framework for investigating electronic structure and many-body physics.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑