Deep Density: circumventing the Kohn-Sham equations via symmetry preserving neural networks

Leonardo Zepeda-Núñez, Yixiao Chen, Jiefu Zhang, Weile Jia, Linfeng Zhang, Lin LinJournal of Computational Physics 2019 · DOI

TL;DR

This paper introduces "Deep Density," a novel deep learning approach that directly predicts electron density from atomic configurations, bypassing computationally expensive Kohn-Sham equations in density functional theory (DFT). By leveraging a symmetry-preserving, localized neural network architecture inspired by Deep Potential, Deep Density achieves high accuracy and excellent transferability across various insulating and metallic systems, including large molecular and condensed matter systems.

What Problem Does This Paper Address?

Kohn-Sham Density Functional Theory (KS-DFT) is a cornerstone of electronic structure calculations, but its computational cost scales cubically with system size, making large-scale simulations prohibitive. While linear-scaling methods exist, they often struggle with metallic systems. Furthermore, improving the accuracy of KS-DFT is limited by the inherent approximations in exchange-correlation functionals. This paper aims to address both the computational bottleneck and the accuracy limitations by developing a machine learning model that directly predicts the self-consistent electron density, thereby circumventing the need to solve the Kohn-Sham equations.

Methods

The core of the Deep Density method lies in its neural network architecture, which is designed to be both localized and symmetry-preserving. The electron density is decomposed into contributions from local atomic environments, inspired by the Deep Potential framework. Key aspects of the architecture include:

  • Locality: The electron density is modeled as a sum of local contributions from each atom and its neighbors. This locality is enforced by defining an interaction list within a cutoff radius ($R_c$), ensuring that the computational cost scales linearly with system size.
  • Symmetry Preservation: The network is designed to be invariant to translations, rotations, and permutations of identical atoms. This is achieved through a descriptor network that encodes local atomic environments in a rotationally invariant manner.
  • Direct Density Prediction: Unlike previous approaches that might predict coefficients for a basis set, Deep Density directly outputs the electron density at specific points.
  • Feature Mapping and Fitting Networks: Local atomic environments are first mapped to a feature space using descriptor networks. These features are then fed into fitting networks to construct the local density contributions.
  • Exponential Decay with Linear Correction: The local density contribution is modeled with an exponential decay term that captures behavior far from the atom, combined with a linear correction and a non-linear term to handle local variations and potential issues with pseudopotentials.

Datasets

The paper utilizes existing datasets for training and evaluation, as well as generating new data for specific systems.

  • One-Dimensional (1D) Model Systems: Custom 1D systems (insulating, metallic, and mixed metallic-insulating) were generated using a simplified Hartree-Fock model.
  • Organic Molecules: Configurations were taken from a dataset [34] for ethane (C2H6), isobutane, and n-butane.
  • Water Systems: Configurations were generated using the DeePMD-kit package [35] for systems of 32, 64, 128, 256, and 512 water molecules.
  • Aluminum Systems: Configurations were generated using the DeePMD-kit package [35] for systems of 32, 108, and 256 aluminum atoms.
  • PWDFT and VASP: Electron densities for these systems were computed using the PWDFT (Perdew-Burke-Ernzerhof functional) and VASP (Vienna Ab initio Simulation Package) codes for comparison.

Key Insights and Ideas

  • "Nearsightedness" Extends to Metallic Systems: The paper demonstrates that even in metallic systems, where the standard "nearsightedness" principle (exponential decay of density response) breaks down, the "screening effect" provides a level of localization that Deep Density can exploit. This is crucial for achieving linear scaling in metallic systems.
  • Direct Density Prediction is Feasible and Accurate: By directly targeting the electron density, Deep Density bypasses the iterative self-consistent field procedure of KS-DFT, offering a significant computational advantage. The results show high accuracy across diverse system types.
  • Symmetry Preservation is Key for Transferability: The explicit incorporation of physical symmetries into the neural network architecture is shown to be vital for the model's ability to generalize to larger systems and different configurations than those seen during training.
  • Localized Contributions Simplify Complexity: Decomposing the total density into local atomic contributions allows the network to learn complex interactions without needing to model the entire system globally, leading to efficient training and inference.

Key Results

  • High Accuracy: Deep Density achieves excellent accuracy for 1D insulating, metallic, and mixed systems, with relative errors typically in the range of 0.01% to 0.43%.
  • Transferability: The model demonstrates strong transferability to larger systems. For water, errors remain around 0.5% (l² relative error) even for systems up to 512 molecules. For 1D systems, a model trained on 8 atoms can accurately predict densities for 32 atoms.
  • Performance on 3D Systems: For small organic molecules (ethane, isobutane, butane), relative errors are around 0.17-0.31%. For water, errors are around 0.5-1.0%.
  • Challenges with Aluminum: The paper notes that accuracy for aluminum systems is lower (relative l² errors up to 2.5% with PWDFT, and higher with VASP), with errors primarily originating from valence electrons. This suggests potential limitations in the screening effect for certain metallic systems or the need for more advanced network architectures.
  • Linear Scaling: The localized nature of the architecture ensures a linear scaling of computational cost with system size.

Limitations and Future Work

  • Aluminum Accuracy: The authors acknowledge that the accuracy for aluminum systems is lower than for other tested systems. They suggest that this might be due to insufficient screening compared to toy models or the need for different neural network architectures.
  • GPU Acceleration: Currently, descriptors are computed on CPUs, which can be a bottleneck. Future work will focus on GPU-based implementation for improved efficiency.
  • Point-wise Evaluation Independence: The paper suggests a modification to make point-wise evaluation independent of system size, further enhancing efficiency.
  • Improved Architectures: Exploring different neural network architectures is proposed as a way to improve accuracy, particularly for metallic systems.
  • Accelerated Training: Investigating methods to accelerate the training process is also a future direction.

Summary updated

BibTeX
@article{zepedanunez2019deep,
  author = {Leonardo Zepeda-Núñez and Yixiao Chen and Jiefu Zhang and Weile Jia and Linfeng Zhang and Lin Lin},
  title = {{Deep Density: circumventing the Kohn-Sham equations via symmetry preserving neural networks}},
  journal = {Journal of Computational Physics},
  year = {2019},
  doi = {10.1016/j.jcp.2021.110523},
}
Figure