Quantum mechanical static dipole polarizabilities in the QM7b and AlphaML showcase databases

Yang Yang, Ka Un Lao, David M. Wilkins, Andrea Grisafi, Michele Ceriotti, Robert A. DiStasioScientific 2019 · DOI

TL;DR

This paper presents a comprehensive dataset of static dipole polarizabilities for over 7,000 molecules, calculated using high-accuracy quantum chemistry methods. This dataset serves as a crucial benchmark for evaluating and developing faster, approximate methods like Density Functional Theory (DFT) for predicting this important molecular property.

What Problem Does This Paper Address?

Predicting how a molecule will respond to an external electric field – its dipole polarizability (α) – is vital for understanding intermolecular interactions, spectroscopy, and developing accurate molecular models. While Density Functional Theory (DFT) is a workhorse for many molecular properties, it often struggles with response properties like α, yielding less accurate results compared to more computationally expensive quantum chemistry methods. The challenge lies in obtaining highly accurate "ground truth" data for α across a diverse range of molecules to properly benchmark and improve DFT methods.

Methods

The authors employed two main computational approaches:

  • High-Accuracy Quantum Chemistry (QC): They used Linear-Response Coupled-Cluster Singles and Doubles (LR-CCSD) with a doubly-augmented correlation-consistent polarized valence double-zeta (d-aug-cc-pVDZ) basis set. LR-CCSD is a sophisticated method for capturing electron correlation, and the augmented basis set helps mitigate errors from incomplete basis sets, making it a gold standard for accurate polarizability calculations.
  • Density Functional Theory (DFT): They used two popular hybrid functionals, B3LYP and SCAN0, with both d-aug-cc-pVDZ and d-aug-cc-pVTZ basis sets. DFT calculations were performed using the finite-field method with orbital relaxation.

The paper also details the processing of molecular geometries, including centering and reordering, to ensure consistency and reproducibility.

Datasets

The study utilizes two primary datasets:

  • QM7b Database: This dataset contains 7,211 small organic molecules, commonly used for machine learning in chemistry.
  • AlphaML Showcase Database: A newly introduced dataset of 52 larger and more diverse organic molecules, including nucleobases, amino acids, carbohydrates, and pharmaceuticals. This dataset was specifically designed to test the transferability of machine learning models for polarizability.

All calculated properties are provided in standard XYZ file format, with detailed information in the "comment line."

Key Insights and Ideas

  • DFT vs. QC for Polarizability: The study quantitatively demonstrates that while DFT methods like B3LYP and SCAN0 can provide reasonable estimates for the isotropic polarizability (α<sub>iso</sub>), they exhibit larger errors for the anisotropic polarizability (α<sub>aniso</sub>), which captures the tensorial nature of the response. This highlights the inherent difficulty DFT faces in accurately describing the subtle electronic rearrangements that govern anisotropy.
  • SCAN0's Improvement: The SCAN0 functional shows a significant improvement over B3LYP in predicting α<sub>iso</sub> when compared to the high-accuracy CCSD/daDZ reference, particularly in the QM7b database. This suggests SCAN0 might be a more robust choice for this property.
  • Basis Set Effects: The comparison between B3LYP/daDZ and B3LYP/daTZ shows that for DFT, increasing the basis set size from double-zeta to triple-zeta has a relatively minor impact on α<sub>iso</sub>, indicating rapid convergence for this property.
  • Molecular Size and Complexity: The paper notes that both DFT methods struggle more with molecules that are larger and more complex, leading to larger errors in polarizability predictions. This is attributed to the increasing non-additivity of electronic effects in such systems.
  • Data for Machine Learning: The curated datasets, especially the AlphaML showcase, are invaluable for training and validating machine learning models aimed at predicting polarizabilities, bridging the gap between computationally expensive QC and practical applications.

Key Results

  • Comprehensive Data: The paper provides static dipole polarizability tensors (α), along with other properties like dipole and quadrupole moments, for 7,211 molecules in QM7b and 52 in AlphaML.
  • Method Comparison:
    • For α<sub>iso</sub> in QM7b, B3LYP/daDZ shows a mean error of ~2.5% compared to CCSD/daDZ, while SCAN0/daDZ reduces this error to ~1.3%.
  • For α<sub>aniso</sub> in QM7b, B3LYP/daDZ shows errors of ~9.3%, and SCAN0/daDZ ~7.9%.
  • The AlphaML database shows similar trends, with SCAN0 generally outperforming B3LYP for α<sub>iso</sub>.
  • Basis Set Convergence: The difference between B3LYP/daDZ and B3LYP/daTZ for α<sub>iso</sub> is minimal (~0.1%), indicating good basis set convergence for DFT.
  • Validation: CCSD/daDZ calculations with the d-aug-cc-pVDZ basis set were validated against the larger d-aug-cc-pVTZ basis set, showing small basis set incompleteness errors (MAPE < 1% for α<sub>aniso</sub>).

Limitations and Future Work

  • Computational Cost: The high computational cost of LR-CCSD limits its application to larger molecules. For the ten largest molecules in the AlphaML showcase, only energy calculations could be performed with CCSD, and polarizabilities were derived using an orbital-unrelaxed finite-field method.
  • DFT Limitations: While SCAN0 shows improvement, DFT methods still struggle with accurately predicting the anisotropic polarizability, especially for larger and more complex molecules.
  • Future Directions: The authors suggest that the generated datasets are suitable for developing next-generation force fields, density functionals, and machine learning models for predicting polarizabilities. Further work could explore dynamic polarizabilities and a more extensive exploration of DFT functionals.

Summary updated

Abstract

While density functional theory (DFT) is often an accurate and efficient methodology for evaluating molecular properties such as energies and multipole moments, this approach often yields larger errors for response properties such as the dipole polarizability (α), which describes the tendency of a molecule to form an induced dipole moment in the presence of an electric field. In this work, we provide static α tensors (and other molecular properties such as total energy components, dipole and quadrupole moments, etc.) computed using quantum chemical (QC) and DFT methodologies for all 7,211 molecules in the QM7b database. We also provide the same quantities for the 52 molecules in the AlphaML showcase database, which includes the DNA/RNA nucleobases, uncharged amino acids, several open-chain and cyclic carbohydrates, five popular pharmaceutical molecules, and 23 isomers of C<sub>8</sub>H<sub>n</sub>. All QC calculations were performed using linear-response coupled-cluster theory including single and double excitations (LR-CCSD), a sophisticated approach for electron correlation, and the d-aug-cc-pVDZ basis set to mitigate basis set incompleteness error. DFT calculations employed the B3LYP and SCAN0 hybrid functionals, in conjunction with d-aug-cc-pVDZ (B3LYP and SCAN0) and d-aug-cc-pVTZ (B3LYP).
BibTeX
@inproceedings{yang2019quantum,
  author = {Yang Yang and Ka Un Lao and David M. Wilkins and Andrea Grisafi and Michele Ceriotti and Robert A. DiStasio},
  title = {{Quantum mechanical static dipole polarizabilities in the QM7b and AlphaML showcase databases}},
  booktitle = {Scientific},
  year = {2019},
  doi = {10.1038/s41597-019-0157-8},
}
Figure