Accurate GW frontier orbital energies of 134 kilo molecules

Artem Fediai, Patrick Reiser, Jorge Enrique Olivares Peña, Pascal Friederich, Wolfgang Wenzel

Abstract

HOMO and LUMO energies are critical molecular properties that typically require high accuracy computations for practical applicability. Until now, a comprehensive dataset containing sufficiently accurate HOMO and LUMO energies has been unavailable. In this study, we introduce a new dataset of HOMO/LUMO energies for QM9 compounds, calculated using the GW method. The GW method offers adequate HOMO/LUMO prediction accuracy for diverse applications, exhibiting mean unsigned errors of approximately 100 meV in the GW100 benchmark dataset. This database may serve as a benchmark for HOMO/LUMO prediction, delta-learning, and transfer learning, particularly for larger molecules where GW is the most accurate but still numerically feasible method. We anticipate that this dataset will enable the development of more accurate machine learning models for predicting molecular properties.

Introduction: The Frontier Orbital Challenge

In the quest to design new organic electronics, catalysts, and functional materials, scientists rely on "frontier orbitals": the Highest Occupied Molecular Orbital (HOMO) and the Lowest Unoccupied Molecular Orbital (LUMO). These energy levels determine a molecule's ionization potential and electron affinity, serving as fundamental descriptors for chemical reactivity and electronic performance.

While Density Functional Theory (DFT) is the standard computational tool due to its efficiency, it systematically and severely underestimates the HOMO-LUMO gap—often by as much as 5 eV. More accurate methods exist but are typically too computationally expensive for large-scale datasets. This has created a significant "data gap" for machine learning researchers who require high-quality labels to train predictive models. This study bridges that gap by providing a massive new dataset of high-accuracy frontier orbital energies for over 132,000 molecules from the QM9 database.

Methodology: Quasiparticles and Extrapolation

The researchers employed the GW method, a many-body perturbation theory approach that treats electrons as "quasiparticles." Unlike standard DFT, GW is specifically designed to calculate excitation spectra accurately. To ensure the highest possible precision, the team implemented several technical strategies:

  • Levels of Theory: Calculations were performed using both "one-shot" ($G_0W_0@PBE$) and eigenvalue self-consistent GW ($GW@PBE$).
  • Basis Set Extrapolation: Because GW results are highly sensitive to basis set size, the team computed values using two different sets (aug-cc-DZVP and aug-cc-TZVP) and applied a linear extrapolation scheme to estimate the "infinite basis set limit."
  • Computational Scale: Utilizing the CP2K quantum chemistry package, the project consumed approximately 7.4 million CPU hours to process the QM9 library.

Results: A New Gold Standard

The resulting GW-QM9 dataset contains converged calculations for 132,151 molecules (99% of the original QM9 set). Validation against the GW100 benchmark showed a mean unsigned error (MUE) of only 28 meV, confirming the implementation's high accuracy.

Distribution of HOMO (a) and LUMO (b) energies across the QM9 dataset, comparing DFT results with various GW methods. The significant shift highlights the systematic correction provided by GW.

As shown in the distributions above, the GW method provides a massive correction to DFT results. On average, GW-calculated HOMO levels were approximately 3.2 eV lower than those calculated with PBE-DFT.

Discussion: The Non-Linearity of DFT Errors

A key insight from this study is that the "DFT gap" is not a simple constant offset. The error in DFT depends on the specific molecule and the orbital energy itself. This non-linearity means that simple empirical corrections to DFT are insufficient for high-precision applications.

Correlation between GW and DFT energies for HOMO (blue) and LUMO (orange). The deviation from the diagonal black line illustrates that the DFT error is non-linear and molecule-dependent.

Furthermore, the correlation between DFT and GW results is significantly weaker for LUMO than for HOMO. Unoccupied orbitals are more sensitive to delocalization and screening effects, making them a rigorous "stress test" for computational methods. The study also confirms that basis set extrapolation is essential; reaching the nominal 0.1 eV accuracy of the GW method requires moving beyond standard finite basis sets.

Conclusion and Future Outlook

While the GW method is computationally demanding, this study demonstrates its feasibility for large-scale molecular datasets. The authors anticipate that the GW-QM9 dataset will become a cornerstone for molecular machine learning, particularly for delta-learning (correcting cheap DFT calculations) and transfer learning. By providing a "gold standard" for 134,000 molecules, this work paves the way for more accurate virtual screening and the discovery of next-generation functional materials.

Summary updated

Abstract

HOMO and LUMO energies are critical molecular properties that typically require high accuracy computations for practical applicability. Until now, a comprehensive dataset containing sufficiently accurate HOMO and LUMO energies has been unavailable. In this study, we introduce a new dataset of HOMO/LUMO energies for QM9 compounds, calculated using the GW method. The GW method offers adequate HOMO/LUMO prediction accuracy for diverse applications, exhibiting mean unsigned errors of 100 meV in the GW100 benchmark dataset. This database may serve as a benchmark of HOMO/LUMO prediction, delta-learning, and transfer learning, particularly for larger molecules where GW is the most accurate but still numerically feasible method. We anticipate that this dataset will enable the development of more accurate machine learning models for predicting molecular properties.
BibTeX
@inproceedings{fediai2023accurate,
  author = {Artem Fediai and Patrick Reiser and Jorge Enrique Olivares Peña and Pascal Friederich and Wolfgang Wenzel},
  title = {{Accurate GW frontier orbital energies of 134 kilo molecules}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41597-023-02486-4},
}
Figure