Abstract
The QM9 dataset has become a standard benchmark for machine learning methods in molecular chemistry, providing geometries and properties for 133,885 compounds at the B3LYP/6-31G(2df,p) level of theory. However, the accuracy of frontier orbital (HOMO and LUMO) energies from density functional theory (DFT) is often insufficient for high-precision applications, typically underestimating the fundamental gap. In this work, we present a new dataset of HOMO/LUMO energies for the QM9 compounds computed using the many-body GW method. This dataset, validated against the GW100 benchmark, provides a more physically accurate foundation for molecular property prediction, delta-learning, and transfer learning, particularly for larger molecules where GW is the most accurate but still numerically feasible method.
Introduction
Predicting a molecule's frontier orbital energies—the Highest Occupied Molecular Orbital (HOMO) and Lowest Unoccupied Molecular Orbital (LUMO)—is essential for designing organic electronics, batteries, and catalysts. These values determine the ionization potential and electron affinity, effectively dictating how a molecule interacts with electrons.
While the QM9 dataset is widely used to train machine learning (ML) models for these properties, its reliance on the B3LYP functional introduces systematic errors. DFT notoriously suffers from the "gap problem," often underestimating the HOMO-LUMO gap by several electronvolts. Consequently, models trained on this data may learn fundamentally flawed physics. The GW method offers a significantly more accurate alternative but has historically been too computationally demanding for large-scale datasets.
Methodology
To bridge this gap, researchers applied the GW method to the entire QM9 collection. The GW approach treats electrons as quasiparticles interacting within a screened field, providing a superior description of electronic excitations compared to standard DFT.
The calculations were performed using the CP2K software package, utilizing the Gaussian Augmented Plane Wave (GAPW) method. The team produced both "one-shot" ($G_0W_0$) and eigenvalue-self-consistent (evGW) results. To ensure results were independent of the mathematical basis set size, they performed calculations at two different levels (aug-cc-DZVP and aug-cc-TZVP) and extrapolated to the infinite basis set limit. The PBE functional served as the starting point for the GW iterations.
Results and Discussion
The resulting dataset represents the largest collection of GW simulations to date, encompassing 133,885 molecules and requiring over 7.4 million CPU hours. Validation against the high-accuracy GW100 benchmark set showed a mean unsigned error of only 28 meV, confirming the high quality of the data.
Accuracy and Distribution Shift
The transition from DFT to GW results in a dramatic shift in the predicted electronic landscape. The mean HOMO energy shifted significantly from -5.79 eV in the starting DFT calculations to -9.91 eV in the final GW results—a correction of over 4 eV. This correction significantly widens the fundamental gap, bringing it closer to experimental values.
Non-Constant Errors and Orbital Reordering
A critical finding is that the discrepancy between DFT and GW is not a simple constant offset. As shown in the correlation plots, the error contains significant molecule-specific components. This variability underscores why simple corrections to DFT are insufficient and why a full many-body treatment is necessary to capture true molecular physics.
Furthermore, the LUMO energies proved to be even more sensitive to the level of theory than the HOMO energies. In certain cases, the GW correction was large enough to reorder the molecular orbitals, which has profound implications for predicting chemical reactivity and optical properties.
Conclusion
By demonstrating that GW calculations scale as $O(N_e^3)$, this work proves that high-accuracy many-body methods are now feasible for "big data" chemistry. While approximately 1% of the molecules presented convergence challenges requiring specialized algorithms, the resulting dataset provides a robust new benchmark. This data is expected to catalyze the development of more accurate ML models through delta-learning and transfer learning, ultimately enabling the rapid and reliable design of novel functional molecules.
Summary updated