Benchmark comparison of dual-basis double-hybrid density functional theory and a neural-network-optimized method for intermolecular interactions

• Published in: Journal of molecular spectroscopy Vol. 376; p. 111406
• Main Authors: Lutz, Jesse J., Byrd, Jason N., Montgomery Jr, John A.
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.02.2021
• Subjects: ACONF; Double hybrid density functional theory; Dual basis SCF; Neural networks; RGC10; S22; RGC10; S22; Neural networks; Double hybrid density functional theory; ACONF; Dual basis SCF
• DOI: 10.1016/j.jms.2020.111406
• ISSN: 0022-2852

[Display omitted] •Double-hybrid density functional theory (DH-DFT) is accurate but expensive for larger systems.•Here an efficient dual-basis DH-DFT is derived and implemented for general use in GAMESS.•Differences between dual-basis and conventional DH-DFT energies are negligible.•Benchmark comparisons reveal dual-basis DH-DFT offers comparable accuracy to state-of-the-art neural-network-based approaches at a reduced computational expense. We present a computationally efficient implementation of double-hybrid density functional theory (DH-DFT) leveraging the dual basis methods of Head-Gordon and co-workers and the resolution-of-the-identity second-order Møller-Plesset (RI-MP2) theory. The B2PLYP, B2GP-PLYP, DSD-BLYP and DSD-PBEP86 density functionals are applied to assess the performance of dual-basis methods on several benchmark test cases, including the CONF set of conformational energy differences in C4-C7 alkanes, the S22 set of noncovalent interaction energies, and the RGC10 noble-gas dimer dissociation curves. The dual-basis DH-DFT approach is shown to give results in excellent agreement with conventional methods at a reduced computational cost. For noncovalent interaction energies, DH-DFT is compared against a leading neural-network-based approach, namely the SNS-MP2 method of McGibbon and coworkers (McGibbon et al., 2017). The DH-DFT and SNS-MP2 methods are shown to produce similar accuracies when compared to the established benchmark values.