Practical Machine Learning Strategies. 2. Accurate Prediction of ωB97X-V/6-311+G(2df,2p), ωB97M-V/6-311+G(2df,2p) and ωB97M(2)/6-311+G(2df,2p) Energies From Neural Networks Trained From ωB97X-D/6-31G* Equilibrium Geometries and Energies
© 2025 Wiley Periodicals LLC.
| Veröffentlicht in: | Journal of computational chemistry. - 1984. - 46(2025), 13 vom: 15. Mai, Seite e70129 |
|---|---|
| 1. Verfasser: | |
| Weitere Verfasser: | , , , |
| Format: | Online-Aufsatz |
| Sprache: | English |
| Veröffentlicht: |
2025
|
| Zugriff auf das übergeordnete Werk: | Journal of computational chemistry |
| Schlagworte: | Journal Article DFT algorithms machine learning molecule structure neural networks |
| Zusammenfassung: | © 2025 Wiley Periodicals LLC. Starting from ωB97X-D/6-31G* geometries and energies, neural network models have been trained to reproduce energies from three density functionals: ωB97X-V, ωB97M-V, and ωB97M(2), using the 6-311+G(2df,2p) basis set. Training sets on the order of 300 k organic molecules were utilized [≈295,000 molecules for the ωB97X-V functional (up to molecular weight 400 amu); and ≈289,000 molecules for both the ωB97M-V and ωB97M(2) functionals (up to molecular weight 380 amu)]. All training and validation molecules comprise uncharged, closed shell singlets including H, C, N, O, F, S, Cl, and Br (only). The resulting models have been assessed using molecules outside the training sets. Total energies obtained from the neural network models rarely differ from the corresponding density functional values by more than 2-5 kJ/mol (RMS), while differences in conformer energies are typically less than 1 kJ/mol |
|---|---|
| Beschreibung: | Date Revised 14.05.2025 published: Print Citation Status PubMed-not-MEDLINE |
| ISSN: | 1096-987X |
| DOI: | 10.1002/jcc.70129 |