Transfer Learning to CCSD(T): Accurate Anharmonic Frequencies from Machine Learning Models

• DOI: 10.1021/acs.jctc.1c00249
• Publication Date: 2021-05-08T13:26:11Z
• Authors: Silvan Käser; Eric D. Boittier; Meenu Upadhyay; Markus Meuwly
• Keywords: Transfer of learning

The calculation of the anharmonic modes small- to medium-sized molecules for assigning experimentally measured frequencies corresponding type molecular motions is computationally challenging at sufficiently high levels quantum chemical theory. Here, a practical and affordable way calculate coupled-cluster quality using second-order vibrational perturbation theory (VPT2) from machine-learned models presented. approach, referenced as "NN + VPT2", uses high-dimensional neural network (PhysNet) learn potential energy surfaces (PESs) different which harmonic VPT2 can be efficiently determined. NN approach applied eight (H2CO, trans-HONO, HCOOH, CH3OH, CH3CHO, CH3NO2, CH3COOH, CH3CONH2) are reported NN-learned MP2/aug-cc-pVTZ, CCSD(T)/aug-cc-pVTZ, CCSD(T)-F12/aug-cc-pVTZ-F12 For largest highest theory, transfer learning (TL) used determine necessary full-dimensional, near-equilibrium PESs. Overall, yields within 20 cm–1 determined close 90% PES available 10 more than 60% modes. MP2 PESs only ∼60% were experiment, with outliers up ∼150 cm–1, compared experiment. It also demonstrated that allows provide correct assignments strongly interacting such OH bending torsional in formic acid monomer CO-stretch OH-bend mode acetic acid.