Reactive dynamics and spectroscopy of hydrogen transfer from neural network-based reactive potential energy surfaces

Abstract

The "in silico" exploration of chemical, physical and biological systems requires accurate and efficient energy functions to follow their nuclear dynamics at a molecular and atomistic level. Recently, machine learning tools gained a lot of attention in the field of molecular sciences and simulations and are increasingly used to investigate the dynamics of such systems. Among the various approaches, artificial neural networks (NNs) are one promising tool to learn a representation of potential energy surfaces. This is done by formulating the problem as a mapping from a set of atomic positions x and nuclear charges Z i to a potential energy V (x). Here, a fully-dimensional, reactive neural network representation for malonaldehyde (MA), acetoacetaldehyde (AAA) and acetylacetone (AcAc) is learned. It is used to run finite-temperature molecular dynamics simulations, and to determine the infrared spectra and the hydrogen transfer rates for the three molecules. The finite-temperature infrared spectrum for MA based on the NN learned on MP2 reference data provides a realistic representation of the lowfrequency modes and the H-transfer band whereas the CH vibrations are somewhat too high in frequency. For AAA it is demonstrated that the IR spectroscopy is sensitive to the position of the transferring hydrogen at either the OCH-or OCCH 3 end of the molecule. For the hydrogen transfer rates it is demonstrated that the O-O vibration (at ∼ 250 cm −1 ) is a gating mode and largely determines the rate at which the hydrogen is transferred between the donor and acceptor. Finally, possibilities to further improve such NN-based potential energy surfaces are explored. They include the transferability of an NN-learned energy function across chemical species (here methylation) and transfer learning from a lower level of reference data (MP2) to a higher level of theory (pair natural orbital-LCCSD(T)).

Full-dimensional potential energy surface for acetylacetone and tunneling splittings

“…In 2020, Meuwly and co-workers using a novel Neural Network approach in the PhysNet software 35 developed a full-dimensional PES for AcAc based on MP2/aVTZ energies and gradients. 1 This PES was used in molecular dynamics calculations of the infrared spectrum. Thus, in 19 years a full dimensional PES for 15-atom AcAc has gone from "impossible" to a reality.…”

“…In this paper we apply the fragmented PIP approach to obtain a new PES for AcAc. Although a NN-based PES for AcAc has recently been reported, 1 as noted above, we decided to augment the data set for that fit with additional MP2/aVTZ energies and gradients to obtain and investigate a new PES. The two approaches are very different and so it is worthwhile to report on this new PES, providing additional insight on accuracy and computational timing.…”

Breaking the Coupled Cluster Barrier for Machine-Learned Potentials of Large Molecules: The Case of 15-Atom Acetylacetone

“…Meuwly and co-workers applied TL using thousands of local CCSD(T) energies to improve their MP2-based neural network PESs for malonaldehyde, acetoacetaldehyde, and acetylacetone (AcAc). 14 We recently proposed and tested a Δ-learning approach that uses a small number of CCSD(T) energies to correct a PES based on DFT electronic energies and gradients. 22 The method was validated for PESs of small molecules, CH 4 and H 3 O + , and for 12-atom N -methylacetamide.…”

“…Finally, it is also worth commenting on the current Δ-ML approach and the recent application of transfer learning (TL) by Meuwly and co-workers 14 to MP2-based PESs for AcAc. The MP2-based PES we reported used a slightly extended database of MP2 energies and gradients from that group.…”

“…Several TL NN models, based on random training data sets (PNO-LCCSD(T) energies) of different sizes, were considered by Meuwly and co-workers. 14 In Table 3 of that paper, results both from a single TL NN model and from an average of several TL NN models were given at the optimized geometries of each model. A TL-NN model gave a barrier height of 0.92 kcal/mol using 100 high-level energies, 2.4 kcal/mol from a single TL-NN model, 1.80 kcal/mol averaged from several TL-NN models with 1000 energies, and 2.66 kcal/mol from a single TL-NN model using 5000 energies.…”