Molecular dissociation of hydrogen peroxide (HOOH) on a neural network ab initio potential surface with a new configuration sampling method involving gradient fitting.

The Journal of chemical physics

PubMedID: 19586096

Le HM, Huynh S, Raff LM. Molecular dissociation of hydrogen peroxide (HOOH) on a neural network ab initio potential surface with a new configuration sampling method involving gradient fitting. J Chem Phys. 2009;131(1):014107.
The O-O bond dissociation of HOOH is investigated on an analytic ab initio potential-energy surface obtained by fitting the energies of 25,608 configurations using neural network (NN) methods. The electronic structure calculations are executed using MP2 calculations with the 6-31G* basis set. A new data-sampling technique is introduced to collect HOOH configurations in the six-dimensional hyperspace. This method is based on a comparison of the NN-computed gradients at configuration points currently in the database with the target gradients. By requiring that the NN gradients closely fit the MP2 target gradients, both the potential and the gradients are more accurately fitted. The selection criteria also ensure a more uniform distribution of configuration points throughout the important regions of configuration space. Molecular dynamics (MD) trajectories are not involved in the sampling. The final NN fitting yields average absolute and root-mean-squared testing set errors of 0.0060 eV (0.58 kJ mol(-1)) and 0.0099 eV (0.96 kJ mol(-1)), respectively. The effectiveness of the support vector machine (SVM) method in fitting large ab initio databases for MD calculations is investigated by using this method to fit the same HOOH database. The SVM fitting quality is tested by comparison to the NN fit. It is found that the average absolute and root-mean-squared testing set errors for the SVM fit are significantly larger than those obtained using NN methods. The total number of parameters in the SVM fit is more than a factor of 11 times the number of parameters in the NN fit. The trajectory computation time using a single NN averages about 1.8 s per picosecond of trajectory time. This increases to 9.0 s per picosecond of trajectory time if a five-NN committee is employed. The corresponding SVM computational time is almost 24 s per picosecond of trajectory time. Consequently, we conclude that a SVM is not as effective in fitting large databases for MD calculations as previously proposed methods, and thus is not employed to conduct MD studies. We employ the five-member NN committee to perform MD calculations at five different internal energies from 3.4 to 4.2 eV, including zero point energy. The rate coefficients are obtained directly from the first-order decay plots. They vary from 0.117 to 0.324 ps(-1). A Rice-Ramsperger-Kassel plot is found to exhibit good linearity.