Molecular dynamics studies on neural network ab initio potential energy surfaces.

摘要

The neural network (NN) method has been employed to construct analytic ab initio PES's for three different chemical reactions that contain four atoms. Two different sampling procedures are used to collect configurations in six-dimensional hyperspace, which are the novelty sampling and gradient sampling techniques. Both methods have been proved to be successful in molecular dynamics (MD) studies.Nitrous acid (HONO) has two reaction channels: cis-trans isomerization and N-O dissociation. 21,584 configurations are sampled using novelty sampling technique, and MP4(SDQ)/6-311G(d) calculations are performed for those configurations. A feed-forward NN fit is performed with 41 neurons in the hidden layers. The reported fitting error is 0.017 eV (1.64 kJ mol-1). MD investigations are conducted for both reactions on this surface, and large intra-mode coupling is observed.The reaction is also a four-body system with a large extension of six-dimensional hyperspace because of molecular collision. With 1,300 points provided from a previous study, it is convenient to sample more configurations. The ab initio potential energy calculations are performed at MP2/6-311G(d,p) level of theory for a final database of 19,208 points. The fitting error of a NN committee (of five NN's) is 0.0046 eV (0.44 kJ mol-1). Fitting gradients are tested, and excellent accuracy is obtained. MD is conducted on the NN surface, and gives a maximum reaction probability of 0.152 in the translational energy range of 0.415 eV to 0.829 eV. Reaction cross sections are also calculated with an impact parameter of 0.265 A with various translational energies from 0.415 eV to 0.829 eV.The last four-body molecular system is HOOH. We introduce a new sampling technique namely "gradient sampling." In this technique, configurations are obtained based on regional gradient analysis of a temporary surface in hyperspace. Data are obtained more uniformly, which helps to improve the fitting accuracy. With 25,608 points being sampled, ab initio calculations are executed using MP2 level with the 6-31G* basis set. A five-member NN committee is constructed (each NN has 34 neurons) and contributes an excellent fitting error of 0.0060 eV (0.58 kJ mol-1). The SVM fitting method is tested on this database, and gives higher fitting error. The SVM surface also costs more computational efforts to execute MD investigations. Therefore, it is not preferred to be used in MD studies. We finally execute the investigation of O-O dissociation on the NN surface at various internal energy levels. The reaction rate coefficients are found based on the first order reaction rate law, and obey the Rice-Ramsperger-Kassel theory. We conclude that three vibrational modes are not effective during the dissociation, and internal hydrogen bonding occurs, which strongly prevents the dissociation.