A high-dimensional neural network potential for protonated water clusters, 2014
In the final semester of my iMOS studies, Dr. Jörg Behler from the theoretical chemistry department accepted to supervise my master thesis. Dr. Behler’s lab specializes in modelling reactive potential energy surfaces (PES) based on neural networks for describing molecular, bulk, and interfacial systems. Such a potential will provide maximum accuracy and efficiency in molecular dynamics (MD) simulations when fitted to the high level quantum chemical data. My project was to develop one such potential for protonated water clusters, which are important model systems in studying proton transfer mechanisms in general. I started with sampling the configurational space of the protoned water clusters from monomer (with one water molecule and a proton) to octamer (eight water molecule with a proton) using stochastic search. Energy and forces acting on these sampled molecular structures were computed with density functional theory (DFT) and it is to these data the neural networks are fitted.
MD simulations were carried out with the preliminary potential in order to sample additional conformations missed in the earlier stochastic sampling. These structures were added to the data set and the potential is refitted in order to improve reliability of the potential. Once a reliable potential was obtained that provided negligible error in the predicted energies and forces compared to the DFT values, it was ready for further analysis and MD simulations. The completed potential was then used to find the minimum energy structures, harmonic frequencies, proton transfer mechanisms and transition pathways between different minima of protonated water clusters as detailed in my thesis.
Link to Master course in Molecular Sciences and Simulation (iMOS) at Ruhr-University Bochum