Neural network approach to the study of dynamics and structure of molecular systems

Neural networks are used to study intramolecular energy flow in molecular systems (tetratomics to macromolecules), developing new techniques for efficient analysis of data obtained from molecular-dynamics and quantum-mechanics calculations. Neural networks can map phase space points to intramolecular vibrational energies along a classical trajectory (example of complicated coordinate transformation), producing reasonably accurate values for any region of the multidimensional phase space of a tetratomic molecule. Neural network energy flow predictions are found to significantly enhance the molecular-dynamics method to longer time-scales and extensive averaging of trajectories for macromolecular systems. Pattern recognition abilities of neural networks can be used to discern phase space features. Neural networks can also expand model calculations by interpolation of costly quantum mechanical ab initio data, used to develop semiempirical potential energy functions.