A neural network approach to the study of internal energy flow in molecular systems

Neural networks are used to develop a new technique for efficient analysis of data obtained from molecular-dynamics calculations and is applied to the study of mode energy flow in molecular systems. The methodology is based on teaching an appropriate neural network the relationship between phase-space points along a classical trajectory and mode energies for stretch, bend, and torsion vibrations. Results are discussed for reactive and nonreactive classical trajectories of hydrogen peroxide (H₂O₂) on a semiempirical potential-energy surface. The neural-network approach is shown to produce reasonably accurate values for the mode energies, with average errors between 1% and 12%, and is applicable to any region within the 24-dimensional phase space of H₂O₂. In addition, the generic knowledge learned by the neural network allows calculations to be made for other molecular systems. Results are discussed for a series of tetratomic molecules: H ₂X₂, X = C, N, O, Si, S, or Se, and preliminary results are given for energy flow predictions in macromolecules.