Neural network predictions of energy transfer in macromolecules

A neural network is trained to predict the dynamical behavior of internal energy in macromolecules. Energy flow out of localized sites within a polyethylene chain is studied as a function of time, excitation level, and temperature. Computations were examined for CH stretch excitation levels between v = 2 and 14, temperatures between T = 20 and 350 K, and time from O to 200 ps. The results show that the neural network predictions are accurate to within 8% error of those calculated from molecular dynamics. This difference is related to statistical fluctuations in the molecular dynamics calculations which result from small ensemble averages. In fact, the neural network acts to "filter out" those fluctuations, providing results which are more concordant with the limit of a very large ensemble average. In addition to this desirable property, the trained network can predict energy flow behavior for any reasonable excitation level, temperature (energies must be bound), and time scale, thereby significantly extending the computational limits of the molecular dynamics method.