Research of feedforward neural network applicability in computer simulation of polymers

In this paper we investigate the adequacy of deep learning force field models for modeling amorphous bodies. A polymer with the studied physical properties, polyphenylene sulfide, was chosen as a test substance. The simulation results shows that the forces predicted by neural networks acting on polymer atoms are significantly different from the forces calculated by ab initio molecular dynamics methods. A qualitative comparison with the force field model of a simpler compound, black phosphorene, shows that feedforward neural networks are unsuitable for modeling complex amorphous substances.