Developing Neural Networks to Represent Anisotropic Molecular Interactions

Efficiency of electricity generation, for instance in solar cells, is determined by the structure of organic molecules in the solar cell material. To determine such defining characteristics molecular dynamic computer simulations are performed, but only to run on simplified models of the molecular structure in order to conserve computational time.

With these simulations we are now applying machine learning (ML) models, specifically artificial neural networks, to encode the molecular interactions between anisotropic rigid bodies. This way polymer and macromolecular systems can be predicted, while lowering computational cost with minimal loss of structural accuracy for these equilibrium systems.

We then test the network structure of these machine learning models and the training datasets they are learning off of. Doing so in order to demonstrate the challenges that arise when moving from spherically-symmetric systems to those requiring orientation specific torque calculations.