Rapid neural network prediction of linear copolymer melt free energies

Rapid computation of polymer free energies has a wide variety of applications in polymer science, particularly in self-assembly, where free energy changes are critical in determining stability and equilibration. Unfortunately, all available free energy calculation methods require integration along the variable of interest, which limits their applicability. Even the Bennett acceptance ratio (BAR) method, which is not as sensitive as other methods, exhibits large errors when the two compared systems are drastically different, such as when two polymer systems have large Flory-Huggins parameter differences. Here, we explore utilizing neural networks (NNs) to predict free energies of linear block copolymers, utilizing dissipative particle dynamics (DPD) simulations to provide relevant training parameters, such as mean per-bead energies, and utilizing the BAR method to provide free energy output values. Furthermore, using gradient tracking, we are able to determine the most relevant parameters towards NN free energy prediction, and extend these analyses to random copolymers.