Applicability of Universal Neural Network Potential to Organic Polymer Materials

With the recent development of materials exploration aided by machine learning (so-called materials informatics), atomistic simulations have been playing an increasingly important role. Density functional theory (DFT) calculations are widely used for accurate atomistic simulations, which usually take a few months due to high computational costs. Molecular dynamics (MD) simulations using classical force fields are much faster than DFT simulations and have been applied for atomistic simulations on nano- to micro-scale. However, the accuracy depends on the parameters of the force field. Therefore, there is a trade-off between computational cost and accuracy between these conventional methods. There are some approaches to calculate materials speedily with accuracy like first-principles calculations, such as density functional tight binding method (DFTB) or reactive force field (ReaxFF), but they have strong parameter dependence and poor versatility because they aim to reproduce specific systems with high accuracy in many cases.<br/>Neural Network Potential (NNP) is a machine learning model that uses the results of first-principles calculations as a data set to infer the energy and force of molecular and crystal structures. While conventional NNPs can only be used in a limited number of material systems for each model because the corresponding structure depends on the training datasets, PreFerred Potential (PFP) was developed with the aim of becoming a universal NNP by efficiently collecting training data and constructing an architecture that reproduces the smooth interatomic potential energy surface. This enables fast calculation of any 72-element combination structure with accuracy equivalent to first-principles calculations, and has already been shown to be effective in material exploration in some areas such as batteries and catalysts. On the other hand, large condensed systems composed of organic molecules, such as liquid solutions and polymers, are conventionally analyzed by MD simulations using classical force fields. Due to its accuracy and universality, PFP can extend the application of atomistic simulation to inorganic/organic interfaces, chemical reactions, and degradation of these systems.<br/>In this presentation, we will show examples of calculations to demonstrate that PFP can be useful not only for inorganic materials but also for organic materials. As a benchmark of the organic systems, we have calculated the density of various organic liquids obtained using PFP. The results are in good agreement with those obtained by DFT. For more practical examples, the thermal decomposition of epoxy resin was analyzed by PFP-based MD simulations. The results were similar to those calculated by ReaxFF, showing that epoxy resins initially produce small hydrocarbon compounds upon fragment decomposition, and that as the reaction temperature increases, the amount of water and hydrogen produced increases through radical reactions. Other examples of calculations made possible by PFP will be discussed.