Physics-aware deep learning models for predicting the heterogeneous mechanical properties of polymeric nanostructured materials

Abstract

Developing innovative materials with superior properties has been a major engineering challenge. Rationally designed polymer nanocomposites represent an emerging field that offers materials with enhanced mechanical properties and added functionalities. Here, we propose a computational methodology for predicting the heterogeneous mechanical behavior of model polymer nanocomposites that combines deep learning and detailed atomistic simulations. Atomistic simulations, while crucial for unraveling the mechanical behavior of composite materials, often focus on global mechanical properties by applying macroscopic strain and calculating the overall stress response. However, exploring the distributions of mechanical properties in heterogeneous polymer systems requires the computation of stress and strain fields for each atom within the simulation box. Our approach is based on a hierarchical data-driven computational framework that involves a nano/micro/macro coupling approach to predict the mechanical properties of polymer nanocomposites. We introduce a physics-aware deep learning method to predict the distribution of mechanical properties of model nanocomposites by directly computing stress and strain at the atomic level. Incorporating a set of physics-based constraints in the loss function encourages the deep learning model to follow the physical symmetries of the underlying system and to accurately predict the values of local stress and strain data in any arbitrarily chosen domain of the model systems. The proposed framework is transferable over systems with different volume fractions of nanofiller while being computationally efficient and effectively agnostic to the underlying microscopic model.