Uncertainty Quantification in Multiscale Modeling of Polymer Composite Materials Using Physically Recurrent Neural Networks

This study investigates whether Physically Recurrent Neural Networks (PRNNs), a recent surrogate model for heterogeneous materials, trained on a micromodel with fixed material parameters, can maintain accuracy for varying material properties without retraining, and propagate uncertainty in a multiscale framework. Unlike conventional RNNs, where parameter changes require training or explicit inclusion of material properties as extra input features, PRNNs embeds material models in their material layer that allow for modification of material parameters after training. When adjusting material properties dynamically according to the input during testing, PRNN shows high accuracy across a wide range of parameters. Therefore the surrogate can be applied to multiscale uncertainty quantification (UQ). Compared to the full-order simulations on an overly coarse mesh, the PRNN-driven model reduces simulation time by over 7000 times while accurately capturing highly nonlinear evolution of the probability density for the macroscopic response as a result of a given distribution for microscale material parameters. A PRNN-driven UQ is demonstrated on a more accurate finer mesh that would be computationally infeasible with the full-order model.