Physically Recurrent Neural Networks for Computational Homogenization of Composite Materials with Microscale Debonding

TL;DR

This paper extends Physically Recurrent Neural Networks to model microscale debonding in composite materials, improving computational efficiency and capturing path-dependent damage behavior for multiscale homogenization.

Contribution

It introduces a novel PRNN architecture incorporating cohesive zone models to simulate microscale debonding effects in composites.

Findings

Enhanced PRNN architecture captures debonding effects.

Different network configurations improve prediction accuracy.

Method reduces computational costs of multiscale modeling.

Abstract

The growing use of composite materials in engineering applications has accelerated the demand for computational methods to accurately predict their complex behavior. Multiscale modeling based on computational homogenization is a potentially powerful approach for this purpose, but its widespread adoption is prevented by its excessive computational costs. A popular approach to address this computational bottleneck is using surrogate models, which have been used to successfully predict a wide range of constitutive behaviors. However, applications involving microscale damage and fracture remain largely unexplored. This work aims to extend a recent surrogate modeling approach, the Physically Recurrent Neural Network (PRNN), to include the effect of debonding at the fiber-matrix interface while capturing path-dependent behavior. The core idea of the PRNN is to implement the exact material models from the micromodel into one of the layers of the network. In this work, additional material points with a cohesive zone model are integrated within the network, along with the bulk points associated to the fibers and/or matrix. The limitations of the existing architecture are discussed and taken into account for the development of novel architectures that better represent the stress homogenization procedure. In the proposed layout, the history variables of cohesive points act as extra latent features that help determine the local strains of bulk points. Different architectures are evaluated starting with small training datasets. To maximize the predictive accuracy and extrapolation capabilities of the network, various configurations of bulk and cohesive points are explored, along with different training dataset types and sizes.