Advancements in Visualization, Post-Processing and AI-Enhanced Modeling of Multi-Scale Micromechanics for Composite Materials

João Guilherme Brites Rei

Key information

Authors:

João Guilherme Brites Rei (João Rei)

Supervisors:

Peter Gustafson; Pedro Miguel de Almeida Areias (Pedro Areias)

Published in

December 10, 2025

Abstract

Composite materials have revolutionized the aerospace industry because of their enhanced mechanical properties and high strength-to-weight ratios. However, their heterogeneity and anisotropy pose significant challenges for predictive modeling, motivating the use of multi-scale micromechanics to connect microstructure characteristics and macroscopic response. Against this backdrop, this study addresses primary deficits of a localization/homogenization workflow by focusing on the visualization and enhancement of models based on the Generalized Method of Cells (GMC) and the High Fidelity Generalized Method of Cells HFGMC), with emphasis on their implementation within NASA’s Multi-scale Analysis Tool (NASMAT) ecosystem. In terms of visualization, this thesis details the development of two dynamic, ParaView-based plugins that convert large hierarchical simulation outputs into interactive insights. The \textit{NASMATDiscreteElement} systematizes Repeating Unit Cell (RUC) creation with transparent parametrization in a reproducible GUI, while \textit{NASMATExplorerRUC} enables inspection of deeply nested datasets through accurate cell picking and synchronizes comparisons between parent and child views. Both plugins contribute to reducing the persistent usability gap, extending beyond user experience and interpretability. Additionally, a shallow U-Net neural network is optimized on HFGMC data to enhance the accuracy of field predictions for Polymer Matrix Composites (PMC) using GMC, considering its limitations in shear-normal coupling. The Machine Learning (ML) model achieved a low validation error (MSE $\approx$5.11$\times10^{-5}$) and improved shear components, with gains of $\sim$52% (strain) and $\sim$170% (stress) on linear regression slopes ($R^2 \geq 0.965$). The average simulation run-time remains lightweight (GMC+ML $\approx$0.1526 s) relative to high-order baselines (HFGMC $\approx$0.19099 s).