Physics-Informed Machine Learning for Damage Localization and Progression in Composite Structures

Abstract

Composite materials are widely used in lightweight structures, but their heterogeneous architecture and interacting damage mechanisms make damage progression difficult to predict. As a result, composite components are often designed with conservative safety margins, increasing weight, cost, and inspection effort. Improved methods for structural health monitoring and damage-progression analysis are therefore needed to support earlier damage detection, residual-life assessment, and more efficient structural design.This research develops physics-informed machine-learning methods for monitoring and modeling damage evolution in composites. We propose a framework that addresses three complementary challenges: (1) damage detection and localization from sparse strain measurements using graph neural networks that generalize to unseen damage locations and sensor layouts, (2) microscale stress prediction in damaged composite microstructures using a deep-learning surrogate that can be integrated into micro-macro simulation workflows, and (3) macroscale damage modeling using a topologically informed neural network. Together, these contributions form a multiscale learning framework linking sparse experimental observations, microscale material response, and macroscale laminate damage evolution.