Sequential prediction of transverse crack paths in CFRP composites via a deep learning approach

Abstract

Transverse cracking is one of the earliest damage modes in carbon fiber-reinforced polymer (CFRP) composites under transverse or complex loading, leading to stiffness degradation and reduced fatigue life. Its initiation and early propagation occur at the microscale and are strongly influenced by random fiber distributions and local streconcentrations. However, transverse crack propagation cannot be fully characterized within a single microstructural window, owing to the combined effects of the local fiber arrangement and crack state inheritance from the preceding region. To address this challenge, a physics guided deep learning framework is developed for sequential prediction of transverse crack paths in CFRP composites. Specifically, the framework directly predicts crack paths from microstructural images and incorporates the predicted crack segment as the initial crack information for the subsequent prediction window, thereby enabling sequential prediction across neighboring regions. An attention enhanced encoder decoder network is employed to capture multiscale features of crack evolution, while physical guidance derived from phase-field fracture analysis is incorporated to improve the consistency of the predicted paths. Compared with phase-field model results, the proposed framework accurately reconstructs transverse crack paths in extended microstructural regions with a prediction accuracy of 93.8%. This study provides an efficient surrogate model for sequential transverse crack path prediction and multiscale damage analysis of CFRP composites.