Optimizing compression after impact prediction in CFRP using machine learning combined with classical composite laminate theory, SHAP analysis
TL;DR
This research investigates Gradient Boosting and Neural Networks for predicting impact damage characteristics and compression-after-impact strength in CFRPs and confirms the models’ physical interpretability, with feature importance rankings aligning with established composite laminate theory.
Abstract
The aerospace industry increasingly relies on modeling to accelerate composite structure development. While finite element methods (FEM) effectively simulate impact damage in carbon fiber-reinforced polymers (CFRPs), they demand specialized expertise and substantial computational resources. Machine Learning (ML) offers an efficient complement to traditional simulations while maintaining physical interpretability. This research investigates Gradient Boosting and Neural Networks for predicting impact damage characteristics (delamination area, indentation depth, residual indentation, and perforation) and compression-after-impact strength in CFRPs. A comprehensive dataset of 500 samples was developed, incorporating mechanical parameters derived from classical laminate theory. Results demonstrate that Gradient Boosting achieves superior accuracy for well-defined experimental measurements, while Neural Networks better handle outputs with greater experimental dispersion. SHapley Additive exPlanations (SHAP) analysis confirms the models’ physical interpretability, with feature importance rankings aligning with established composite laminate theory, thereby validating their potential for damage tolerance assessment in aerospace applications.