Machine learning-based damage characterization in composite plates subjected to low-velocity impact

Composite materials are employed in engineering due to their high strength-to-weight ratios and versatility. However, they are susceptible to low-velocity impacts, which present notable challenges. Traditional techniques for damage assessment, such as non-destructive testing and numerics-based methods, often face obstacles including high costs, material-specific limitations, and computational demands. To address these issues, we introduce an innovative machine learning-based framework designed to diagnose damage and anticipate structural responses in composite plates experiencing low-velocity impacts. This methodology employs a convolutional neural network to characterize damage, and a feed-forward neural network to predict the maximum deflection of composite plates under various impact scenarios. The framework was validated against experimental tests on composite plates made with different combinations of aramid and S2-glass fibers. Damage was accurately characterized following data augmentation and hyperparameter tuning, enabling precise predictions of damage presence, extent and position. Similarly, the structural response was satisfactorily predicted, with an average prediction error of 2.95% and 1.89% over two stacking sequences not seen during training. This approach marks a significant progression in composite material diagnostics and performance prediction by offering rapid predictions, thus bridging the existing gap between experimental constraints and computational efficiency.