Machine learning-based real-time velocity prediction of projectile penetration to carbon/aramid hybrid fiber laminates

Abstract

Composite laminates subjected to dynamic impacts are usually investigated by experimental or numerical techniques. Numerical simulations, as an excellent complementary tool to experiments, are capable of reproducing microscopic results that cannot be observed in experiments, but require time-consuming calculations. Therefore, this work demonstrates the capability and efficiency of neural networks and decision tree models for the real-time prediction of projectile penetration of aramid/carbon hybrid fiber laminates impacted at variable angles for different initial velocities. To obtain accurate prediction models, a combination of experimental and finite element methods has been adopted and the experimentally validated finite element models have been used to provide data for training the prediction models. Consequently, the prediction model is able to accurately predict the residual velocity after projectile penetration of unknown hybrid laminates. The research demonstrates that using a large dataset generated by finite element analysis can help the prediction model to give more accurate predictions. The decision tree model outperforms the neural network model in known datasets, but the neural network model has better generalization capabilities of handling unknown feature inputs and giving accurate results.