Flexural and hardness performance prediction of hybrid composites via machine learning and experimental validation

Abstract

This study has investigated the flexural strength and hardness of a hybrid composite reinforced with jute, kenaf, and glass fibers and filled with aluminium oxide (Al2O3) nanoparticles. The mechanical performance of the composites with (1) different fiber orientations (0°, 45°, and 90°), (2) different stacking sequences (1–3 layers), and (3) different Al2O3 contents (3–5 wt%) has been systematically evaluated. Response surface methodology (RSM) has been utilized to determine parameter interactions and statistical significance and to optimize the experimental conditions. Artificial neural networks (ANN), generative adversarial networks (GAN), and feed-forward neural networks (FFNN) have been employed for predictive modelling. Among these approaches, FFNN has demonstrated the highest prediction accuracy, closely matching the experimental results and outperforming GAN and RSM. The optimal combination of parameters has been identified as a 90° fiber orientation, 5 wt% Al2O3 content, and the stacking sequence G/J/J/K/K/K/K/J/J/G, which has enhanced the flexural strength by 18% and hardness by 32%. The effective predictive capability of FFNN has been confirmed by the lowest prediction errors, with a Mean Absolute Error (MAE) of 15.05 and a mean squared error (MSE) of 47.65, demonstrating that FFNN has provided superior predictive performance.