Optimized Boosting machine learning approaches for tensile strength prediction and failure classification of mono, bi and tri fiber reinforced composites

Fiber reinforced composites (FRCs) are important to aerospace, automotive and marine industries because of their good strength-to-weight ratio and flexibility in design. However, their highly reliable design is defined by a complex interaction of numerous processing parameters which makes challenge for predicting mechanical properties. Conventional destructive testing is expensive and time consuming and existing predictive models are often limited to single fiber systems or a limited set of the many properties of interest, failing to cover the holistic system performance of hybrid composites. To overcome these limitations, in this work a robust machine learning framework is developed for simultaneous prediction of tensile strength and primary failure modes classification of the mono, bi, and tri fiber reinforced epoxy laminates. Experimental results found carbon-dominated 8 cross ply tri-fiber hybrid laminate (CGK) as the balance condition with a maximum tensile strength 389.79 MPa while delamination is found as main failure. Consequently, five boosting models like AdaBoost, Gradient Boosting, XGBoost, CatBoost and LightGBM were trained, validated, optimized and tested on tensile experimental data. SHAP analysis showed that tensile strength is dominated by design parameters such as lay-up configuration and actual thickness while the failure mode is more sensitive to as manufactured features such as the lay-up type and lay-up configuration. After tuning hyper-parameters, CatBoost proved to be best for regression task having R 2 of 0.74, MSE of 1648.90 MPa, MAE of 29.65 MPa indicating good agreement with experimental observations while LightGBM proved to be sound classifier having Accuracy, Recall or sensitivity and F1-score of 1.0 for multi-class categorical failures. The proposed model have competitive and strong performance compared to the existing works on fiber reinforced composite, which usually claim high accuracy in single fiber systems and do not consider the more complex mono, bi and tri fiber hybrid systems. This work presents a broad, twin objective reliable, physically consistent predictive tool for capturing highly complex structure-property relations in hybrid composites, making competent structure design optimization and experimental testing less dependent on generous experimental data.