Machine Learning Models for Strength Prediction of Carbon Fiber Reinforced Polymers

Abstract

This paper applied and benchmarked machine learning regression algorithms for predicting the tensile
strength of carbon fiber-reinforced plastics. A comprehensive statistical analysis revealed significant
linear correlations between the input parameters (ratio of matrix and filler concentration, composite
density) and the target variable (tensile strength). The dataset consisted of 150 experimental data
points. However, nonlinear models significantly outperformed linear models in terms of prediction
accuracy. The highest accuracy on an independent test dataset was demonstrated by the neural
network model (R² = 0.996351), as well as the Gaussian process regression models (R² = 0.996246)
and the ensemble method (R² = 0.996096). Classical linear regression yielded a significantly lower
result (R2 = 0.89498), confirming that the overall dependence is a significant, nonlinear,
multidimensional function. The key result is the detection of nonlinear behavior in the
«polymer−fiber» system: the tensile strength reaches a maximum in the optimal reinforcement phase
(10–20 wt.%) and then decreases sharply due to fiber agglomeration and porosity. The success of
nonlinear ML models is based on their ability to effectively model these complex, hidden
dependencies, paving the way for accelerated materials design.