TITLE:
A New Method to Predict the Mechanical Behavior for a Family of Composite Materials
AUTHORS:
Viorel Mînzu, Iulian Arama
KEYWORDS:
Composite Materials Modeling, Regression Models, Machine Learning, Simulation
JOURNAL NAME:
Journal of Materials Science and Chemical Engineering,
Vol.13 No.9,
September
23,
2025
ABSTRACT: This paper introduces a method to develop a common model based on machine learning (ML) that predicts the mechanical behavior of a family with three composite materials. The latter are structures composed of the same carbon fiber and different matrices. The developed algorithms predict the stress given the strain value and eight physical parameters of the matrix. Creating a prediction algorithm (PA) for each composite alongside its training dataset is the straightforward solution. First, the paper develops three separate regression models that model the composite family together. As a main contribution, a model conversion is achieved, turning the set of individual regression models into a single model representing the entire family. This model conversion is advantageous because the training dataset of the family model is much smaller than the combined datasets of the three models. This reduction is possible because the family has a common “feature” that defines the family: the same Carbon Fiber. According to our knowledge, this research topic has not been addressed in the existing literature. An iterative procedure develops regression models for the entire family. For the composites family, the models are trained and tested using data sets generated by a high-performance simulation software. Complex design sessions in MATLAB, which have used Multiple Linear Regression, Support Vector Machines, Decision Trees, Regression Neural Networks, and Gaussian Process Regression, have identified effective PAs. The selection is based on the accuracy of the prediction, mainly indicated by Root Mean Square Error. A family regression model runs 60 - 70 times faster than the simulation software. This helps it work more efficiently and be integrated more easily into an optimization or design program.