Accurate Prediction of Microstructure of Composites using Machine Learning

In this prospective work, a machine learning (ML) model based on multiple independent random forest models to predict the configuration of binary composite bars is developed. The input variables to the ML model are elastic wave signals collected at one end of the composite bar, while the targets of the ML model are binary vectors representing the configuration of the bars. This study results indicate: First, a short period of elastic wave propagated through a composite bar can collect and carry the detailed information of the entire bar; second, the patterns hidden in the collected signals can be detected, extracted, and used by the ML model; finally, the ML model can be well trained using a relatively small dataset pool (less than 0.1% of all possible samples), and make accurate predictions. For the 30 sections of bars used in this study, the average prediction accuracy for each section of this bar can reach 95% and even higher. This ML guided technique can be modified and used in different functionalities and applications such as composites characterization, structure health monitoring, limestone determination, and archaeological detection.