Investigating the efficacy of machine learning tools in modeling the continuous stabilization and carbonization process and predicting carbon fiber properties

Abstract

Carbon fiber reinforced composites are finding increased usage in a wide variety of applications, ranging from aerospace to energy harvesting to sporting goods and more. Carbon fiber manufacturing is an extremely time- and cost-intensive process with more than 70 processing variables. To better understand the processing-structure-property relationships and accelerate the improvement in carbon fiber mechanical properties, the integration of machine learning (ML) and experimental techniques is critical. The goal of this work is to investigate the use of ML models to map precursor information and carbonization process parameters to mechanical properties -- specifically, tensile strength and tensile modulus -- of carbon fibers. The experimental data consisted of 600 distinct points with 31 features. Four ML models were investigated viz. support vector regression (SVR), multi-layered perceptron (MLP) neural network, gradient boosted regression trees (GBRT), and recurrent neural network (RNN). The results indicate that ML can be used to approximate the underlying function describing the effect of the manufacturing process parameters on the carbon fiber tensile properties, with the RNN model outperforming all other models under consideration (R2 scores of 0.85 and 0.67 for tensile strength and tensile modulus, respectively).