A machine learning case study with limited data for prediction of carbon fiber mechanical properties
School of Engineering
Predicting success before full scale manufacturing is the cornerstone of every high-tech industry. Many factors should be considered when designing a product. Design of Experiments (DOE) has been a major, conventional method for such purpose. Moreover, nowadays, Machine Learning (ML) provides fast processing, real-time predictions with a continuous quality improvement for complex industrial design processes. ML with big data is known to handle multi-dimensional and multi-variate data in dynamic industrial systems. However, the ability of the ML methods against limited training datasets has been addressed scarcely.
In the present study, a Taguchi DOE approach combined with two selected common ML approaches, the Support Vector Regression (SVR) and Artificial Neural Network (ANN), have been assessed against producing a robust prediction tool for mechanical properties of carbon fibers with a limited training dataset. To this end, empirical models based on three rendering parameters of the stabilization reaction process of oxidized Polyacrylonitrile (PAN) fibers (OPF) have been created. These rendering parameters were obtained according to the explanation of whole Fourier transform infrared attenuated total reflectance (FT-IR ATR) spectra. ML revealed a very promising agreement between the trained empirical models and experimental data. Namely, in terms of the fibers Young’s modulus, the results showed that the SVR empirical model with the average error of less than ±2.4% had a better performance than ANN-LMA (Levenberg –Marquardt algorithm) with the average error of less than ±2.7%. However, in terms of the fibers tensile strength, it was concluded that ANN-LMA empirical model with the average error of less than ±3.7% had slightly surpassed the SVR with the average error of less than ±4.1% in this case study.