Machine learning based inverse framework for predicting the transverse and shear modulus of carbon fiber

Evaluating the transverse and shear modulus of carbon fibers experimentally is challenging and are usually obtained by an inverse approach using a micromechanical model. This paper presents an inverse approach for predicting the fiber properties from experimentally evaluated properties of the unidirectional (UD) lamina, using a machine learning (ML) based surrogate model. The ML framework is based on Gaussian process regression (GPR), which also provides a measure of uncertainty in the predictions. The ML model is trained using synthetic data generated using Finite Element (FE) homogenization considering different fiber and matrix properties, volume fractions, and fiber distribution. The proposed inverse framework is demonstrated to predict the elastic properties of polyacrylonitrile (PAN)-based fibers used in carbon-epoxy composites that have significant variations (low to high modulus) in its properties (T300-M60J). The fiber properties predicted using the GPR surrogate shows good agreement with the values reported in the literature, with a difference of less than 1.5%. The computational framework is further extended to predict the elastic properties of the woven fabric laminate using multi-scale homogenization. In summary, the results show that the GPR-based surrogate model offers an accurate and computationally efficient alternative to FE-based forward model for predicting the properties of the fiber.