Determining damage initiation of carbon fiber reinforced polymer composites using machine learning

Abstract

Computational progressive failure analysis (PFA) is vital for the analysis of carbon fiber reinforced polymer (CFRP) composites. The damage initiation criterion is one of the essential components of a PFA code to determine the transition of a material’s state from pristine or microscopically damaged to macroscopically damaged. In this paper, data-driven models are developed to determine the matrix damage initiation with the objective to save computation time. For 2D plane stress states, the computational cost for determining damage initiation can be dramatically reduced by implementing a binary search (BS) algorithm and predictive machine learning models. Machine learning models are evaluated against a regression problem of predicting damage crack angle as well as against a classification problem for predicting damage initiation outright. It is found that regression models perform much better for plane stress and 3D stress states when generating failure envelopes. In both cases, a neural network is able to produce a failure envelope that is over 99% accurate while reducing computational cost by over 90%.