Exploring Underlying Correlations in Multi-Fidelity Finite Element Simulations of Open-Hole Tension Tests: A Comparative Study Using Machine Learning

Abstract

Certification of composite aerostructures is typically achieved via analysis supported by testing, following the building block approach. Analysis of failure and crash at different scales, as well as design iterations and optimization, require fast and validated numerical models. Given the complexity and interactions of multi-failure mechanisms in composites, high-fidelity FE models are often needed to explicitly simulate the behavior of individual layers and their interlaminar interfaces, using a mix of continuum and discrete damage models. However, high-fidelity FE models require a considerable amount of time to set up, calibrate, and perform. On the other hand, low-fidelity modeling approaches, such as laminate-based smeared models, may provide reduced accuracy in some cases, while offering shorter simulation times and thereby lower computational costs. In this study, progressive failure of Open-Hole Tension (OHT) tests on HEXCEL IM-7/8552 quasi-isotropic laminates are investigated using both low- and high-fidelity FE models. A Machine Learning (ML) technique is then employed to perform a comparative study between the inputs and outputs of these models. The ultimate goal is to identify areas in the FE parameter space where the low-fidelity model can be used as a substitute for the high-fidelity model with reasonable accuracy and to discover the relationships among main parameters to reproduce high-fidelity results from the low-fidelity model in an accelerated manner. To generate the low-fidelity model, a laminate-based FE model was created in LS-DYNA, using a continuum elastoplastic damage-based material card, MAT081. The model parameters were varied randomly, and simulations were performed to generate a large database of low-fidelity FE results. Similarly, a ply-based FE model employing MAT261 for continuum intralaminar damage modeling combined with cohesive tiebreak contacts to capture discrete interlaminar delamination was created. This high-fidelity model was also used to generate another database of simulation results. Both databases were used to create a training set for an ML model, aiming to study the correlation between low- and high-fidelity inputs that result in similar failure responses. The results enable increasing use of low-fidelity FE models to generate data for the training of ML models.