Multiscale Thermal-Mechanical Analysis of Tow-Steered Composite Plate Structures Using a Mixed-Fidelity Neural Network Model

Abstract

One challenge of design optimization of tow-steered composites is the significantly increased computational cost with more design variables in realistic structures. Emerging machine learning models, such as artificial neural network (ANN) models, offer an efficient alternative to expensive finite element analysis (FEA). However, the requirement for extensive training data can limit the application of ANN models in structural design using tow-steered composites. In this paper, we will utilize a recently developed Design Tool for Advanced Tailorable Composites (DATC) to generate training data. Single-fidelity ANN models will be developed to take design parameters of tow-steered composites and predict the critical buckling load of structures. The accuracy and efficiency of the ANN models will be validated with high-fidelity FEA. To reduce the cost of generating training data, we propose a composite ANN model to take mixed-fidelity data from simulations at different mesh densities. Two examples of tow-steered composite structures will be employed to develop the single-fidelity and mixed-fidelity ANN models. The results show that the proposed mixed-fidelity ANN model delivers accuracy comparable to high-fidelity FEA, while also significantly enhancing efficiency in comparison to the single-fidelity ANN model.