Crashworthiness Analysis of An Aircraft Subfloor Structure using Artificial Neural Networks

Abstract

The goal of this thesis is to build an artificial neural network(ANN) surrogate model, that predicts the crashworthiness performance of a structure. The structure used in this research is a thermoplastic fibre-reinforced composite aircraft subfloor section which is part of the SmarT multifUNctioNal and INteGrated thermoplastic fuselage (STUNNING) fuselage demonstrator section. The structure will be analysed numerically, by making use of the finite element software LS-Dyna by Ansys. Within this research, a surrogate model is defined as a model that has been fed data from a computationally expensive model, yet is computationally considerably more inexpensive to run to obtain similar results as the expensive model. The main methodology to do so in this research is to build a two-dimensional design space, which is sampled using optimal latin hypercube sampling. By using an existing model for the STUNNING fuselage subfloor section and by automatic generation of geometry changes of the structure through a Python script, input files can be created according to the sampled design space. The simulations are then run on a high performance cluster. Once the simulations are run the results can be fed to an artificial neural network that uses this data to predict crashworthiness performance of the structure within the design space.

To build towards the goal of the thesis, a small comparative study has been performed on material models MAT054, MAT058, and MAT 261 in LS-Dyna. These material models are based on different failure criteria, and thus have different performance in coupon simulations. Using these simulations as a basis, a simple design space was created, sampled, and modelled so that a first ANN surrogate model could be built that could predict the failure displacement and the ultimate load of a composite coupon in tension. It was found that for a small dataset (<10 samples), the neural network could easily fit to the data, and could predict accurately on samples within the design space that were withheld from the network during training.

Then, the STUNNING subfloor section model was introduced and a number of models that section this structure are investigated. The main purpose of this investigation is to find out if the model can be reduced in size without changing the resulting failure mode. When a satisfactory reduced model was found, the research carried on by doing a sensitivity analysis on a proposed two-dimensional design space, and similarly a sensitivity analysis on the parallelization settings. From this study, a second failure mode of the model was discovered and the effect of the parallelization settings was quantified. From here the limits of the design space could be formed and sampled. Finally, different variations of ANN surrogate models were defined, trained, and tested. In the end, a root mean square discrepancy of around 10\% was achieved during testing of the network. This means that the methodology could prove useful in design applications, depending on the design stage and needed accuracy.