A deep learning approach to predict crashworthiness behaviour of mechanical meta-material

  • Belinda Babu Joseph

Student thesis: Doctoral ThesisDoctor of Philosophy

Abstract

With “The 2030 Agenda for Sustainable Development” aiming to minimise road traffic accidents by halve, due to 1.3 million people dying every year in road accidents, the need for improvement in vehicle crashworthiness is required. One feasible way to improve crashworthiness is by designing mechanical meta-material structures for the design of car bumpers. However, conducting finite element analysis for a crash scenario is time consuming and expensive due to the complexity of the model. Therefore, the aim of this project is to investigate whether a deep learning neural network could be used to predict crashworthiness behaviour of mechanical meta-material structures. Artificial intelligence is capable of making smart, informed decisions in a short span of time, thereby, reducing computational cost compared to current computational methods.

A novel computational framework was proposed to develop a user-defined auto-generated algorithm capable of modelling mechanical meta-material structures, consisting of thin-walled hexagonal cells and solid hexagonal cells, that undergoes crush simulation and output crashworthiness parameters. The auto-generated algorithm was created using Python and ABAQUS®/Explicit. Multi-objective simulated annealing optimisation was also used to control the tessellation of the design space of the mechanical meta-material. A Pareto front containing all the optimum mechanical meta-material structures for increased specific energy absorption and peak crush force was obtained. The solutions on the Pareto front was used as the data pool for training the deep learning neural networks designed. Levenberg-Marquardt Backpropagation (LMB) neural network and densely connected convolutional neural network (DenseNet) were used to predict the crashworthiness behaviour of optimum mechanical meta-material structures. LMB and DenseNet were also used to predict optimum meta-material structure when peak crush force was given as input for the network. It was found that LMB neural network performed better for predicting crashworthiness parameters, and predicting optimum mechanical meta-material structure for improved crashworthiness.

Date of AwardJul 2023
Original languageEnglish
Awarding Institution
  • Queen's University Belfast
SponsorsNorthern Ireland Department for the Economy
SupervisorGiuseppe Catalanotti (Supervisor), Zafer Kazancı (Supervisor) & Brian Falzon (Supervisor)

Keywords

  • Deep learning
  • neural networks
  • mechanical meta-materials
  • crashworthiness
  • multi-objective optimisation

Cite this

'