Neural Network Modelling for Composite Damage Pattern Generation

This research project was initiated as a result of a curiosity and desire to investigate the applicability of surrogate modelling to analyse complex non-linear behaviour in aircraft structures. The study chose to focus on modelling damage in composite plates, and through a literature review, deemed that the generation of graphical outputs was a domain worthy of attention. Thus, the research questions subsequently formulated were centred around the modelling of artificial neural networks for the generation of damage patterns on composite plates. Data for training and evaluating these neural networks was first generated through 20 finite element models solved using Abaqus 2017. Standard neural networks trained to directly reproduce these damage patterns, as well as reduced-image neural networks trained to reproduce a reduced formof these patterns (obtained using convolutional neural networks) were analysed, and both were found in want ofmore training data. The generation of a further 421 finite element models resulted in a striking improvement in the performance of both networks, but with the standard neural network outperforming the reduce-image neural network. Thereafter, it was discovered that a hybrid network that combined facets of the standard and convolutional neural networks performed superior to both. In the process of training these networks, it was recognised that while the performance metrics served as an indicator of the resemblance between the predicted and actual outputs in terms of colours and contours, the same trends did not apply to the image quality. In order to improve the visual quality of outputs from the hybrid network, the use of the Structural Similarity Index (SSIM) was explored. It was eventually determined that pre-training the network using the Mean Square Error (MSE) as its optimising metric before then doing the final training using SSIM resulted in a model with impressive results. This fine-tuned model carried out predictions with a mean error of 0.0014 on theMSE metric and 0.9804 on the SSIM metric when evaluated on an independent dataset. Finally, the reliability and computational efficiency of the hybrid model was measured. It was found that approximately 95% of the MSE values on the independent dataset were within a value of 0.0040, while the same percentage of SSIM values were over 0.9100. The computation speed, meanwhile, improved by a factor of roughly 34 times on average, with the figure rising to over 443 on specific models.