written by Prof. Ian Hamerton
I am delighted to report a publication success from one of our recent Aerospace Engineering graduates, Mr Ji Dong, who conducted his final year research project with me (and ably co-supervised by Dr Ali Kandemir) following a summer internship within BCI.
The mechanical performance of discontinuous fibre composites is heavily dependent on the quality of fibre alignment and the ability to gauge this parameter rapidly and reliably (and thus the potential properties of the resulting composite) is key to successful application of the technology, particularly in a production environment.
Ji’s study investigated the application of deep learning-based image segmentation using 2D optical imaging for the microstructural characterisation of composite materials with hybridised fibres, potentially offering a cost-effective and more rapid alternative to computed tomography/3Dimaging. Laminates were produced using the HiPerDiF method, combining discontinuous high modulus carbon and basalt fibres to reinforce a poly(L-lactic acid) (PLA) matrix. Ji found that the Generalised Dice Loss function significantly outperformed others, particularly when discriminating voids, achieving a 19% improvement in Dice Similarity Score on an unseen dataset for full image characterisation. Additionally, volume fraction, relative fibre and void ratios, and fibre alignment computed from the segmentation results closely matched his ground truth data.
Figure: Image of a cross section of hybrid composites containing discontinuous carbo and basalt fibres. The x-axis and y-axis units are expressed in numbers of pixels (a) Image in RGB (b) Greyscale (c) Greyscale image after filtering and CLAHE (d) Ground truth (e) Class segmentation results based on the traditional Multithresh method, where the Class4 and Class5 represent the noise being classified as a category rather than the desired region of interest (f) to (h) are the segmentation results obtained from U-Net model trained under CE, GDL and Comp1 respectively, where it can be seen that the degree of misclassification is significantly less compared to the multithresh method .
A manuscript produced from his research dissertation “Microstructural characterisation of fibre-hybrid polymer composites using U-Net on optical images” has just been accepted for publication in a forthcoming special issue of Composites Part A (Machine Learning and AI in Composite Science and Manufacturing), to be edited by Assist. Prof. Navid Zobeiry and Assoc. Prof. Marco Salviato (both of the University of Washington, Seattle). The quality of the work is apparent when I note that the manuscript was received at the editorial office on 13th September, rigorously reviewed, revised (with additional work conducted by Ji) and returned by 20th October, and accepted on 30th October.
Ji is currently pursuing an MSc in Engineering Mathematics and is seeking to secure a PhD primary focus on opportunities within bioengineering/biomedical engineering and AI (sadly beyond my research interest areas). He is an exceptional researcher and I’m sure that this will be the first of many such papers in his career.