Quantitative micro-structural characterisation of hybrid polymer composites with convolutional neural networks
Topic(s) : Special Sessions
Co-authors :
Ji JI DONG (UNITED KINGDOM), Ali KANDEMIR (UNITED KINGDOM), Ian HAMERTON (UNITED KINGDOM)Abstract :
With properties like high specific strength and tailored structural properties enabled by the lay-up sequences, polymer composites have gained popularity in both industry and academia in recent decades. To compensate for the shortcomings of traditional composites, such as poor vibration damping, high carbon footprints, and production cost [1], the use of fibre hybridisation or incorporating natural fibres alongside carbon or glass in composite materials is becoming increasingly popular.
Common micro-structural characterisation for polymer composites generally involves destructive tests such as thermogravimetric analysis (TGA) or less destructive/non-destructive tests like microscopy combined with traditional image processing. However, TGA has poor compatibility with plant-based natural fibres, due to the high temperatures involved [2], and it suffers several limitations: it yields only a fibre volume fraction; quantitative determination is not achievable; it does not offer informative visualisation. While traditional image processing can address some of these issues, characterising micro-structure of hybrid composites is constrained by these labour-intensive processes, where only selected image sections are analysed [3]. By contrast, deep learning algorithms can automate hierarchical feature extractions, eliminating excessive manual intervention while maintaining high accuracy. For instance, Badran et al. [4] demonstrated the use of U-Net [5] to segment constituents and characterise the micro-structure of ceramic composites from micro-CT (Computed Tomography) slices.
Limited research exists on the utilisation of U-Net image segmentation in hybrid polymer composites and this study focuses on the micro-structural characterisation of hybrid polymer composites with deep learning techniques. The U-Net model is adapted and trained on optical images of hybrid composite cross-sections. This approach automates the quantitative and efficient estimation of micro-structural properties such as volume fraction, relative volume ratio between fibres, and void or fibre lumen content. In this study, the U-Net model in the initial trials showed promising results, achieving an accuracy of 0.98, precision of 0.8, and a dice score of 0.78 on the validation set for 5 classes (two fibres, matrix, void, and potting resin/background) segmentation. Successfully characterising hybrid composite micro-structures from optical images would quantitatively automate characterisation, which would be otherwise challenging with traditional image processing. This deep learning approach provides valuable insights e.g. estimating the relative fibre ratio, a key factor influencing pseudo-ductility [6]. The model’s training on optical images offers foundational understanding despite relying on 2D spatial information. Moreover, this work could lay a crucial foundation for future U-Net training on hybrid composite CT slices, promising a comprehensive 3D spatial quantification.
Common micro-structural characterisation for polymer composites generally involves destructive tests such as thermogravimetric analysis (TGA) or less destructive/non-destructive tests like microscopy combined with traditional image processing. However, TGA has poor compatibility with plant-based natural fibres, due to the high temperatures involved [2], and it suffers several limitations: it yields only a fibre volume fraction; quantitative determination is not achievable; it does not offer informative visualisation. While traditional image processing can address some of these issues, characterising micro-structure of hybrid composites is constrained by these labour-intensive processes, where only selected image sections are analysed [3]. By contrast, deep learning algorithms can automate hierarchical feature extractions, eliminating excessive manual intervention while maintaining high accuracy. For instance, Badran et al. [4] demonstrated the use of U-Net [5] to segment constituents and characterise the micro-structure of ceramic composites from micro-CT (Computed Tomography) slices.
Limited research exists on the utilisation of U-Net image segmentation in hybrid polymer composites and this study focuses on the micro-structural characterisation of hybrid polymer composites with deep learning techniques. The U-Net model is adapted and trained on optical images of hybrid composite cross-sections. This approach automates the quantitative and efficient estimation of micro-structural properties such as volume fraction, relative volume ratio between fibres, and void or fibre lumen content. In this study, the U-Net model in the initial trials showed promising results, achieving an accuracy of 0.98, precision of 0.8, and a dice score of 0.78 on the validation set for 5 classes (two fibres, matrix, void, and potting resin/background) segmentation. Successfully characterising hybrid composite micro-structures from optical images would quantitatively automate characterisation, which would be otherwise challenging with traditional image processing. This deep learning approach provides valuable insights e.g. estimating the relative fibre ratio, a key factor influencing pseudo-ductility [6]. The model’s training on optical images offers foundational understanding despite relying on 2D spatial information. Moreover, this work could lay a crucial foundation for future U-Net training on hybrid composite CT slices, promising a comprehensive 3D spatial quantification.