AI prediction of the architecture of 3d textile fabrics

  • Koptelov, Anatoly (BCI, University of Bristol)
  • El Said, Bassam (BCI, University of Bristol)
  • Thompson, Adam (BCI, University of Bristol)
  • Hallett, Stephen (BCI, University of Bristol)

Please login to view abstract download link

Weaving is one of the most common techniques for manufacturing of textile preforms. To reduce the number of labour-intensive manufacturing trials, it is important to be able to predict the mechanical properties of the woven fabric based on its initial geometry and architecture. The conventional approach to fabric compaction simulation is extremely time consuming. It could take days and weeks of computational time to produce a reliable deformed model. To tackle that problem, the kinematic multi-filament software suite was developed in Bristol Composites Institute [1], [2]. It proved to provide a significant advantage in predicting 2D and 3D fabric geometries over the conventional finite element approach. The resulting compacted architecture contains the information about yarns’ features in three-dimensional space, such as fibre volume fraction and fibre orientation. This work employs an AI approach to drive the capabilities of textile modelling even further. The idea of the proposed system is to train a deep learning model to mimic the fabric’s deformation process and to predict the distribution of the important features of a compacted textile unit cell based on its initial architecture. Four thousand case studies of various fabric weave architectures were generated to provide a sufficient training dataset for the AI model. Each case study contained time-distributed evolution of a 3D textile unit cell throughout weaving process. The analysis of the resulting spatiotemporal data was carried out by a combination of convolutional and recurrent neural networks. Such approach facilitates the extraction of relevant features from the deformed yarns geometry at every step of a weaving process. The resulting trained model was able to predict three-dimensional feature distribution of a random textile unit cell at different stages of compaction process based on the provided initial architecture only as shown in Figure 1. Such approach allows to significantly reduce the computational time required for the modelling as the prediction for each solution step is nearly instantaneous. Rapid weaving simulation lays the ground for design and optimisation of textile preforms, as such problems require to solve many different woven architectures.