Towards machine learning accelerated modelling of orthotropic composite materials​

Licentiate thesis, 2026

Laminated unidirectional composites and random short fibre composites have been homogenised with analytical methods for engineering design for decades. However there are no general analytical methods for woven (particularly 3D-reinforced) composites. There has been extensive work on the prediction of the elastic properties of 3D-reinforced textile composites through process modelling, but lately it has become popular to infer the as-manufactured geometry from X-ray computed tomography (XRCT). To derive the meso scale geometry from a XRCT volume the individual material phases must be segmented, which is a challenging problem. In this work the gap in the literature is being addressed by investigating machine learning (ML) based segmentation methods.

The first paper addresses the challenge of XRCT image segmentation, and more precisely the need for annotated data for training of ML based segmentation models. A fully automated pipeline for the generation of synthetic, automatically labelled, training data for ML based segmentation models of XRCT scans of 3D-textile reinforced composites is presented. A segmentation model trained on a dataset utilising the proposed pipeline shows 88% pixel-wise agreement when compared with a hand-segmented XRCT scan.

The second paper expands on the pipeline presented in the first paper with the aim to perform elastic homogenisation of a 3D-textile reinforced composite sample. The segmentation model is upgraded to a 3D U-Net architecture. The pipeline is also expanded to create voxel meshes from the segmentations including periodic boundary condition application. A novel material mapping routine is implemented that allows the mapping of the local yarn orientation and fibre volume fraction to the voxel elements, even when it is not fully possible to separate individual yarns. Combining the ML segmentation model with the material mapping routine, finite element models are generated which can be used to predict the statistical variation of elastic properties through computational homogenisation. An application of the pipeline results in accurately predicted homogenised elastic stiffnesses, with a deviation from experiments of less than 6.5%.