Tomography-based modelling of fiber material with the help of machine learning

Important results from the project

The purpose was to automate the image processing of the fiber networks investigated in a beamline together with BillerudKorsnäs. The manual segmentation was significantly improved and a process description was created. However, it is likely that further speed and accuracy improvements will be required before manual segmentation is a realistic alternative for segmentation of relevant volume sizes. The automatic segmentation method works as intended, but the data generated is not sufficiently accurate data to be directly applicable.

Expected long term effects

The project helped identify appropriate methods for segmenting natural fibers and will inform future efforts to correctly segment volumes with many fibers. Together with machine learning experts work is ongoing to develop methods that give machine learning algorithms the best possible starting point. The lessons learned during the project have contributed to two scientific publications and a software library.

Approach and implementation

Two methods to generate training data were investigated. The method development can be used by stakeholders in academia or industry to manually segment natural fiber networks. The work continues in cooperation with DESY in Hamburg. For machine learning to work in the segmentation task, more work is needed to improve all steps of the process. Parts of the results are presented in two articles under consideration by the journals Experimental Mechanics and Cellulose.