Membrane Characterisation using Machine Learning

The synthesis and evaluataion of synthesized membrane structures as virtual parts of digital twins provide a data basis for training machine learning algorithms and for enabling them to predict structure-property correlations when applied to real structures. The research data infrastructure Kadi4Mat [1] developed at KIT offers a platform for combining both, data management and data processing as a prerequisite for structured data storage as a database for machine learning.
The present work introduces a workflow that models the development process of a machine learning algorithm for the determination of the geometric anisotropy in porous polymer membranes and is implemented in Kadi4Mat in an automatable way. In the current use case, the structures to be investigated are diagnostic membranes used in lateral flow tests.
As a basis for the machine learning approach, more than 12,500 structures with a specifically configured microstructure are generated algorithmically. These synthesized microstructures are designed to mimic the macroscopic properties of the real diagnostic membranes and are additionally imprinted with different geometric anisotropies.
... mehr
The generated structures are prepared for the CIDS [2] machine learning framework, an integrated part of Kadi4Mat. A supervised learning algorithm is selected and trained with the prepared structures aiming to predict anisotropic physical properties of the specific membrane structure. We show that the resulting anisotropy measurement tool can be applied to the digitally reconstructed images of the real diagnostic membranes.


In order to save the process knowledge and keep it accessible for similar applications, a digital workflow holistically models the development process of the ML tool from the generation and preprocessing of the data to the training of the ML architecture. The KadiStudio workflow editor, an integral part of the research data infrastructure, allows the creation, visualisation and recording of the workflow.

The focus of this contribution lies on the implementation of such a reproducible and automatable workflow for the characterization of porous microstructures by means of machine learning. The relevance of a research data infrastructure for accelerating and archiving research efforts will also be addressed.