GeoDict 2024 ships with pre-trained neural network models for binder and fiber identification which operate directly on segmented micro-CT images. These networks produce good results for a wide range of fibrous and granular materials. For materials which are too different from the original training data, e.g. fibers with a very large diameter or non-circular cross-sections, however, they might not produce acceptable results.
In this breakout session, we will demonstrate a workflow based on GeoDict-AI which enables you to train your own custom neural network which is optimized to analyze the materials relevant to your specific application in a matter of days. Key to this approach are the advanced structure generation capabilities of GeoDict, which enable us to produce synthetic datasets suitable for training AI models.
As a case study, we will begin by generating a digital statistical twin of a nonwoven material. We then record this generation process as a GeoPy macro and introduce parameters which allow us to vary material properties such as fiber diameter, porosity or fiber orientation.
In the next step, we use GeoDict-AI to generate a Design of Experiments which samples uniformly from a given range of values to produce a parameter table which is representative of the variation seen across real material samples.
We then use “Create Training Data” to automatically generate a set of training structures corresponding to this parameter table. These structures are then used as a training data set to train an optimized custom neural network model to identify and isolate individual fibers within these structures. Finally, we demonstrate how this model performs on real materials by applying it to a micro-CT scan of a nonwoven.

The video is part of the GeoDict Innovation Conference 2024 and the simulations were run with the GeoDict 2024 release.
—————————————————————————————————————————-
For more information, contact us directly at support @math2market.de.
Please also visit our website at https://www.math2market.com.
Find us on LinkedIn: https://www.linkedin.com/company/math2market-gmbh
Math2Market GmbH 2024

source