Artificial Intelligence and Machine Learning

Artificial Intelligence - particularly deep learning techniques - has proven to be highly effective in image segmentation tasks. This technology assists in transforming 3D gray-value images, obtained from FIB-SEM and micro-CT, into digital representations of materials. Within GeoDict, this capability is implemented in the ImportGeo-Vol module. Essentially, AI serves as a powerful tool in accurately defining the material phase associated with each voxel, contributing to the precise characterization of complex structures within the images.

Machine Learning is used also to enhance image features and to improve the image quality of 3D gray-value images in the ImportGeo-Vol module of GeoDict. This capability finds application in scenarios where high-quality results are desired, but are constrained by factors like short scan times or initially low-quality images. Additionally, super-resolution techniques are used to enhance image detail and clarity beyond the inherent resolution limitations.

Distinguishing between the fiber material and the pore space is possible when analyzing segmented 3D images featuring a fibrous or nonwoven material . However, directly assigning a solid voxel to a specific fiber presents challenges for conventional algorithms, such as those based on watershed methods. In this scenario, the pre-trained neural networks in GeoDict offer precise single fiber identification and are the foundation for highly accurate computations of fiber diameter, orientation, curvature, and length. The single fiber identification based on machine learning is found in the FiberFind module.

In a gray-value image, it is a challenge to distinguish between fibers and binders in a gas diffusion layer (GDL) of a fuel cell or between active material and binder in a battery electrode microstructure. GeoDict comes equipped with pre-trained neural networks designed to accurately identify the binder phase within 3D images. This enables the creation of a precise digital material representation and facilitates the incorporation of distinct properties of individual constituent materials. Consequently, this approach allows for accurate microstructure-based simulations. These trained neural networks are available in the FiberFind and GrainFind modules of GeoDict.

GeoDict comes equipped with powerful pre-trained Neural Networks designed for specific applications.

With GeoDict-AI, we offer an user-friendly solution for our customers to train their own Neural Networks. The networks are finely tuned to tackle diverse challenges in the analysis of 3D images. For example, to identify contact areas between individual grains in materials such as rocks or even battery electrodes.

Additionally, the capabilities of GeoDict in microstructure generation serve as an ideal solution to create virtually limitless amounts of training data for machine learning and, in this way, advance digital material development.