Identification of fibers with Artificial Intelligence (AI)

FiberFind-AI

The FiberFind module is the next step towards precise object recognition in µCT images, combining the power of fast parameter prediction with the speed of neural networks for accurate fiber and binder detection. FiberFind embodies the work in GeoDict that aims to understand 3D scans of fibrous materials, like nonwovens and fibrous composites.

3D models are obtained after importing and segmenting computer tomography or FIB/SEM scans of a material, and three distinct approaches are followed in FiberFind:

  • Identification of individual fibers by classical image processing methods 
  • FiberFind-AI and BinderFind-AI: Identification of individual fibers and binder by Artificial Intelligence (AI) approaches 
  • Analysis of statistical properties of fibers: Diameter distribution, orientation distribution, curvature distribution

Examples of applications

  • Modelling: In material modelling, together with the FiberGeo module, to create structure models matching a physical sample. 
  • Analysis of binder content: Identify and separate binder from fibers in 3D scans and study the effect of varying binder content on the performance of a fibrous material (e.g., nonwoven) or granular material. 
  • Material optimization: Reduction of experimental costs by simulating and determining the predicted effects of changes in material parameters before manufacturing material prototypes. 
  • Material quality control: Study of inhomogeneities and deviations from target values of diameter, orientation, and curvature of fibers during production.

FiberFind Features

Fiber orientation distribution

The orientation distribution computes an orientation tensor characterizing the orientation of fibers. This analysis can be performed globally, over the whole sample, or for individual sub-regions in the model. The latter functionality can be used e.g. to analyze each layer of a layered material separately or to study heterogeneity in fiber orientation across the sample volume. As explained before for fiber diameter, the orientation tensor can be entered directly into the FiberGeo module to reproduce structures with those orientation distributions. Orientations can also be computed per voxel and stored as a 3D orientation field. This field can be loaded into the ElastoDict or ConductoDict module to consider transverse isotropic material behavior (different material constants along vs. across the fiber).

Curvature estimation

The curvature estimation produces a histogram of fiber curvatures by extracting individual fibers of the µCT image.

 

Fiber diameter estimation

The fiber diameter estimation computes the average fiber diameter for a chosen number of different fiber types, as well as its standard deviation which can be sufficient for unimodal distributions. More detailed results are provided in the form of a diameter histogram plotting the fiber diameter vs. the volume fraction of fibers of that diameter. Fiber diameter distributions (discrete or continuous) can then be entered in the FiberGeo module to reproduce structure models with matching distributions.

 

BinderFind-AI and FiberFind-AI

BinderFind-AI and FiberFind-AI are based on neural networks trained to differentiate between fibers and binder and to identify separated fibers. The unique structure generation capabilities of GeoDict (FiberGeo module) provide the ground truth data to train the neural networks.

 

Examles of Applications

FiberFind in the Material Development Process

By classical image processing methods, pores are separated from solids. The solids consist of individual fibers and (possibly) some binder which usually have the same gray values in the images, so that gray value-based identification is not possible. FiberFind makes it possible to differentiate binder from fibers (BinderFind-AI) and, then, to identify and separate the individual fibers (FiberFind-AI) automatically through trained Neural Networks or by classical image processing approaches.

The results of FiberFind separation and identification can be used as input to reproduce isotropic and anisotropic fibrous structures using the FiberGeo module. By being used together, FiberFind and FiberGeo are intended to close the digital fibrous material design loop and to help in reducing the number of laboratory experiments needed.

Subsequently, the statistical parameters of the modelled structure may be easily varied to investigate the effect of material structure on the performance of materials by using one of the property prediction modules (-Dict) in GeoDict. For example, the fiber orientation can be computed for each material voxel, producing an orientation field. This makes possible to simulate materials with transverse isotropic properties in studies with the ElastoDict (mechanics) and ConductoDict (thermal conductivity) modules.

FiberFind is particularly well suited for the analysis in fibrous structures made of long non-hollow fibers with circular cross-section.

GeoDict Applications

Additional modules needed?

  • The GeoDict Base package is needed for basic functionality. 
  • Some installation steps are necessary to set up FiberFind-AI and BinderFind-AI for GPUs and CPUs. 
  • The ImportGeo-Vol module is needed to import and segment µCT images and create the structure models for analysis. 
  • The FiberGeo module can be used to model structures that (statistically) match the analyzed µCT images. 
  • The fiber orientation field computed by FiberFind can be used by other modules, such as ElastoDict and ConductoDict, to carry out simulations of deformation and damage or thermal and electrical conductivity for transverse isotropic materials.