AI in the future of materials research

GeoDict-AI

AI and machine learning are central components of GeoDict. Researchers and engineers work more efficiently and accurately by using AI and machine learning, ultimately accelerating the introduction of innovative materials to the market. In GeoDict, this technology helps our customers to convert 3D grayscale images - from sources such as FIB-SEM and micro-CT - into digital material representations. Additionally, GeoDict is equipped with powerful, pre-trained neural networks designed for specific applications, e.g. for

  • Identification of individual fibers or grains,
  • Separation of material phases, or
  • Enhancement of a grayscale image.

With GeoDict-AI, our customers have access to a user-friendly solution to effortlessly train their proprietary, custom neural networks. These neural networks are tailored specifically to almost any task involved in the analysis of 3D images. A concrete example of application is the detection of contact areas between grains in materials such as rocks or even battery electrodes.

Typically, a large amount of sample data are required to train a neural network. However, manually labeling or annotating materials or objects in 3D scans is extremely tedious and time-consuming.

Facilitating this task, GeoDict-AI enables training a neural network using artificially generated structure models. Here, GeoDict's ability to generate 3D structures provides the ideal solution to create practically unlimited training data for machine learning applications in digital materials development.

Powerful structure generators such as FiberGeo, for fibrous structures with binder, and GrainGeo, for granular structures with binder, are used for this purpose. GeoDict-AI handles the training with practically any structure generated with the GeoDict structure generators.

After training, the trained neural network is used to analyze 3D scans of real materials. This approach saves time and resources, since the network automatically extracts the necessary information from the scans without requiring additional manual work.

Application areas of GeoDict-AI

The potential applications for AI in materials research are nearly limitless. Here are some examples of networks trained with GeoDict-AI that we use.

Identification of binders

It might be difficult to distinguish between fibers/grains and their binders in a grayscale image of a fuel cell's gas diffusion layer or in the microstructure of a battery electrode. Neural networks may be trained with GeoDict-AI specifically to accurately identify the binder phase within 3D images. This results in an accurate digital representation of the material, later enabling precise simulations at the microstructure level.

Identification of individual fibers or grains (labeling)

When analyzing a 3D grayscale image of a nonwoven fabric or a fiber-reinforced composite, it is possible to distinguish between the fiber material and the pore space. However, for classical algorithms like the Watershed algorithm, assigning a single voxel directly to a specific fiber is challenging. In this case, the pre-trained neural networks in GeoDict enable precise identification of individual fibers, which is necessary for accurate calculations of fiber diameter, orientation, curvature, and length.

Individual grains are also identifiable with customized neural networks. For example, in 3D images of a battery electrode.

Separation of material phases (segmentation)

Artificial Intelligence, particularly deep learning techniques, has proven to be extremely effective in image segmentation. Within GeoDict, this technology helps our customers convert 3D grayscale images from sources such as FIB-SEM and micro-CT into digital material representations. Essentially, AI serves as a powerful tool to accurately define the material phase assigned to each voxel in the 3D model, thus contributing to the precise characterization of complex structures within the images.

Enhancement of grayscale images

Machine learning is also used within GeoDict to enhance image features and the image quality of 3D grayscale images. This capability is applicable in scenarios where high-quality results are desired but limiting factors exist, such as short scan times or initially lower-quality images.

Following Modules are often used in combination with GeoDict-AI:

Image Processing and Image Analysis ImportGeo-Vol            
Material Analysis GrainFind FiberFind PoroDict + MatDict        
Modeling & Design FiberGeo FoamGeo GrainGeo PaperGeo WeaveGeo GridGeo PleatGeo
Simulation & Prediction BatteryDict ConductoDict SatuDict ElastoDict FlowDict    

Suitable modules depend on the concrete application.