Wednesday, June 24, 2020.
Dr. Rolf Westerteiger, Senior Software Engineer at Math2Market GmbH, in an interview with Steffen Schwichow, Math2Market GmbH
Edited by Dr. Barbara Planas, Rabea Rett, and Franziska Arnold, Math2Market GmbH
Last update 07.07.2020
What can GeoDict do in the field of AI-aided image processing and object identification to truly understand a material?
GeoDict represents the digitalization of the materials laboratory. 3D image data in the micro and nanometer range are more accurate and affordable thanks to achievements in imaging techniques that make possible to experience and improve materials in a new dimension.
GeoDict offers the possibility to visually display 3D image data, and also allows experiments to determine the properties of the material to be conducted digitally. These digital property predictions bring insights into previously hidden processes of the microstructure of materials. But not only properties of existing materials can be predicted - new and revolutionary materials can be developed digitally through the material design modules in GeoDict.
However, until now, a gap existed between material analysis and material development. The generation of a statistical model (digital twin), based on the analyses of the material, was essentially limited by information gathered through manual measurements, such as the curvature of a fiber. The use of artificial intelligence based on neural networks has brought the decisive breakthrough.
The FiberFind-AI module of GeoDict provides a bridge to generate a true statistical (digital) twin for fiber media using AI-based image recognition of fibers directly on 3D image data. With the AI approach, fibers can be recognized and analyzed as independent objects, and information on the fiber media becomes available automatically with one click - fiber thickness, fiber length and fiber curvature.
In GeoDict 2021, this innovation has been taken a step further and a general object identification based on user-defined neural networks is ready to use.
The digital twin
A digital twin mirrors its real counterpart precisely. Thus, predictions can be projected back onto the original by simulations on the digital twin. For example, the time period of a material stress can be scaled down in days or hours and the material behavior can be simulated. Weak points can thus be identified and eliminated in advance.
In GeoDict, a digital exact representation of the original is also called digital twin, that is obtained by importing 3D image data into GeoDict to form a 3D model. All analysis and simulation possibilities of GeoDict can be applied on this digital twin.
The statistical twin
The particularity of microstructures is the complexity in the acquisition of representative material samples. Usually, several samples need to be taken, analyzed, and evaluated to compensate for local irregularities. This results in information defining a statistical model that can be given form using the material design modules (so-called generators) in GeoDict and, then, visualized.
The statistical twin is created and, unlike the digital twin, it does not represent the data of the material, e.g. the 3D image data, but the information contained in it, e.g. fiber lengths, grain size distribution, porosity, and much more. Using this statistical model, countless statistical twins of different sizes and resolutions can be generated, and their properties can be predicted by simulation.
The digital prototype
The great advantage of the statistical twin is the underlying statistical model, which describes the material using information. This information can be modified and adapted, displayed through the material design modules in GeoDict and their properties can be predicted by simulation. In this way, countless digital prototypes can be obtained quickly and easily, and their properties determined - without production costs or logistic waiting times.
In which areas does GeoDict already support the user with Artificial Intelligence?
Reliable segmentation of objects and binding agents in CT scans
With classic thresholding methods, image structures in CT scans are recognized and segmented based on their different gray levels. Since objects, such as fibers or grains, have the same gray value as the surrounding binder in the CT scan, it was previously hardly possible to separate them from each other. The Artificial Intelligence in BinderFind uses a different method - it looks at the geometric shape of the objects and can reliably segment them from the binder. In the next step, the object types in the material structure are identified. Up to this point, the detection of binders has been designed for fiber media.
Unique identification of object types in CT scans
A material can also contain several object types in its structure, defined by different geometric properties. Objects are marked as one cohesive component in threshold value procedures if they touch each other in the material or lie on top of each other. In this case, object types cannot be detected automatically, and manual measurement of the objects is difficult or impossible.
Our FiberFind-AI module uses Artificial Intelligence to identify and analyze the different fiber types in fiber materials, such as in the gas diffusion layer (GDL) of a fuel cell. This provides valuable information, e.g. on the number of fibers, length distribution, diameter distribution, curvature, and orientation distribution in three-dimensional space. With this information, a realistic microstructure, the so-called statistical twin, is generated with the FiberGeo module. In GeoDict 2021, the information provided by the analysis of these statistical data (postprocessing) directly from FiberFind has been extended to more graphical representations and videos.
For granular materials, the approach in the GrainFind module is not based on AI. In this case, powerful algorithms allow to identify and analyze grain types based on their size, shape, or position in space with comparable precision. With this statistical information a realistic statistical twin is then created in GrainGeo..
Realistic simulation based on the digital / statistical twin
The precise segmentation of CT scans by AI is the prerequisite for an exact digital twin. This means that all experiments based on this digital twin will afterwards provide realistic predictions on the behavior of the material.
The analysis and Find modules of GeoDict provide the way from digital to statistical twin. Wit them, 3D image data is converted into statistical information, such as fiber distribution and grain sizes, using smart algorithms and neural networks.
The statistical twin is the starting point for the development of new materials. The underlying statistical models of the statistical twin are modified by changing the generation parameters in the material design modules. In the next step, the properties of the new material - the digital prototype - are predicted by simulation and evaluated by the user. For example, first, the ratio of thick and thin fibers in the statistical model is changed and the effect of this modification is analyzed. Afterwards, for fiber media, the flow resistance in filtration or its mechanical properties are simulated.
The information in the statistical model is modified and analyzed via script with the global Python interface in GeoDict - GeoPython. It automates simulation workflows, so that, countless digital prototypes are created and analyzed, and properties predicted, without having to manufacture them at high costs. The automation of simulation workflows with GeoPython saves considerable time and money in the development of a material.
Thus, the entire chain of tools for material development is directly at your fingertips, in a comprehensive and easy to use software.
What does GeoDict 2021 deliver in terms of Artificial Intelligence?
New: Segmentation through interactive labeling of 3D CT scans
The classical method, to gather training data for a neural network, is to obtain sufficient CT images in advance for the training structures, in which objects must be manually marked in each CT layer. For 3D data sets, this procedure is hardly feasible.
This interactive labeling has, so far, only been applied to CT data sets in 2D, often in medicine. With this method, one or more interesting objects are roughly marked in the layers of a CT data set, each in any color. The idea is that the algorithm in the software learns incrementally from this to recognize and segment objects more and more effectively.
With the ImportGeo-Vol-AI module in GeoDict 2021, the interactive labeling method can be applied to 3D CT data sets. Starting with manually marking different objects in color on a CT layer, the algorithm learns and offers suggestions for completing the manual marking, resulting from what it has learned so far. These suggestions are checked and corrected, if necessary, or additional objects are selected to achieve the desired result.
In this way, the algorithm is interactively trained to recognize the objects precisely in all CT layers and to apply this knowledge to new CT data sets in the future.
New: Training of own neural networks
Even if GeoDict already provides ready-to-use neural networks, there is certainly always the need for the user to train his/her own neural networks. This may be the case if special geometries or forms in the microstructure have to be recognized.
In GeoDict 2021, the module GeoDict-AI helps in this task:
1. Creating training structures
GeoDict provides unique generators for complex objects allowing to create training structures for all kinds of objects (which can be displayed with GeoDict) by the user himself/herself within a very short time. The object generator is configured by the user via a Python script so that the data match the real material or CT data. Based on this, training structures are generated in GeoDict.
The exceptional advantage of GeoDict is that the material information is preserved during this process and a difficult manual labeling is not necessary.
2. Training a network
he neural network is now trained on the generated training structures. In our experience, this entire process takes a maximum of 4 days on a computer with a consumer Nvidia GTX 1080.
The neural network can now be used to recognize objects in CT scans.
What are the benefits of these improvements?
Saving a lot of time and money because...
- the production of prototypes is reduced to a minimum
- the manual labeling / marking in the CT scans is omitted
- the digital training data can be reused
- reliable simulation results due to high quality output data
You might also be interested in this!
GeoDict calculates on huge 3D structures in a very short time.
One of our highlights - In a client project, a 3D image data set of 18000 x 4000 x 2000 voxels, 144 billion data cells, was evaluated with FiberFind-AI within 24 hours.
This extraordinary computing power of GeoDict is certainly unique in the field of digital material analysis.
We support our customers with our expertise.
GeoDict-AI offers unimagined possibilities, but also requires some experience in using GeoDict and Python programming. Customers, who do not have the necessary know-how in this field, have the possibility to access the experience of our experts within the scope of a project.
Besides the project work, our workshops and online seminars provide the opportunity to build up the necessary knowledge by means of practical examples from your application area. Especially in the AI area, there will certainly be much more to see in the next months.