Tuesday, April 21, 2020.
Steffen Schwichow interviews Andreas Grie▀er of Math2Market GmbH.
Edited by Dr. Barbara Planas, Rabea Rett, Franziska Arnold and Sina Schwichow, Math2Market GmbH
By founding Math2Market GmbH in 2011, we set ourselves the goal to turn our passion for mathematics and advancement in computer science into an easy-to-use and versatile simulation software for everyone. We are very proud of the resulting software, by the name of GeoDict. Our dedication to its constant improvement drives us to top performance. Luckily, we are exposed to the unusual mix of closeness to cutting-edge scientific research and development, and the commercial practicality of bringing a truly useful simulation software to our customers. We are grateful for the daily challenges and the inspiration from our customers and colleagues to make GeoDict an even better tool.
But far too rarely do we take the time to share the new features and our know-how with others. When we do it, we often hear "What, GeoDict can do that?!" Sharing and communicating the stunning possibilities of GeoDict is our motto for this year and for the upcoming GeoDict release 2021 - Do good and talk about it!
Therefore, let's start by asking and answering:
With the module ImportGeo-Vol, GeoDict offers a rich toolbox to map the complete image processing workflow for image data sets: from image processing up to segmentation and the resulting 3D microstructures.
With GeoDict, this workflow is conveniently executed in a single software and fully automated. In recent years, we have paid special attention to the improvement and further development of our own image filters in the field of image processing. These new filters are specific for the removal of artifacts in grayscale images.
For GeoDict 2021, we have focused on simplifying segmentation and incresing its quality. For image analysis, and moreover for material development, GeoDict offers a wide range of tools in the form of modules. For example, the PoroDict module is applied in the analysis of pore space, and combined with the MatDict module, for the analysis of the complete material structure. The FlowDict module is used to compute flow properties and ElastoDict is optimal to predict plastic deformations and, thus, damage. And this is only a small selection of GeoDict modules and simulation possibilities.
GeoDict is not bound to a certain hardware or device. GeoDict runs on all common Windows and Linux computer systems. Faster computer systems are especially advantageous for simulations. Because GeoDict is designed to work in parallel, more calculation units not only bring faster results, they also save time and costs. The limiting factor of the hardware is the main memory. To compute on very large microstructures, a corresponding amount of RAM is needed. So far, the largest raw structure that we have imported from a 3D image data set into GeoDict was about 18000 x 4000 x 2000 voxels. But there is no size limit for GeoDict.
GeoDict is also not bound to a device type or manufacturer for the image capture. For GeoDict, the output data format is important and not the device for image capture, be it ÁCT, CT, FIB-SEM or Synchrotron). GeoDict covers a wide range of formats and, of course, the most common image formats (raw, vol, rek, txm, am, vox, iass, tif(f), jpeg, png, and many more), no matter if in a batch or as single images.
A prototypical workflow does not exist, but slightly depends on the input 3D data sets. Are these already segmented or are they still available as raw data? It also plays a role if only the segmentation is done in GeoDict and, afterwards, work is continued with the segmentation results in another software. Assuming all simulation work should be done in GeoDict, the workflow of image processing, cleaning of the data, segmentation, and analysis of material properties, can be accomplished in this order in GeoDict.
GeoDict includes all tools needed to obtain the best results out of the 3D data set.
As mentioned at the beginning, the main memory is the limiting factor for the size of the data set. Large data sets require a lot of main memory. If this is given, GeoDict can work with it.
The process is demanding for materials with different components that cannot be separated in the image, mostly because of very similar density and absorption properties. In this case, an alternative image capturing technique and approach may be worthwhile. We often see this for composite materials, fiber materials containing a binding material, or electrode materials in which the carbon black is difficult to distinguish from the active material.
The different dimensions of the data sets may make things even more interesting. So far we have often mentioned 3D data sets, but it is also quite common to work with 2D data sets, e.g. micrographs. We are planning to present a preview of working with 4D data sets this year. Not only the processing, but especially the analysis of these data is very exciting.
In GeoDict 2021, the user interface has been vastly improved, especially regarding the user guidance. Filters and settings are comprehensibly sorted as a workflow and can be recorded, saved and reused as workflow macros.
This workflow starts in the user interface with the optimization of the images. For the material analysis later on, it is crucial how precisely the data set has been segmented. This means that the images have to be easily recognizable for GeoDict in order to provide the best results during segmentation. To save time, we recommend to start with the determination of the region-of-interest. During the image capture, there is always more taken than the region of interest. The reason for this is that it is important not to damage the capture area while preparing the sample. The uninteresting border areas can simply be cut off later, making the data set somewhat smaller and work of the image filters faster. The selection of the region-of-interest is manual work, because the machine cannot select the area of interest to the user. However, when all images are fairly similar, a workflow macro can be defined, saved, and applied to the data sets at any time.
With the image filters, the image noise is reduced or the edges are sharpened in the next step. This step helps to better assign the image areas during the segmentation and improves the quality. Also here, the right parameters for the image filters are found manually and saved as a workflow macro. Great breakthroughs are already included in GeoDict 2020, specially for acquisition artifacts, such as streak and ring artefacts which are difficult to avoid during the capturing process and disturb the quality enormously. For example, we have developed a so-called alignment filter especially for FIB-SEM images, because the technical conditions of the image capturing process can cause the individual images to slip against each other. The alignment filter corrects this slippage in the images.
Personal preferences for image filters and proprietary image filters are easily implemented by GeoDict users via the Python interface in GeoDict.
A great deal of fine tuning and adjustment is necessary, especially in segmentation, to avoid assigning too much or too little gray value to the materials. How can GeoDict help with this issue? And how do you make sure to have found the right parameters?
There are different methods, depending on the quality and on the complexity of the 3D data set.
If the user is already familiar with the samples and with the imaging device and knows from experience the gray value ranges at which the material is located, these limit values for the materials can be entered manually into GeoDict. Here, an automation would be worthwhile to save time, either via Python interface in GeoDict 2020 or directly via workflow macros, as a novelty in GeoDict 2021,.
However, if you are dealing with changing material samples or new imaging devices, we recommended using the segmentation filters.
In the best case, the Otsu-method alredy allows segmenting the 3D data set in a fully automated way, with one click. GeoDict calculates the best way to separate the materials or phases and applies it to the data set. The Otsu-method does not require any user input, which makes the result reliably reproducible.
With more complicated data sets, e.g. with several materials or phases in one data set, it becomes more sensitive. The Otsu-method can still be used here, but then the contrasts in the data set are usually no longer sufficient to reliably separate different materials. To better support the user, we have developed an AI (Artificial Intelligence) segmentation filter for GeoDict 2021.
In the first step, the different materials in the data set are manually labeled. The same material gets the same label. Thus, the AI learns which values represent which material and reliably segments the data set from then on. Marking is done conveniently in the ImportGeo-Vol module using a brush tool, by painting the materials with the corresponding label for the material directly on the data set. The user can freely choose the slices or images of the data set to work on. After all, it is quite possible that not all materials are available on a single image of the data set.
In the second step, a preview of the segmentation is displayed. Areas that are incorrectly recognized by the AI can be marked as training data to further improve the AI recognition. As soon as the result is satisfactory, the trained AI model is saved and ready to be applied to similar data sets.
But any segmentation is only good to the extent that it digitally represents the actual material. Since there is no uniform measurement value or key figures by which to measure the quality of a segmentation, it is complex to determine how representative a statistical Digital Twin is. GeoDict can help here when a specific characteristic value for the material is known, either statistically (e.g. the solid volume fraction in the sample is known from documentation) or experimentally determined (e.g. the pressure drop through the material or its conductivity). In GeoDict, these characteristics are determined later on the statistical digital twin and compared to the original material. If the deviations of these values are outside the tolerance range, it means that the segmentation needs to be adjusted. The deviation is used as an indication for segmentation improvements.
At this point, the Session Macro or GeoDict's command history is truly helpful. The Session Macro automatically records and saves all steps of a GeoDict session and can be edited and executed again as a Python script. This also allows to quickly test and validate a number of different import parameters.
At the end of the segmentation, there is a digital representation of the material on voxel basis in GeoDict, which we call statistical digital twin.
We know from customers that the workflow of image processing, segmentation, and validation described above is fully automatically integrated into their material development workflow.
What we also experienced was an instance when experiment and simulation diverged greatly. This difference could not be explained by normal measurement fluctuations or statistical model deviations. Nor had we received any such feedback from other customers. On the contrary, the simulation results are highly reliable. We tested GeoDict strenuously searching for a possible mistake in the simulation, and the customer took another close look at the experiment. It turned out that a defective component in the experiment setup was responsible for the error, and it was explicitly found through the simulation. In this case the simulation validated the experiment and not the other way around.
Nevertheless, we do not advise against laboratory experiments. Only when two independent sources are compared with each other, can realistic and reliable results be obtained. However, GeoDict is able to assist to evaluate whether a cost- and time-intensive experiment may pay off.
Here we have a wide range of information that you can get from your material through GeoDict. Starting with the geometric image analysis, which is based on the result of the segmentation, one can determine and compare properties like pore size distribution, fiber diameter, wall thickness, the connectivity of the pore network, and much more. These geometric analyses are uncomplicated and can be read out from the geometry of the material, for the most part directly or with minimal time expenditure.
Thanks to the determination of material properties using simulations in GeoDict, a much more detailed picture of the material can be drawn. Be it to determine the filtration performance of its fiber medium, to simulate thousands of charge cycles on its electrode to determine the durability of its battery, or to determine the relative permeability by two-phase flow. There are practically no limits to the determination of physical properties. And if there are, we are always happy to receive feedback to make GeoDict even more efficient for our customers.
As already mentioned, this analysis data can then be used for validation or to predict how the material will behave under certain conditions and influences. This is where we have already reached deep into materials research and development.
Material development is the keyword here for the last question. What is the advantage of material development in GeoDict. Especially from the point of view of image analysis, how does the workflow in GeoDict look like?
The advantage of simulation is that promising materials can be identified more quickly in advance during material development. The detailed digital view into the interior of the material, which the experiment cannot provide in this way, also provides a better understanding of which parameters of the material are particularly important for the later properties. These findings save time and thus money in the development process. There is no need to produce and test countless expensive prototypes, only the most promising materials are specifically validated in laboratory experiments.
In case GeoDict is not the end of the pipeline and you want to perform other simulations in another software afterwards, you can convert the data from GeoDict into other common file formats via the export module (ExportGeo). An example would be to translate the microstructure of voxels into a surface mesh. This way, GeoDict can be flexibly integrated into your existing workflow.
Of course, it would be beneficial to do all steps in one software to avoid friction losses by export and import. In this context, GeoDict offers further tools for material development, such as on the one hand, the analysis and simulation modules, with which the statistical material properties can be determined and on the other hand, the material design modules, which combine these statistical material properties into a material model. What you get in the end is a statistical representation of the original material, vividly visualized in GeoDict. It may be visually different from the original, but statistically it is the same material, a statistical twin. As if one had simply taken his material sample a little further to the side.
In GeoDict 2020, we have already taken the first step towards intelligent, machine-supported material development with the Find modules.
In GeoDict 2021, we have now taken a few more steps forward. The Find modules transform data into information using neural networks which can be adapted to the respective data sets by training. Thus, we get image recognition in addition to image processing and image analysis. FiberFind, for example, provides us with an exact digital twin from our fiber medium data set, which is no longer just a voxel representation, but each individual fiber is recognized as an independent object. This is what we mean when we talk about turning data into information. We now have the information of each individual fiber, know its length, diameter, curvature, etc. But what was previously only possible for fibers (FiberFind) and limited for grains (GrainFind), we have made available in GeoDict 2021 a general find module for any kind of material shape. With ObjectFind you can train your own neural network on an object shape and apply it to your data sets.
For the material development in GeoDict 2021 this means that the results from the find modules are directly entered into the corresponding material design modules as generation parameters. This saves the user from having to collect the relevant generation parameters manually and provides a more precise result.
Finally, a digital prototype is generated from the statistical twin. Now that the parameters defining the material are known, they can be changed as desired in the material design module. The effect of the modification on the material properties is then determined by simulation and compared with the desired result. This loop of design and simulation can be fully automated via the Python interface in GeoDict. You have the enormous advantage of not having to wait for the production and then the laboratory. Yes, also with GeoDict you have to wait for the simulation, but we are talking about hours or days, not weeks or months. You can easily set up this workflow with GeoPython and let GeoDict work for you over the weekend, automatically.
In summary, GeoDict is not only the ideal software for material analysis,for prediction of material properties, and for optimization of material design, but also a powerful tool for advanced image processing,which is indispensable for correct material analysis. Both in combination make possible the development of the materials of tomorrow. With our AI for material segmentation and object recognition, which has been further developed in GeoDict 2021, you too can take your material to the next level quickly and easily!