Hundreds of synthetic graphite anodes were generated as digital material models during the project and used as training data for the Neural Network (NN).

A typical graphite anode contains different material phases, which we call constituents: the graphite particles, the so-called Carbon Black & Binder Domain (CBD), and the pore space, which is filled with electrolyte. CT-scans of various graphite anodes were taken as part of the structur.e project. On these scans, the microstructure of the anodes is displayed in 3D and can be examined.

µCT 3D image of a graphite anode
Fig. 1 µCT 3D image of a graphite anode. The individual material phases (constituents) are usually well separated by their gray value. In graphite anodes, however, the graphite particles and the so-called carbon black & binder domain (CBD) often have the same gray value, because the two constituents consist mainly of carbon. The pore space (dark) is clearly different from the two solid phases.

The constituents are distinguishable by their gray value, as shown in Fig. 1. However, in a graphite anode, it is usually impossible to separate the graphite particles from the CBD, because both constituents are largely composed of carbon and therefore, display a very similar gray value. This is where Artificial Intelligence (AI) can help: Neural Networks (NN) can learn to differentiate the two constituents based on geometric features, for example.

Artificial digital model of a graphite anode
Fig. 2 Artificial digital model of a graphite anode statistically showing the same geometrical and physical properties as the "original". Graphite particles are shown in gray and the CBD in red. The right table shows the comparison of the geometric and physical properties of the pore space.

Hundreds of artificial graphite anodes were generated using the digital material models created during the project. The artificial graphite anodes were then used as training data for the NN. These artificial structures resemble the real anodes (see Fig. 2) in all essential characteristics. Through the training data, the NN learns to distinguish the graphite particles from the CBD, because in generated structures the "correct solution" is always known.

µCT image of a graphite anode
Fig. 3 µCT image of a graphite anode. On the left the gray scale image and on the right the digital model segmented using artificial intelligence. All constituents, the graphite particles (gray), the CBD (red) and the pore space are visible.

This allows the NN to distinguish the constituents on the CT images of the real anode as well (see Fig. 3). After the trained NN has successfully separated particles and CBD, Math2Market and our project partners use the material model of the real anode to analyze and improve the properties. Faster-charging batteries for the next-generation of E-cars are the goal.

The trained Neural Network is available to our clients as part of the GrainFind-AI module of the GeoDict 2023 release.