Fraunhofer Institute for Industrial Mathematics ITWM
Fraunhofer Institute for Industrial Mathematics ITWM
Deep Learning für die 3D-Rekonstruktion hochporöser Strukturen aus FIB-REM-Bildstapeln
© Fraunhofer ITWM
Deep Learning für die 3D-Rekonstruktion hochporöser Strukturen aus FIB-REM-Bildstapeln

Scanning Electron Microscopy (SEM) combined with Focused Ion Beam (FIB) spatially images material structures on the nanoscale. We reconstruct porous structures using morphological image processing or Machine Learning (ML).

© Fraunhofer ITWM

Deep Learning for 3D Reconstruction of Highly Porous Structures From Fib-Sem Image Stacks

3D Images With Nanometer Resolution

Complex materials such as gas diffusion layers for fuel cells, electrodes for lithium-ion batteries, filter media, ceramic materials with active components or active-reactive coatings have microstructure components that decisively influence their material properties. Such structures can be imaged three-dimensionally with resolutions between 5 and 100 nm using the FIB-SEM serial sectioning technique.

 

Backscatter Artifacts Make the Reconstruction of Porous Structures Difficult

However, at high porosity, the 3D structure reconstructed from the 2D SEM images of the FIB sections does not correspond to the real structure. The high depth of focus of the SEM allows structural areas behind the current cut surface to also become visible through the pores and appear just as bright – causing shine-through artifacts. Reconstructing the undistorted 3D structure thus becomes a difficult image segmentation task. We solve it using both classical image processing and Machine Learning.

Zirconia Sample
© Sören Höhn, Fraunhofer IKTS
Figure 1: SEM image of a nanoporous zirconia sample.
Random pack of cylinders
© Fraunhofer ITWM
Figure 2: Example of synthetic SEM image data from realizations of stochastic geometry models; random packing of cylinders.
CoxBoole model of spheres.
© Fraunhofer ITWM
Figure 3: Example of synthetic SEM image of realizations of stochastic geometry models; CoxBoolean model of spheres.

Synthetic Image Data as a Basis

For the development of the classical algorithms and for the training of the learning procedures, it was essential to generate FIB-SEM image stacks. We physically simulate the electron-material interaction. A machine learning model trained on this data is faster.

Publications and References

Related Links: