Understanding increasingly complex materials
A major industrial challenge
Modern materials, such as metal alloys, composites, and recycled polymers, have complex microstructures consisting of grains, fibers, pores, and defects that are invisible to the naked eye. However, it is precisely these microscopic details that determine mechanical strength, fatigue life, and the risk of cracking.
Traditionally, the study of these phenomena has relied on complex physical models, the development and calibration of which requires lengthy, cumbersome, and costly experimental campaigns. Conventional mechanical tests, such as fatigue or tensile tests, require significant resources and do not allow for the exhaustive exploration of all possible configurations.
In this context, AI opens up new possibilities: by drawing on ever-increasing amounts of data while strictly adhering to the laws of physics, it makes it possible to reduce the number of tests, speed up analysis, and more effectively target the parameters that are most decisive for material behavior.
Active material particle constituting the positive electrode of NMC-type lithium-ion batteries, used for example in electric cars and phones. The sample was prepared by ion beam in a very powerful scanning electron microscope capable of seeing the internal organization of the material, then characterized by EBSD. The colors correspond to different crystal orientations, which do not all behave in the same way during the battery’s charge and discharge cycle. Analyzing the arrangement and orientation of these crystals makes it possible to perform realistic multiphysics simulations, i.e., to create highly realistic digital models to better understand what happens inside the battery during operation, providing a better understanding of the mechanisms involved in battery cycling and paving the way for improvements in battery performance and lifespan.
Authors: Clémence Pinot (CMAT), Jean-Michel Scherer (CMAT), Cyrille Colin (CEMEF), Fabrice Gaslain, Domitille Giaume (Chimie ParisTech – PSL), Aubin Geoffre (CMAT), Cécilie Duhamel (CMAT)
Automating image analysis to reveal the invisible
A first major area of research concerns the automated processing of experimental data, in particular images from X-ray tomography, a non-destructive technique that allows the reconstruction of “cross-sectional” images of a three-dimensional object, microscopy, or electron backscatter diffraction (EBSD – EBSD), a technique that maps the orientation of crystals within a material in order to understand its internal structure and mechanical behavior.
Using specialized neural networks, CMAT researchers are developing tools capable of:
- Automatically segmenting 3D images of composite materials
- Detecting defects (pores, cracks, inclusions)
- Identify crystallographic anomalies invisible in conventional analyses
For example, U-Net-type architectures, a symmetrical U-shaped architecture that allows for accurate image segmentation while preserving important contextual details, now achieve greater accuracy than a human expert, while drastically reducing analysis time. Furthermore, advanced models such as Vision Transformers are capable of capturing both local and global information, which is essential for understanding the mechanical behavior of a material under stress.
These tools pave the way for faster, more detailed, and more systematic characterization of industrial materials.
Generating realistic microstructures for non-destructive testing
Another major scientific obstacle is access to representative data, particularly when it comes to rare but critical defects. To address this, CMAT uses generative models, in particular GANs (Generative Adversarial Networks) and diffusion models.
