S T R U C T U R E O F M A T E R I A L S

S C I E N T I F I C H I G H L I G H T S

1 2 4 H I G H L I G H T S 2 0 2 2 I

PRINCIPAL PUBLICATION AND AUTHORS

Structure of an amorphous calcium carbonate phase involved in the formation of Pinctada margaritifera shells, T. A. Grünewald (a), S. Checchia (b), H. Dicko (a), G. Le Moullac (c), M. Sham Koua (c), J. Vidal-Dupiol (d), J. Duboisset (a), J. Nouet (e), O. Grauby (f), M. Di Michiel (b), V. Chamard (a), Proc. Natl. Acad. Sci. 119 (45), e2212616119 (2022); https:/doi.org/10.1073/pnas.2212616119 (a) Aix-Marseille Univ., CNRS, Centrale Marseille, Institut Fresnel, Marseille (France) (b) ESRF (c) Ifremer, ILM, IRD, Univ. Polynésie française, EIO, Taravao, Tahiti, French Polynesia (France) (d) IHPE, Univ. Montpellier, CNRS, IFREMER, Univ. Perpignan Via Domitia, Montpellier (France) (e) GEOPS, Univ. Paris-Sud, CNRS, Univ. Paris-Saclay, Orsay (France) (f) Aix-Marseille Univ, CNRS, CINaM, Campus Luminy, Marseille (France)

REFERENCES

[1] F. Mastropietro et al., Nat. Mater. 16(4), 946-952 (2017). [2] J. Duboisset et al., Acta Biomater. 142, 194-207 (2022). [3] H. Dicko et al., J. Struct. Biol. 214, 107909 (2022).

This study opens the door to more fundamental understanding of the biomineralisation process by accessing the crucial first ACC states while providing sub-micrometric spatial resolution. Eventually, the

tools developed for this experiment will help to study other amorphous/crystalline systems and disentangle the complexity of crystallisation in non-classical processes.

Automating segmentation of complex damage in aerospace composites tomography by deep learning

Machine learning techniques were applied to X-ray computed microtomography imaging of multiclass microdamage in heterogeneous composite materials used for aerospace structures. The machine-driven segmentation effectively eliminated the time-consuming human effort associated with tomography segmentation as well as new damage discovery in large datasets.

Advanced composite laminates comprised of carbon fibre reinforced polymer (CFRP) have become widespread in modern high-performance aerospace structures, providing a superior combination of tailorable mechanical properties and low weight. However, CFRPs also exhibit complex damage mechanisms that lead to difficult-to-predict failure, limiting even more widespread use. As a state-of-the-art visualisation technique, in-situ synchrotron radiation computed tomography (SRCT) couples mechanical testing with non-destructive 3D X-ray imaging to reveal composite morphology, manufacturing defects, and complex damage progression in real-time [1]. Yet, objective insights are presently extremely challenging to extract from the large (∼10 GB/mm3), typically damage-sparse SRCT datasets due to time-intensive, subjective semi-automatic (human-driven) damage segmentation techniques, culminating in a data-to-knowledge bottleneck. Machine learning techniques using convolutional neural networks (CNNs) offer the possibility of automated segmentation

but have never been applied to multiclass microscale damage segmentation.

Generalised automated damage segmentation that would promote objectivity, speed, and feature flexibility has thus far been impossible to code using traditional rule-based programming, which employs digital image processing tools (e.g., filtering, thresholding, clustering). CNNs are a spatially invariant class of deep artificial NNs (i.e., NNs for deep/multi-layered learning) that especially excel in scalable discriminative ( example-based ) learning of large, complex imagery datasets. To date, CNN-based CT segmentation of material degradation has been leveraged by several fields, including medicine and materials engineering (e.g., concrete and advanced composites), although to the authors knowledge, multiclass, micron-scale damage (and progression) that underpins advanced composites failure remains entirely unexplored and a critical open research question in the space of heterogeneous materials generally. Figure 117 presents a comparative view of human-driven and (the new) machine-driven (CNN) end-to-end approaches for CFRP damage segmentation.

In this study, CNN machines were trained to segment complex, sparse (<<1% of scan volume) multiclass microdamage in CFRPs using 65 000 human-segmented tomograms from 30 SRCT scans comprising six different specimens and two different CFRP types, enabling a robust study of performance generalisability. The SRCT dataset, collected at beamline ID19 by the authors in collaboration with the University of Porto (Portugal) and the University of Southampton (UK), 4D, i.e., spatial and temporal