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 6 H I G H L I G H T S 2 0 2 2 I
data, of damage progression within nanoengineered (nanofibres used in concert with microfibres) CFRP specimens under load. The tomography data revealed 3D damage progression throughout the volume of notched test specimens as a function of increasing load with sub- micron pixel size .
Following a high-level hyperparametric optimisation study involving 20 different machines (each featuring the common encoder-decoder CNN architecture shown in Figure 117), the highest-performing trained machine was down-selected based on key machine learning efficacy metrics for deeper subsequent assessment focused on test set segmentations/inferences, given that the machines were not trained on the test set. The selected machine segmentation results were proven via 2D and 3D quantitative and qualitative analyses to be in excellent agreement (~99.99% class binary accuracies on validation and test datasets) with time-intensive, subjective human-driven semi-automatic segmentation methods, while concurrently introducing significant improvements in efficiency and consistency, and in some cases improving upon the trained-human identification of damaged regions. Particularly, machine-discovered
PRINCIPAL PUBLICATION AND AUTHORS
Deep learning unlocks X-ray microtomography segmentation of multiclass microdamage in heterogeneous materials, R. Kopp (a), J. Joseph (b), X. Ni (a), N. Roy (a,b), B.L. Wardle (a,c), Adv. Mater. 34, 2107817 (2022); https:/doi.org/10.1002/adma.202107817 (a) Department of Aeronautics and Astronautics, MIT, Cambridge, MA (USA) (b) MIT Quest for Intelligence, MIT, Cambridge, MA (USA) (c) Department of Mechanical Engineering, MIT, Cambridge, MA (USA)
 X. Ni et al., Compos. B. Eng. 217, 108623 (2021).
Imaging 3D preserved soft tissues and organs in a 380-million-year-old fish
Fossils of extinct animals usually comprise only mineralised tissues, teeth, scales, spines and the skeleton but, in rare instances, soft anatomy is also preserved. X-ray phase-contrast microtomography was used to image soft tissue remains from a 380-million-year-old fish, revealing the absence of lungs and thus refuting the hypothesis that lungs are ancestral in jawed vertebrates.
The living jawed fishes comprise two groups: the cartilaginous fishes (sharks, skates and rays) and the bony fishes (ray finned and lobed finned), and both show a mosaic of ancestral and derived traits. Placoderms stem-jawed vertebrates are an important group for
determining which characters are ancestral but, until now, this determination has been limited to the skeletal characteristics. The Gogo Formation in the Kimberley region of Western Australia represents an ancient reef complex that today yields not only skeletal remains but also soft tissue remains from an extinct group of early jawed vertebrates, the placoderms. In the year 2000, small amounts of mineralised muscle tissue were recognised between two dermal plates of placoderm armour. It became apparent that the muscle tissues were being destroyed through acid preparation, which had been the standard technique to remove the fossil from the matrix. However, attempts to scan the fossils using computerised tomography (CT) proved unsuccessful.
Using X-ray phase-contrast microtomography at beamline ID19 to image a fossil sample, the musculature of the neck and abdominal region were identified and
human segmentation error manifested primarily as new damage discovery and segmentation augmentation/ extension in artefact-rich tomograms. Figure 118 presents examples of machine vs. human segmentation, comparing their ability to classify damage.
Machine-driven segmentation can accelerate discovery within new datasets, including basic structure-property relationships underpinned by failure mechanisms across a spectrum of heterogeneous materials (e.g., biological and biomimetic materials), as the machine effectively eliminated the time-consuming human effort associated with tomography segmentation. For each 3D scan, what used to be performed in ~10 hours by a trained human can now be done in ~1 hour by the trained machine (without human intervention). For example, for the 30 scans captured as part of a two-day beamtime mission, it takes 60 working days for a human vs two days for a computer to segment. This utilises 1 GPU as considered here for the trained machine, and it is noted that further acceleration is achievable with multi-GPU machines, or multiple machines working in parallel, as well as by potentially initialising future machines with the presented model to aid learning efficacy.