Super-resolution X-ray tomography using deep learning applied to the 3D quantification of defects in lattice...

Lattices are intrinsically multiscale materials, forming large structures composed of repeated unit cells, while also including relatively small defects such as pores or grooves. Those defects are detrimental to their mechanical properties and must be quantified. X-ray tomography (CT), a 3D non-destructive imaging technique, is an excellent candidate for this task but it is limited by a trade-off between spatial resolution and scan time. Hence, fully imaging the lattice at a resolution providing a clear depiction of the defects of interest is not compatible with its multiscale aspect, as it would require a prohibitive amount of scan time. Recently, deep learning-based super-resolution has shown remarkable advances in improving spatial resolution of low-resolution CT. However, the validation of these super-resolution workflows is often based on visual quality metrics that do not directly assess the ability to capture critical metrics or phenomena relevant to material scientist. The present study aims at redefining image quality from a material science and task-based perspective, and to quantify the measurement uncertainties associated with super-resolution applied to the 3D characterisation of defects in lattice structures. To address this issue, we have designed a comprehensive super-resolution workflow using a mixed-scale dense network, covering data acquisition, preprocessing, and tailored algorithm validation. The method was tested experimentally on a steel lattice produced by laser powder bed fusion. Super-resolution volumes were computed and their quality was assessed from global greyscale data down to the local scale, investigating both key features of interest: porosity and surface roughness. This approach enhances image quality and improves the morphometric depiction of defects, while enabling a significant reduction in scan time, reaching several orders of magnitude. Thus, we demonstrate that defects inspection in multiscale material such as lattices is now feasible within a reasonable timeframe through deep learning-based super-resolution.