3D multiscale characterisation of cement-based building materials

点击翻译按钮获取中文摘要

The demands on modern building materials are constantly increasing. The development of innovative solutions has produced a multitude of binders and concretes that often have a very complex, multiscale heterogeneous composition. As the demands on building materials increase, so do the demands on characterisation. The cross-scale recording of structural characteristics is still an unsolved, but nevertheless central task. The cross-scale characterisation provides an improved understanding of essential building material properties (strength, fluid and gas transport, durability) on the one hand and is also the basis for realistic modelling on the other. In the proposed project, a combination of X-ray computed tomography (sub-µ-CT) and nanotomography in the scanning electron microscope (FIB-REM-nT) will enable a cross-scale 3D representation of the concrete structure. The evaluation of the collected data will be brought to a completely new level by applying the latest strategies for image enhancement and analysis (including machine learning algorithms). Within the first work package, an optimal sample preparation and data acquisition will be developed for both tomographic methods on uniform samples. The parallel procedure should allow a registration of the high-resolution FIB-REM-nT data within the sub-µ-CT volume. This is the basis to improve the data evaluation and thus segment more phases than possible before by applying machine learning algorithms. In the second work package, algorithms for image registration, analysis and segmentation will be developed. For example, in order to increase data quality, a swin transformer block is to be integrated into the encoder-decoder convolutional network UNet during the denoising of SEM and CT data in order to be able to model and remove different types of noise by means of synthesis of artificial training data. In order to achieve high-resolution segmentation results at the pixel and voxel level, a transformer-based neural network Swin-UNet++ will be redesigned to enable higher predictive accuracy and robustness in micro-feature identification. In the third work package, the developed procedures and algorithms will be applied to characterise the nano- to macrostructure of at least two mortar samples. These are a standard mortar and a mortar with a very low porosity. These two samples will be used as examples to develop and demonstrate the extended characterisation options that are now available, and these will also be compared with conventional methods.