Method for quantifying the reaction degree of slag in alkali‐activated cements using deep learning‐based electron microscopy image analysis

Abstract

In this paper, we present a methodology for measuring the reaction degree of ground granulated blast furnace slag (GGBFS) in alkali‐activated cements using neural network based image analysis. The new methodology consists of an image analysis routine in which the segmentation of the back scattered electron (BSE) (SEM) images is based on a deep learning U‐net. This methodology was applied to and developed for NaOH‐activated slag cements and validated against independently measured XRD results. In a next step the developed method was applied to NaOH‐Na2SO4‐activated systems, to check the broader applicability. The neural networks based image analysis results were shown to correlate well with the XRD results. Once the model was trained, it segmented images fast and accurately. Furthermore, the model trained on the NaOH‐activated systems was readily applicable on NaOH‐Na2SO4‐activated system indicating that the model generalises well. As such, the developed methodology and models can be more performant and robust than conventional threshold‐based image segmentation. The method's accuracy, replicability and transferability make it a promising tool for material analysis and characterisation.

Segmentation of SEM images of multiphase materials: When Gaussian mixture models are accurate?

“…There are numerous segmentation methods for SEM images proposed in the literature. These methods may be either supervised [4][5][6] or unsupervised. 7,8 Supervised segmentation models are trained in a specific data set and therefore have limited applicability taking into account the large spectrum of used acquisition parameters and the differences in the composition of analysed samples.…”