Fig. 7: ML training and segmentation results for the three datasets with different contrast mechanisms.

a ML trained models overlaid on the three raw datasets. b 3D rendering of the segmented volumes at the three studied hydration ages. The DoH values determined from microtomography (bold) are compared to the ones from calorimetry (italics). The number of quantified components in the Lab-μCT and Syn-μCT datasets are four: i) porosity (air and water), ii) LDH (low-density hydrates: mainly C-S-H gel and ettringite), iii) HDH (high-density hydrates: mainly portlandite and calcite), and iv) UCP (anhydrous cement particles: all unreacted clinker phases). The number of quantified components in the PXCT datasets is seven: i) air porosity, ii) water porosity, iii) LDH (low-density hydrates: mainly C-S-H gel and ettringite), iv) portlandite, v) calcite, vi) C₃A/C₃S/C₂S, and vii) C₄AF. It is noted that with the quality of the data reported in this study (spatial resolution and electron density contrast), it is not possible to disentangle low-density from high-density C-S-H.