Cement and Concrete Research 141 (2021) 106327

Contents lists available at ScienceDirect

Cement and Concrete Research

journal homepage: www.elsevier.com/locate/cemconres

edxia: Microstructure characterisation from quantified

SEM-EDS hypermaps
Fabien Georget *, William Wilson, Karen L. Scrivener
Laboratory of Construction Materials, LMC, EPFL-STI-IMX, Station 12, CH-1015 Lausanne, Switzerland

A R T I C L E I N F O A B S T R A C T

Keywords: The characterisation of cement paste microstructure is an important step towards understanding durability
SEM-EDS mechanisms in cementitious materials. Scanning electron microscopy (SEM) coupled with energy dispersive
Microstructure spectroscopy (EDS) is a widely used technique to analyse the microstructure at the micron-scale. However, it is
Image analysis
challenging, notably because the characteristic size of many phases is found on a scale smaller than the EDS
Quantitative analysis
Data analysis
interaction volume. This work presents a new image analysis framework to identify phases and quantify the
microstructure of cementitious materials from SEM-EDS hypermaps. By leveraging domain knowledge, repre­
sentative points are attributed to phases and mixtures of phases based on ratio plots. Then, quantitative analysis
of the microstructure can be carried out (chemical composition, particle size distributions, volume fractions, …).
We demonstrate the abilities of the framework, and we present possible applications and extensions of the
method. The framework is available as both a graphical interface and a Python code.

1. Introduction qualitative mappings). However, this approach has limitations. First, the
particles can only be identified one by one, limiting the understanding of
Scanning electron microscopy (SEM) coupled with electron disper­ the distribution of phases, as well as the estimation of their volume
sive spectroscopy (EDS) has many applications in cementitious mate­ fractions. Secondly, representative quantitative information about the
rials. For example, it can be used to obtain the C-A-S-H composition [1], chemical composition may not be easily retrieved. Thirdly, from XRD
to analyse the composition of complex SCMs [2,3], to calculate the de­ analysis, we know that the ettringite content should be around 12% (in
gree of reaction [4,5], or to interpret the results of other methods such as mass), although this phase cannot be identified on the BSE image! We
nanoindentation [6,7]. A particular use of SEM-EDS is the acquisition of are looking for a method which is able to solve all of these limitations.
multispectral or hyperspectral mappings. In these mappings, each pixel The identification of phases is a clustering problem [6,7,13–16],
contains a complete EDS spectrum which can be quantified to obtain the which can be solved with the use of automatic clustering algorithms,
chemical composition, and then combined with spatial information to such as the decision-tree [14], the SVM [15], the k-means [16], or the
provide mappings of all quantified elements. In combination with Gaussian mixture clustering [6,7] algorithms. Although good results can
backscattered electron (BSE) micrographs, such mappings can be used to be obtained, these techniques are not widely used in practice, even if
study hydration [2,3,8], and/or durability issues [2,9,10]. The amount some tools (e.g. [16]) are available in well-established software, such as
of information in these hyperspectral images is large but, unfortunately, ImageJ [17,18]. In the opinion of the authors, this is due to the intrinsic
it is not easily accessible due to intrinsic difficulties of displaying and complexity of both the algorithms and the cement paste microstructure,
interpreting this information. The challenge is twofold: the phases must compared to the usefulness of the information that can be directly ob­
first be identified [2,3,8,11], and a meaningful quantification must then tained. The clustering algorithms have been developed to analyse a large
be obtained [3,11,12]. volume of data with a lower signal to noise ratio than those available in
As an example, Fig. 1 shows an annotated version of a BSE image of a practice.
limestone calcined clay cement paste (LC3) sample that will be studied in For SEM-EDS mappings, time is the main limiting factors. Studies
this article. An expert can already identify phases in this image, espe­ based on a large number of hypermaps are rare. For example, the work
cially by using the information from EDS (either point analysis or presented here is based on just 19 hypermaps. More maps would have

* Corresponding author.
E-mail address: [email protected] (F. Georget).

https://doi.org/10.1016/j.cemconres.2020.106327
Received 1 June 2020; Received in revised form 2 October 2020; Accepted 30 November 2020
Available online 16 December 2020
0008-8846/© 2020 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

improved the statistical significance of our results, but this is the con­ presented here as a step in the data analysis. The importance of these
dition of many studies using SEM-EDX as a tool for a mean. The main scientific questions over the clustering problem, has led to the combined
limitation is the time required to prepare the sample, analyse the sample use of multispectral images (generally qualitative) to study the spatial
in the microscope (a couple of hours to a day for a high-quality map) and distribution of phases and, EDS point analyses to study their composi­
quantify the maps (typically 10 h). Therefore, it is important to use a tion [5,7,9]. However, to obtain quantitative measures of the spatial
reproducible and reliable analysis approach (that can be checked), distribution, a bridge between a robust phase separation and quantita­
rather than a fully automated analysis that can treat a large amount of tive chemical analyses is required.
data. The importance of constraining the algorithms by domain- In this article, we propose a new approach that allows phase sepa­
knowledge was previously recognized by Munch et al. [16], who fully ration, chemical analysis, and visualization on the BSE image. It ad­
integrated inputs from the expert into their algorithm. A second dresses both the physical and the practical challenges of analysing SEM-
requirement for the adoption of these tools is the use of a graphical EDS hypermaps of complex cement paste microstructures. The proposed
interface [12,16], which can provide a direct feedback to the user. Many method is based on a combination of (1) a denoising of quantified maps,
of these algorithms require a large number of steps and parameters to be (2) a user defined phase clustering based on domain knowledge and (3) a
adjusted by the user. A contrast can be made with the Rietveld method visual tool to interact with the user (i.e., with the domain-knowledge
for quantitative X-Ray analysis: it is similar in complexity, but the ease- and the research problem). First, using a segmentation of the compos­
of-use of common software makes it a tool available to researchers ite image [3,7,11], we extract representative points from the quantita­
without a programming background [19]. In addition, the clustering tive EDS hypermap. The emulation of the point analysis method using
problem and the separation of phases is just a first step. It is important to ratio plots [1,19] allows the composition of the phases to be studied, but
provide to the user analysis tools [11] that can be used to answer specific also identifies mixtures of these phases from representative points. In
scientific questions. For example, below is a non-exhaustive list of addition, phase masks can be defined by translating the point clusters
research questions (from different fields of the cement material science) onto the maps. Then, particle shapes and spatial distribution can be
that could benefit from a powerful hypermap analysis tool: analysed from these phase masks. This method was implemented into a
new framework called edxia, and added into the Glue software [20,21]
• Where are the minor elements in polyclinker phases? to make this tool accessible to all researchers. This accessibility as well
• Where is the ettringite in the microstructure? as the flexibility of our method represent an important advantage
• Where does strätlingite precipitate when it forms? compared to previous approaches. Thanks to the versatility of the
• Does exposure to chlorides change the morphology of AFm phases? approach, the scientific questions presented above were all studied with
• What is the composition of AFm phases after exposure to chlorides? our software as part of the development process (although they could
• Can we directly quantify chloride binding in C-S-H? not be described in detail here).
• What are the characteristic features of a carbonation front? In this paper, we present the experimental protocols for sample
• What is the impact of sulfates on the hydration process? preparation, SEM-EDS hypermaps acquisition and quantification. Sec­
• How is the microstructure impacted by sulfate attack? tion 3 presents in details the image analysis methods and its justifica­
tions. Section 4 gives some examples of microstructural investigations
All these questions are currently under study in our lab using the tool enabled by quantitative phase maps, chemical analysis of phases,

Fig. 1. Phase annotation from BSE and elemental maps of an LC3 sample.

2
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

quantification of phase volume fractions and estimation of microstruc­ vacuum).

ture descriptors. Section 5 presents some additional applications for our
tool: investigation of profiles, coping with lower-efficiency EDS de­ 2.2.2. SEM parameters
tectors, filtering data at the spectrum level, and incorporating machine The majority of SEM-EDS analyses were performed with a Zeiss
learning. Overall, this paper not only describes the method and some Schottky SEM equipped with a Gemini®2 column. The accelerating
current applications, but also aims at setting the stage for further de­ voltage was 15 kV to enable quantification of iron (to access K lines of
velopments in an open-source framework. iron). Previous work showed that with such a voltage, 95% of the
characteristic X-rays escaping cementitious materials were generated
2. Materials and methods within a depth of ~900–1700 nm (depending on the phase, it’s density
and on the measured element) [7]. The beam current was adjusted to 2
2.1. Materials nA (using a Faraday cup) to obtain the maximum signal without satu­
rating the EDS detector. Although this value is higher than the recom­
To illustrate the robustness of the edxia approach, a CEM I Ordinary mendations by Rossen et al. [1], the damage to the sample was limited
Portland Cement (OPC) and blended system were chosen as examples of by the ultra-rapid scanning of the surface enabled by an Oxford Ultim­
cementitious microstructures with different levels of complexity. More Max 170 EDS detector having a silicon drift detector surface of 170 mm2.
specifically, the blended system is a limestone calcined clay (LC3) Unless otherwise specified, the EDS hypermaps were acquired using
cement paste [22]. LC3-50 paste incorporates 53% of OPC (CEM I), 30% the Aztec software at a resolution of 1024 × 768 pixels over regions of
of calcined clay (fired and ground in India, with a calcined kaolinite 280 μm by 210 μm. To avoid image quality losses due to possible drift of
content of ~45%), 15% fine commercial limestone powder and 2% the surface, the Autolock drift correction was employed with a mea­
chemical-grade gypsum. The chemical composition of the cement and surement every 30 s and predictive correction every 10 s. The hypermap
calcined clay obtained by XRF are presented in Table 1. The LC3 cement acquisition parameters were set as a compromise to obtain sufficient
paste was mixed with deionized water. counts at each pixel, while limiting both surface damage and the
required machine time. The combination of 256 s dwell time per pixel
2.2. Methods with an averaging over 12 frames was found to provide good results
(although other satisfactory combinations could be possible). In this
2.2.1. Sample preparation condition, the acquisition time of a map is around 1 h.
To ensure the homogeneity of the samples and the representativeness To ensure the consistency of EDS analyses, the beam intensity was
of the areas investigated, the cement pastes were mixed for 2 min with a calibrated on a copper tape before each analysis (even if Schottky SEMs
5 cm axial flow impeller at 1600 rotations/min. This high-shear mixing are extremely stable). The deadtime was kept under 30–35% for the
was followed by 1 min of vacuum mixing at 450 rotations/min to different types of polished sections investigated. To achieve quantitative
remove entrapped air. A typical water-to-binder ratio of 0.4 was chosen EDS analyses, certified standards were used for all elements of interest:
and a polycarboxylate-based superplasticizer was employed to obtain CaSiO3 for O and Ca, Jadeite for Na, Al2O3 for Al, SiO2 for Si, CaSO4 for
sufficient workability for casting. The paste was cast in cylindrical S, KCl for K and Cl, MgO for Mg and Fe2O3 for Fe (standards from Micro-
polypropylene molds of 33 mm diameter, which remained sealed for the Analysis Consultants Ltd). After acquisition, each hypermap was quan­
first 20 h before demolding. They were then water cured for 28 days in tified using the Aztec built-in QuantMap function with the processing of
slightly larger containers. In addition to limiting the amount of water all elements, no normalisation, corrections for the window artefact and
with the container size, any possible leaching was further avoided by pulse pile up. Although the oxygen may not be perfectly measured, it
adding to the container a sacrificial finely-ground specimen acting as a was still directly quantified from the spectrum which was deemed the
leaching buffer protecting the sample. best option for cement pastes: stoichiometric oxygen from oxides would
At 28 days, 2 mm thick slices were cut and dried by isopropanol lead to large errors of matrix correction, because of the important non-
solvent exchange for 7 days (with renewal of the isopropanol after 1 h, 1 quantified hydrogen and carbon contents. Nevertheless, the content of
day and 3 days) followed by isopropanol evaporation under slight vac­ water and carbonates was still estimated using the sum of oxides (SOX),
uum for more than 7 days (with CO2 and H20 absorbing agents). After which sums the mass of all measured elements and their stoichiometric
drying, smaller pieces of the slices were polished with a silicon carbide oxygens (considering the oxides expected to occur in the cement paste,
paper #1200 (to remove any possible carbonation and provide a flat e.g., CaO, SiO2, Al2O3, Fe2O3, etc.)
surface). The pieces were then impregnated under vacuum using the In addition to the main analyses carried out with approach, the
EPO-TEK 301 epoxy into 25 mm diameter molds, to support the delicate choice of parameters and the possibility of using lower-efficiency de­
microstructure of the cement paste hydrates. The embedded samples tectors was explored with a Quanta 200 tungsten filament SEM equipped
were then polished using a Struers RotoPol-25 automatic polishing with a 30 mm2 EDS detector (Brucker XFlash 4030). The same standards
machine with Struers MD Largo discs and DP-Spray M diamond sus­ were employed with a similar calibration and quantification approach.
pensions, with deodorized petroleum as lubricant. A high-quality Different sets of parameters were used, with two types of instruments as
mirror-like surface was obtained by successive polishing for 15 min described in Table 2.
with a 9 μm diamond suspension under a pressure of 15 N, followed by 1
h with a 3 μm diamond suspension at 20 N and 2 h with a 1 μm diamond 2.2.3. Additional characterisation of the cement paste
suspension at 20 N. Between each step (and after each hour of polish­ In addition to SEM-EDS analyses, X-ray diffraction (XRD) and ther­
ing), the samples were cleaned in an isopropanol ultrasonic bath for 2 mogravimetric analyses (TGA) were also done on the same LC3 paste to
min. After 2 days under slight vacuum to evaporate the remaining iso­ provide reference phase fractions. XRD analyses were performed on
propanol, the samples were carbon coated with a thickness of about 15 fresh discs of 33 mm diameter, using a PANalytical X’pert Pro Bragg–­
nm (to make the surface conductive and enable SEM analyses at high Brentano diffractometer. The parameters were chosen according to the

Table 1
Chemical composition (mass%) of the cement and the calcined clay used in this study. LOI = loss of ignition.
Material Na2O MgO Al2O3 SiO2 SO3 K2O CaO TiO2 Fe2O3 LOI

CEM I 0.4 1.1 5.4 20.7 2.9 0.7 65.0 0.2 1.9 1.8
Calcined clay 0.1 0.0 36.7 56.1 0.1 0.2 0.2 2.4 3.4 0.5

3
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

Table 2 SEM-EDX hypermaps, and (2) relate this ratio plots to the BSE image,
EDS detectors and hypermaps acquisition parameters used in this study. including the analysis of the mixed phases. After these processing steps,
Set EDS Resolution Beam Dwell time/ Total different types of analyses can be run on these masks depending on the
detector current frames time specific scientific problem. The flowchart of the complete approach is
A Brucker 30 1024 × 768 ~6 nA 512 μs/10 ~1 h presented in Fig. 2, and each step is further described in a corresponding
B Brucker 30 1600 × ~0.8 nA 512 μs/30 ~8 h section of this paper.
1200
Ca Oxford 170 1024 × 768 2 nA 256 μs/12 ~1 h
3.1. Denoising of chemical maps
D Oxford 170 2048 × 2 nA 256 μs/12 ~4 h
1536
E Oxford 170 1024 × 768 4 nA 256 μs/6 ~30 min Unless the user has access to a high-end EDX detector for a long
F Oxford 170 512 × 384 2 nA 256 μs/12 ~15 min period of time, the number of counts per pixel is usually below the
a
Set C is the default parameter set used unless otherwise stated. threshold required for a reliable point analysis [1], i.e., the maps are
noisy. The noise manifests itself as brightness variations on the maps (i.
e., a high-frequency texture on homogeneous regions). As an example,
equipment available and to the recommendations of [19]: a copper tube
this texture is visible on the raw calcium map presented in Fig. 3. This
operated at 45 kV and 40 mA; Soller slits of 0.04 rad for incident and
texture significantly lowers the output quality of the image analysis al­
diffracted beam; incident anti-scatter and divergence slits fixed at 0.5◦ ;
gorithms. Therefore, it must be removed before further analysis.
beam knife; X’Celerator linear position-sensitive X-ray detector with a
Although some pixels may not be reliable on their own, we can use the
length of 2.122◦ 2θ; scanning from 5 to 70◦ with a 0.017◦ 2θ step;
information distributed over the large number of pixels to refine the
spinning stage at 15 rotation per minute; 30 s per integrated step, for a
maps.
scan duration of about 15 min. The patterns were analysed using the
Many denoising algorithms exist. As a guideline, a good denoising
Rietveld method with an external standard [23], along with the iterative
algorithm should remove the high frequency texture from the map while
approach proposed by Li et al. [24]: the cement phases were first refined
keeping sharp edges. The smearing of the edges leads to an apparent
from the pattern of the anhydrous cement and then transferred to the
increase of the interaction volume effect. In addition to the raw map,
refinement of the cement paste; this latter quantification was then
Fig. 3 shows the results from three denoising algorithms on the Ca map:
normalised with the K-factor obtained using a rutile standard and the
the total variation algorithm [32], a Gaussian filter, and the joint
mass attenuation coefficient of the sample. This approach provides ab­
bilateral filter [33,34]. The joint bilateral filter is an edge-preserving
solute mass fractions for crystalline phases and total amorphous content.
filter using a different (but related) image for edge detection. The
TGA analyses were made with a Mettler Toledo TGA/SDTA851 instru­
reference image for edge detection in our workflow is the BSE image.
ment, considering the approach described in [19]. A sample of about 50
The maps are displayed in the left column of Fig. 3, while the histograms
mg was placed in an alumina crucible and tested with a 10 ◦ C/min
corresponding to these maps are shown in the right column. Each al­
temperature ramp from 30 ◦ C to 1000 ◦ C, in a nitrogen flow of 30 mL/
gorithm depends on parameters controlling the extent of the denoising.
min. The recorded curves were evaluated using the tangent method,
The total variation algorithm depends on a weight parameter (here
mainly to estimate the portlandite Ca(OH)2 content from the water loss
equal to 0.1), the Gaussian filter depends on a scale parameter (here
centered around 450 ◦ C.
equal to 0.5) controlling the width of the Gaussian window. The joint
bilateral filter depends on a space parameter (here 10.0) and an intensity
2.3. Implementation of the edxia method parameter (here 2.5). For typical SEM-EDX hypermaps of cementitious
materials, we observed that the total variation algorithm provides a
The new framework developed during this work, edxia, is coded in better denoising than the standard medium filter, that is robust with
python [25], using the well-established scipy [26,27], matplotlib [28] respect to the parameters. On the other hand, the Gaussian filter is a
and scikit-image [29] libraries. The Python scientific stack was chosen difficult trade-off between the removal of the noise and the smearing of
for its flexibility and extensibility. The framework is developed as an the information with no clear optimal points. Because the bilateral filter
Application Programming Interface (API). It means that the user can was developed to reduce the smearing across edges, it is a good choice to
develop their own analysis tools, and work-flows adapted to their limit the impact of the denoising. Therefore, it can reduce the need for a
problems. The open-source framework is freely available [30]. compromise, but it is quite sensitive to parameters, which should be
To make the framework available to researchers without coding adapted to the resolution, the field of view, and also the contrast in the
experience, it was included as a plugin to Glue [20,21]. This python BSE image.
software was developed to easily visualize and create links between Another method to judge the effect of denoising is to analyse the
datasets. For example, Glue was initially developed to identify stars in histogram of the intensity of the images. The denoising results in a
telescope pictures from star catalogs. We used it here to identify the refinement of the main peak, and the appearance of secondary peaks.
phases on the BSE map, using the chemical information provided by the The main central peak corresponds to the main hydrates phases such as
EDS analysis. Documentation of the interface is also available [31]. the C-S-H and AFm phases (layer double hydroxides of the hydro­
calumite family) [35]. The peak at high Ca corresponds to the clinker
3. The edxia method phases. The shoulder on the C-S-H/AFm peak is the portlandite. This
shoulder is more visible for plain cement (CEM I) where the portlandite
To answer the challenges identified in the introduction, our method content is higher. The peaks at low Ca are the metakaolin and the mixes
follows a sequential approach summarized as follows: between metakaolin and hydration products. Even if the map is more
refined after denoising, the peaks are still overlapping. Therefore, they
1. Maps are denoised and combined cannot be used to quantitatively select phases on their own. It should be
2. Representative points are extracted from the maps noted that the denoising cannot be optimized based on just the width of
3. Ratio plots are used to identify phases by the user the peaks in the histogram. Because, it would lead to over-smoothing,
4. Masks are created from the regions identified in the ratio plots, to where mixes of phases cannot be resolved anymore.
identify phases in the microstructure. Denoising is not an operation without compromise, as the denoising
might introduces a bias. However, ratios of composition (Al/Ca,Si/Ca,S/
The novelty of the method is (1) to propose a reproducible combi­ Al, …) were found to be quantitative on average, as mentioned by
nation of methods to obtain a ratio plot than can be analysed from a Harrisson et al. [36]. This point is developed further in Section 5.2. Since

4
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

Fig. 2. Flowchart of the edxia method. Numbers indicate the section of this paper.

the total-variation algorithm provides the best results in terms of The resulting composite image is shown in Fig. 5a. This image can be
robustness and final aspect of the maps, it was used in the remainder of analysed to separate the phases by color. The main phases are identified
the examples presented in this paper. It is also the default option in the in the legend of that figure. To better identify phases by colors, the
implementation of our method. background color of the legend is set to an average color of the outer C-S-
H matrix.
Phase masks could be obtained from this composite image, as usually
3.2. Combination into composite maps performed by transforming the RGB channels into hue, saturation and
brightness (HSB) channels [37]. The phases are then defined by
A common approach to analysing chemical information, e.g. thresholds on the hue. Additional thresholds might be defined on the
[3,11,37], is the creation of a composite map. The composite map is also saturation, brightness, and/or on the BSE value, to separate between
used as a pre-processing step in our workflow to increase the signal to anhydrous phases and hydrates. The analysis of the composite image is
noise ratio. This composite map is obtained by merging scaled elemental useful to quickly analyse a map. However, the main limitation is that it is
EDS maps assigned to different color channels. A convention for creating limited to 3 components, or 4 if a gray channel is added instead of the
composite maps is the use of Ca, Si and Al elemental maps, which allows BSE map. Another limitation is that the threshold is based on a non-
the best separation between the main cement phases [3,38]. physical parameter, the hue rather than the composition. Although it
The individual colored maps, for the LC3 cement paste are shown in is sufficient to distinguish C-S-H from portlandite, it is not directly
Fig. 4. By assigning these maps into the corresponding channels of a RGB possible to separate hemicarboaluminate from the monosulfoaluminate,
image (red = Si, green = Al, blue = Ca), a composite map can be ob­ for example. An analysis using all the available elements is necessary
tained. Each individual map can further be scaled to obtain an adjusted [16].
range of colors. The composite map can also be displayed as a trans­
parent layer over the BSE map to highlight the boundaries of the par­
ticles, or mixed with another gray-level map (e.g. Fe, Mg) to increase the
visual separation of specific phases.

5
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

Fig. 3. Effect of the denoising algorithm on the Ca map. The left column presents the calcium maps, the right columns are the corresponding histograms. The red line
was positioned at the maximum value of the histogram corresponding to the Ca map denoised with the total variation algorithm. Parameters for the denoising
algorithms are described in the text. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

3.3. Segmentation of regions with similar compositions of particles, or homogeneous sections of particles or gel. Standard al­
gorithms can provide an adequate result for the anhydrous clinker
To facilitate the identification of the phases, the composite map can grains. However, they usually fail for the hydrated phases as their
be separated into regions using a superpixel segmentation algorithm characteristic sizes are smaller (sometimes even smaller than the reso­
[39]. This segmentation identifies adjacent pixels of similar color (i.e. lution), and the gray levels are similar and dependent on the sample and
similar composition for a composite map) and group them into contin­ the microscope. Therefore, it is preferable to carry out the segmentation
uous regions. on the composite map. The segmentation algorithm creates region of
A segmentation algorithm could be applied to the gray BSE image in similar composition. The degree of similarity considered here is usually
order to separate the regions without information about the composi­ defined by the number of regions created by the algorithm. In most al­
tion. These regions would correspond to particles, grains, agglomerate gorithms, the number of regions is a function of a user-defined

6
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

Fig. 4. Components of a composite image. These denoised maps are each assigned to the channels of an RGB image to create the composite image.

Fig. 5. a) Composite image formed from the combination of the individual channels presented in Fig. 4, and, b) corresponding segmentation using the SLIC al­
gorithm. The dots represent the geometrical center of each segmented region.

parameter. 3.4. Emulation of point analysis

In edxia, the SLIC segmentation algorithm [39] is used by default. Its
effect is shown in Fig. 5b. As observed in this figure, the main features of Chemical analysis with SEM-EDS is usually done with point mea­
the composite map are kept after the segmentation. The main advantage surements where the results are plotted as ratio plots [19,36]. For
of this method is that we can transform a large number of pixels (e.g., example, the composition of C-A-S-H is usually determined using an Al/
1024 × 768 = 786, 432) into a much smaller number of regions (e.g., Ca vs. Si/Ca scatter plot [1]. Our framework emulates that approach.
<20, 000). Although it is not mandatory, this step is important to The main challenge is to select representative points in the map using a
simplify further analysis. It should be noted that the loss of information reliable and reproducible method. For ideal systems made of particles in
during this step is minimal, because neighboring points will have a a gel matrix, the most representative points are at the center of the
similar composition (due to the overlapping interaction volumes, and particles, because the effect of neighboring phases in the interaction
the denoising). On the other hand, the signal-to-noise ratio is increased volume are minimized. Unfortunately, as described in the previous
because many points at the interface between two phases, or in pores are section, it is not possible to identify perfectly each particle that can be
not considered. detected visually in the backscatter micrograph. One of the main reasons
is that superpixel algorithms constrain the superpixel sizes to be of

7
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

similar sizes, even if the particles sizes in the systems vary by a few identification of each pixel to a particular phase is not possible without
orders of magnitude. However, we have identified regions of similar strong assumptions. Instead, the characterisation of these mixtures is a
composition by segmentation. As an approximation, each of these re­ feature of this framework.
gions is assimilated to a particle, a section of a particle, an agglomerate To achieve this goal, we adopt a manual approach using ratio plots.
of particles, or a gel of relatively homogeneous composition. Therefore, In a ratio plot, the points from the EDS analysis are plotted on axes of
the representative points are taken at the geometric center of these elemental ratios. The most common is the Al/Ca vs. Si/Ca plot, which
segmented regions. The reasoning is that for perfectly circular regions in can be used to identify the main phases of cementitious materials. An
a matrix of outer C-S-H, only the pure phases are sampled, and not the idealized version of this plot is displayed in Fig. 7. The colored circles
mixture of phases. Of course, real cementitious systems are not made of represent the common end points, representing “pure phases” at the
circular regions. However, this approach has been shown to provide an microvolume level, and the gray points represent the main mixture lines.
adequate filtering as long as the number of points is adjusted to the In practice, the gray points were obtained by using random samples from
resolution and the field of view of the map. The filtering can be observed a triangular distribution centered along the pure phases in the direction
in the ratio plots presented in this article where most points are clustered of the binary mixtures. It should be noted that these mixture lines are
between the pure phases (e.g. Fig. 6). An ideal situation would be to generally not well defined and lines from different mixtures can overlap.
have one point per particle, or per agglomerate of particles. However,
due to the large heterogeneities in cementitious materials, it is not
possible. In practice, values between 5000 and 20,000 segmented re­
gions were found to be a good compromise for standard maps.
A dataset made from the chemical composition at all of these points
can then be created. These points can also be filtered based on their BSE
values, for example to remove points corresponding to pores. The
composition of these points can then be analysed using the standard
ratio plots. An example of the points obtained with this method is pre­
sented in Fig. 6 for the OPC and the LC3 samples. The challenge is now to
understand how these points are related to the cement phases as iden­
tified by XRD or by thermodynamic modelling.

3.5. Characterisation of mixtures

Due to the interaction volume of the electron beam, the nano-

crystalline phases and the many solid solutions typical of cement
pastes [40], as well as the inherent low signal-to-noise ratio of EDS
hypermaps, most pixels usually have to be considered as mixtures of
phases rather than attributed to single pure phases. The main example Fig. 7. Ideal phases and average anhydrous phases in the Al/Ca vs. Si/Ca.
Examples of binary mixtures were created from triangular distributions around
corresponds to the C-S-H matrix, the glue of the cement paste which
the position of the pure phases. Real mixtures points would include noise and
incorporates many small particles (unhydrated clinker grains, quartz,
ternary mixtures.
limestone, AFms, Ettringite, SCMs, etc.). Therefore, an automatic

Fig. 6. Representative points extracted from the EDS hypermaps and plotted in the Si/Ca-Al/Ca ratio plot for the OPC and the LC3 samples. Taylor phases correspond
to the stoichiometries provided by Taylor [35], Table 1.3, p. 10.

8
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

For example, a mixture of C3A and C3S (in a polyphase clinker grain) One of the main consequences of the intermixing is that pure phases
can be found at the same position in this plot as a mixture of AFm and are hard to isolate. It is especially true for the C-S-H which is finely
outer C-S-H. Therefore, it is important to develop combinations of filters intermixed with other phases such as ettringite, hydrotalcite, or even
by using several representations. For example, these two points can be fine particles of metakaolin. As such, the properties of pure C-S-H are
separated by adding an additional condition on the BSE intensity, or on hard to obtain through this type of analysis [1]. To make this point
the sum of oxides to distinguish between anhydrous and hydrous phases. clearer in the rest of the paper, we call this fine mix of C-S-H and other
The crude representation of Fig. 7 is sufficient to understand Fig. 6. phases the C-S-H matrix. The properties of this matrix are an average of
However additional effects, such as random noise, background and the true C-S-H gel composition as well as a proportion of the other mixed
ternary mixtures must be considered to obtain the clouds of points of phases at a scale lower than the interaction volume of the electron beam.
Fig. 6. In Fig. 6 it can be observed that the density of points and the known

Fig. 8. Phase separation in 28 day cured samples of an OPC and a LC3 samples. a) and b) are the Al/Ca vs. Si/Ca ratio plots for the OPC and LC3 samples. c) and d) are
the phase masks overlay on top of their respective BSE micrograph. Taylor phases correspond to the stoichiometries provided by Taylor [35], Table 1.3, p. 10.

9
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

position of the main cement phases is sufficient to understand the main clusters as well as the AFm phase which is present in larger masses and,
mixture lines. By defining thresholds along these lines, it is possible to in higher amounts. These observations are consistent with our current
relate each point to a particular phase. The selection of points can be understanding of the LC3 microstructure.
defined by a series of inequalities (e.g. Si/Ca < 0.2 and Al/Ca < 0.2 for Although not all the points can be attributed to the phases listed, the
portlandite), or graphically by using the Glue interface [20,31], to draw main grains and regions are identified, as observed on the BSE overlay in
a surface around the region of interest (ROI). These inequalities are then Fig. 8c and d. An advantage, but also a limitation of this approach is that
used to identify the corresponding regions in the maps. Specific in­ not every pixel is identified to a phase. For example, the inner C-S-H
equalities for OPC-based cementitious systems are described in the next close to aluminate clinker phase is not detected as inner C-S-H because
section. its aluminate content is high due to the interaction volume of the elec­
tron beam. Instead, it is defined as the outer C-S-H matrix, which cor­
responds to a phase mixture. This effect is stronger in the LC3 sample due
3.6. Qualitative phase separation
to the intermixing with fine metakaolin grains. This can be observed in
Fig. 8, where the inner C-S-H matrix identified for the LC3 sample has
As explained in the previous section, it is possible to identify most of
higher Al/Ca ratio, and it is moved towards the metakaolin. The choice
the cement paste phases visible at the BSE micrograph scale using a
of the threshold is an advantage for chemical composition identification,
manual selection of points. This section applies this approach to the two
but an inconvenience for volume fraction quantification. Thus, the user
cement pastes as examples. The separation of phases is presented in
should choose between assigning all pixels to a phase, or separating
Fig. 8 for the OPC and the LC3 samples, in the Si/Ca and the Al/Ca ratio
phases with relatively pure compositions. Both extremes (very pure
plot as well as overlays of the phase masks on the BSE image.
phases, or all pixels identified) can be appropriate according to the
The separation was made according the mixtures identified in Fig. 7
problem at hand. This choice is intrinsic to the physical limitations of the
and guidelines detailed in Table 3. Table 3 summarizes the main in­
experiment. As such, is not unique to our method. It is present in all
equalities used to separate phases based on the Si/Ca, Al/Ca, BSE, Sum
methods, even if it is not recognized.
of OXides (SOX) and others. The BSE histogram is roughly separated into
Only qualitative comparisons have been made so far. The next sec­
two main peaks (clinker and hydrates). The actual values are dependent
tions will focus on extracting quantitative data and, on comparison with
on the sample, the microscope and the microscope operator. This table
other measurement methods.
does not provide strict guidelines, but it gives rules of thumb for
thresholds to be adapted to each sample, similar to the decision tree
4. Microstructure characterisation
presented by Bentz et al. [13].
The Al/Ca vs. Si/Ca plot is the most useful representation to separate
4.1. Chemistry of phases
phases, but it is not sufficient. As an example, the calcite and portlandite
were separated using the BSE vs. SOX plot: both phases are very similar
Once a phase has been identified, it is possible to study its compo­
in backscattered contrast and chemical ratios, but they exhibit different
sition. As an example, the composition of the AFm solid solution formed
absolute contents of calcium oxide (i.e., different SOX). Therefore, they
during chloride ingress was found to be a very good case study for the
can still be separated. This separation can be verified on the BSE image
edxia approach. The AFm phases are part of the layered double hy­
due to the morphology difference between calcite (rounded particles)
droxide family. Their interlayers contain anions that can be exchanged
and portlandite (elongated clusters).
with the pore solution. In particular, in contact with chloride, the sulfate
Fig. 8c and d demonstrates the main microstructural differences
and carbonate- AFms of the cement paste can transform to Friedel’s salt
between the OPC paste and the LC3 paste. The inner C-S-H is a main
[41,42]. The conversion to Friedel’s salt is only partial at the commonly
hydrate phase in the OPC paste with the portlandite linked by the outer
employed concentrations (e.g. 0.5 M NaCl) and it is important to mea­
C-S-H matrix. The AFm grains are small and they have precipitated in
sure the chloride stoichiometry in the solid solution to fully characterize
spaces left by dissolved clinker grains. The LC3 paste is more heteroge­
the chloride binding.
neous and compact at this scale. One visual difference is the dominance
A method based on SEM-EDS point analysis was recently developed
of the yellow and green hue compared to the blue and purple of the OPC
in Sui et al. [43]. It is based on extrapolating the AFm/C-S-H mixture
paste. This is due to the metakaolin (Al, Si) present in large and small
line to Al/Ca = 0.5 in the Cl/Ca vs. Al/Ca ratio plot, enabling the esti­
mation of the Cl/Ca ratio of the solid solution between Friedel’s salt and
Table 3
another AFm (Cl/Ca = 0, but Al/Ca = 0.5). Using the edxia approach
Example of boundary definition for common phases in cement pastes. ε repre­
instead of EDS point analysis for this task has two main advantages.
sents a tolerance dependent on the noise in the map, phases in the sample, and
the interaction volume of the electron beam.
First, it is useful to check that the selected points are indeed AFm based
on morphology. Secondly, the mixture line is fully captured, and a
Phase Si/Ca Al/Ca BSE SOX Others
higher precision can be obtained. Finally, the map further contains
peak
additional information, such as the chloride sorption on C-S-H. As an
C3S 0.33 ± ε Clinker 1±ε
<0.1
illustration, Fig. 9 shows the Cl/Ca vs. Al/Ca ratio plot for an LC3 sample
C2S 0.5 ± ε <0.1 Clinker 1±ε
C3A/C4AF <0.2 >0.5 Clinker 1±ε
cured 28 days before and after being exposed to a 0.5 M NaCl solution. In
AFm <0.2 0.5 ± ε Hydrate <0.8 S/Ca,Cl/Ca, … this case, the Cl/Ca ratio is around 0.24 ± 0.01 in the AFm phases after
Portlandite <0.2 <0.2 Hydrate 0.76 exposure, which is a little less than half the maximum binding capacity
±ε compared to formation of pure Friedel’s salt. Additional details on this
Calcite Hydrate 0.48
study is available in [44,45], such as the estimation of the solid solution
<0.2 <0.2
±ε
Strätlingite 0.5 ± ε 0.5 < x < Hydrate <0.8 by XRD which compare very well to our method.
1
Metakaolin >1.0 ± ε ∝ Si/Ca Hydrate 1±ε 4.2. Sub-resolution phases
Slags Variable Variable Hydrate 1±ε Mg can help [5]
Slags Variable Variable Hydrate <0.8 Mg can help [5]
hydrates Another added value of the full chemical analysis is the possibility to
Quartz >1.0 <0.2 Hydrate 1±ε Can filter on Si detect sub-resolution phases. Ettringite is one of the main phases in
Fly ash Variable Variable Hydrate 1±ε cement paste. However, it is not identified in Fig. 8. The reason for this is
C-S-H matrix Variable Variable Hydrate <0.7 Depends on that ettringite crystals are below the resolution of a typical hypermap, as
cement blend
ettringite commonly precipitates in needles of less than a hundred

10
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

Fig. 9. Composition of the Friedel’s salt-hemicarbonate solid solution as measured by SEM Image analysis. Example of a LC3 sample before and after exposure to a
0.5 M NaCl solution.

nanometers compared to the micrometer-scale interaction volume, and points are detected in the Al/Ca vs. Si/Ca ratio plot, as shown in Fig. 8.
about half a micrometer BSE resolution. Some clusters of ettringite could However, the presence of ettringite can still be detected through the
be identified on the BSE image by a trained expert, especially in the case EDS chemical composition. The process is presented in Fig. 10. First, the
of OPC. However, this does not tell us about the real distribution of outer C-S-H gel is selected as a reference in the Si/Ca vs. Al/Ca ratio plot.
ettringite. Sub-resolution refers here to the fact that no pure ettringite Then, in the S/Ca vs. Al/Ca, the Ettringite/C-S-H mixture line is

Fig. 10. Subresolution detection of ettringite in the OPC sample at 28 days.

11
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

identified. Points along that line, but outside the main C-S-H matrix measured area should be big enough to obtain a big enough average over
cluster of points are selected. The selected pixels do not correspond to the heterogeneous material. However the magnification should be big
pure ettringite pixels, but to mixtures of C-S-H and ettringite. This enough to be able to separate the hydrate phases of interest. The largest
process is of course highly dependent on the threshold chosen. However, source of heterogeneities in the cement paste is the large anhydrous
it helps to detect the presence of sub-resolution phases. In the case of grains. As seen in Fig. 8, the presence or the absence of a clinker grains
ettringite, this method can be used to study delayed ettringite formation would make a large difference on the volume fractions. One very time-
or sulfate attack to detect the precipitation patterns. Although fractions consuming way to solve this problem would be to acquire many maps at
cannot be computed accurately for these phases intimately intermixed, high magnification. A simple method is to choose a phase to normalise
point densities can be used to compare samples qualitatively. This the results to a known quantity, and thus to remove the influence of the
detection is of course dependent on the level of the intermixing. For larger grains. Thus, our approach is the following: (1) first, we separate
example, in the OPC sample, small ettringite grains can be observed as the big grains, and we compute the surface fractions in the remaining
depicted in Fig. 10. However, in the LC3 samples, the regions detected by area; (2) and then, a phase present in that volume is used to renormalise
this method are not easily identified using only the BSE, as seen in the data to include the big grains. This approach requires another
Fig. 11. Detection of mixed phases is a common use of ratio plots. The experiment to carry out the normalisation but it avoids having to take
added value of our framework is that their spatial distribution can also too many high resolution maps with a very high field-of-view.
be studied. The total volume fraction of small particles in the LC3 system is
defined as:

4.3. Quantitative phase fractions ϕsmall ≡ 1 − ϕclinker − ϕmetakaolin − ϕquartz − ϕpores

The partial volume fraction of phase X, with small characteristic size,

One of the simple microstructure descriptors that can be obtained
can then be computed by:
from a phase mask is the volume fraction. The use of image analysis to
quantify the volume fraction of anhydrous phases (and the reaction ϕsmall =
ϕX
degree) is well documented in the literature as early as 1986 [3,5,46].
X
ϕsmall
The volume fraction of hydrates is less studied. The main reason is the
The normalisation factor can be computed for an easily quantifiable
difficulty to separate the hydrates phases easily.
mineral Y by XRD or by TGA:
As an illustration of the potential of our framework, we propose a
way to quantify the volume and mass fraction of AFm phases in a LC3 ρY ϕsmall
Y
αY =
cement paste, which are generally considered difficult to quantify. These wY
phases are important as they fill the pore space in the microstructure,
and thus contributes to the compressive strength. They are also impor­ where ρY is the density of solid phase Y, and wY the mass fraction of this
tant for durability-related issues as they are the main phases able to bind phase as measured by another method.
chlorides [41]. The mass fraction of phase X is then given by:
A main challenge is to obtain a representative scale to analyse. The

Fig. 11. Subresolution detection of ettringite in the LC3 sample at 28 days.

12
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

ρX ϕsmall
x
4.4. Microstructure descriptors
wX =
αY
Describing the microstructure of random heterogeneous porous
In addition to the error of the estimation of the phase fractions from
materials is a challenging problem as only a statistical approach can be
the hypermaps, this method is also sensitive to the error made during the representative [48]. Our phase separation analysis provides binary
XRD analysis and the estimation of the density. In this example, we
masks representing each phase. From these masks, it is possible to
choose calcite as the normalisation phase, as it is present in large compute statistical representations of the microstructure. For example,
quantity and it can be measured by XRD or by TGA. Its density is also
the point-point correlation function, S2, [11,48], or the chord-length
well known. In order to obtain representative estimates of the volume distribution function [48] can be easily computed in our framework.
fractions, the analysis was done over fields of view larger than the
These functions provide information about the volume fraction, the
feature of interest (here ~300 μm by 200 μm) and with a sufficient particle size distribution, the particle shapes and their distribution in the
number of replicates to cover the local heterogeneity of the material
material. This feature is directly available in the graphical interface.
(here 13 hypermaps, set of parameters F in Table 2). Mass densities were As an example, Fig. 12 presents a comparison of the S2 correlation
obtained from Balonis and Glasser [47].
functions for the portlandite and the AFm phases in the OPC and the LC3
The mass fractions estimated by this method are presented in samples. S2(r) is the probability of finding a phase i at a distance r from a
Table 4. The mass fractions obtained for calcite and portlandite by XRD
random location in the sample. From binary phase masks, it is
and TGA agree very well. The mass fraction of portlandite found by straightforward to compute as described in [48]. By construction, S2(0)
image analysis after normalisation is also very close. Although por­
corresponds to the volume fraction (ϕ), and S2(∞) = S2(0)2 = ϕ2.
tlandite and calcite precipitate very close to each other, and their Therefore, to better identify the length scales, Fig. 12 shows S2(r) − S2(0)
boundaries are not well defined (see Fig. 8), it indicates that our phase as a function of r. In this view, S2(r) − S2(0) = 0 corresponds to an
separation using the sum of oxide works well. absence of spatial correlation, where the probability of finding the phase
Due to overlapping peaks, the AFm content cannot be measured is only determined by the volume fractions. From these plots, we can see
accurately from TGA, and only the XRD values can be used as compar­ that portlandite have similar characteristic lengths in both systems.
ison. The mass fractions obtained by image analysis are higher than the However, it’s volume fraction is lower in the LC3 sample due to the
XRD values. There are two possible explanations for this. Either the XRD pozzolanic reaction. On the other hand, the volume fraction of AFm in
underestimate the amount of AFm phases. The AFm peaks are difficult to the LC3 system is higher due to the synergetic reaction with limestone
fit accurately, and the choices for the Rietveld analysis are not obvious.
[22]. The grains of AFm are also significantly larger in the LC3 system.
In addition, some of the AFms might be poorly crystalline. On the other These microstructure descriptors (volume fraction, correlation
hand, the image analysis can also overestimate the AFm content. First,
functions, …) can be used to study the effect of the raw materials pro­
the AFm precipitates in large pores in what used to be clinker grains. cessing or curing conditions of the assemblage of phases at different ages
Therefore, they are surrounded by porosity which can be included in
to optimize the mix design for compressive strength and/or durability
part in our analysis. Another source of error is the internal porosity of [49].
these grains. Although these grains are smooth in the OPC case, they
have a texture in the LC3 case. The densities used are the solid densities
5. Going further
from the crystal structure. The density that should be used is the
apparent density, taking into account the internal porosity (due to
5.1. Profiles
cracked or non compact grains) and the mixed phases. However, this
value is not available. The impact of this density is important as high­
The field of view of a typical map is usually less than a couple of
lighted by the difference between assuming hemicarbonate or mono­
hundred micrometers. The typical scale of penetration profiles, relevant
carbonate as the AFm phase.
for are a few centimeters. Therefore, maps taken along the profile of a
As such, this quantification by image analysis is interesting for
sample can be considered homogeneous (with respect to the durability
several reasons: (1) it provides the apparent volume fraction directly,
issues). As a consequence, all the different analysis shown in the pre­
and not a mass fraction to be transformed into volume fraction (the
vious section can be done at different depths in a sample. It can be used
volume fraction is of interest for validating homogenization models to
to study the changes occurring during exposure to environment. For
predict the compressive strength); (2) it provides an additional mea­
examples, we can obtain the chemical composition of the AFm for a
surement to verify/confirm other methods; and (3) it enables estimation
cement paste sample subjected to an electro-migration test [50]. Specific
of volume fraction for amorphous phases that cannot be directly
details of the test run are available in [45]. The result of the analysis for
measured by XRD-Rietveld, or by TGA due to peak overlapping. If the
an OPC sample of 1 cm is presented in Fig. 13. Additional methods (e.g.,
approach is exemplified here for AFm phases, a similar protocol
microXRF) can be used to obtain the total chloride concentration.
considering the scale of particles could be adapted for other hydrates
Therefore, both the total amount, and the composition of the phases can
and anhydrous phases.
be obtained as a function of specimen depth. In addition, the method
used to obtain the quantitative volume fractions could be used in such a
profile to get both the composition and the amount of a specific phase.
This combination of information is necessary to correctly validate
Table 4 reactive transport models without overfitting parameters.
Mass and volume fraction obtained from 13 small hypermaps and a normal­
isation factor computed by XRD or by TGA. Uncertainties for the XRD corre­ 5.2. Using lower-efficiency EDS detectors
spond to the standard deviation on 5 scans using the same Rietveld fitting
process.
The maps presented so far were obtained with a high-end EDS de­
Solid phase (mass%) wXRD wTGA wIA
XRD wIA
TGA
ϕsmall
X tector. In practice, detectors with a lower efficiency might be the only
(mass%) (mass%) (mass%) (vol%) hardware available. By lower efficiency, we mean a detector with lower
Calcite 9.7 ± 0.2 9.9 9.7 9.9 10.9 detection surface, lower sensitivity, higher dead time or any other factor
Portlandite 3.2 ± 0.3 3.3 3.1 3.2 4.2 that requires a higher exposition time to obtain the same number of
AFm (as hemicarbo) 9.8 ± 0.5 12.4 12.7 19.1 counts. In this section, we discuss the influence of the detector to un­
AFm (as 9.8 ± 0.5 13.6 13.9 19.1 derstand to what extent information can be obtained from lower quality
monocarbonate)
maps. Fig. 15 compares different signals obtained on the same region of

13
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

Fig. 12. S2 correlation function for (a) the portlandite and (b) the AFm phase in a OPC and a LC3 samples.

Table 5
Peak signal-to-noise ratio (dB) for various maps of the LC3 sample according the
parameters.
Map Ca Al S

A 22.9 33.0 40.2

B 19.5 28.8 36.2
C 29.6 35.1 41.4
D 26.4 32.8 39.1
E 28.8 34.9 42.6

1 ∑M ∑ N
MSE = {I(i, j) − D(i , j)}2
MN i=1 j=1

The effect of the parameters is shown in Figs. 14 and 15. A main

feature is that the absolute values (%Ca, SOX) are a lot more sensitive
than the ratios (Ca/Si) to detector quality. While the absolute values
present a clear bias after filtering, the Ca/Si ratio is similar between map
A and map C. However, the ratio presents a slight bias for the noisiest set
Fig. 13. Chemical composition of the AFm in a OPC sample subjected to an of hypermaps, set B. This is the main a posteriori justification for the
electromigration test. selection of phases based on ratios in our general approach: even if the
maps are quite noisy, the ratios remain a reliable measure [36].
interest using two detectors and different instrument parameters as Nevertheless, the absolute values provide valuable information for
described in Table 2 and Section 2.2.2. The set of maps studied until now microstructure characterisation. For example, the sum of oxide is the
is the set C. Sets C, D and E have been obtained with the same high-end most reliable method to separate portlandite from limestone. Selecting
EDS detector, while sets A and B have been obtained with a detector low calcium phases (e.g., quartz) on absolute chemical values is also
with a lower surface of detection. Table 5 shows the peak signal-to-noise more reliable than selecting it on chemical ratios (e.g., Si/Ca) due to the
(PSNR) for these maps. The PSNR of a M × N EDS map is estimated from division by small values.
the map I, and the denoised map D. The PSNR is defined as:
( ) 5.3. Filtering at the spectrum level
1
PSNR = 10⋅log10
MSE
In this section, we present a method to improve the signal-to-noise
And the mean squared error (MSE) is defined as ratio. The main source of noise is the low electron counts during the
map quantification. The filtering process reduces the error, but it does

14
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

Fig. 14. Comparisons of the Ca, and the Ca/Si maps (after denoising) for the set B (high noise) and the set C (low noise) of hypermaps.

Fig. 15. Comparison of EDS hypermaps of the same region in the same sample with various signal-to-noise ratio. All curves were obtained from the filtered maps.
The red dashed curve is obtained from filtered spectrums (see Section 5.3). (For interpretation of the references to color in this figure legend, the reader is referred to
the web version of this article.)

not address the root cause of the problem. A better method would be to the pixels of a 5 × 5 square centered at the representative points was
filter the hyperspectral image directly. This is unfortunately often then summed. Then, the reconstructed spectrums were quantified using
restricted by the proprietary format used by the microscope manufac­ the same quantification algorithm as used for the hypermaps. The results
turer, and the difficulty of implementing the quantification algorithm are shown as the red dashed curve in Fig. 15. It can be observed that the
[51]. However, to demonstrate the potential of this method, we used a Ca/Si ratio is similar, but the bias on the Ca amount has been greatly
simpler yet effective approach. For the set A of hypermaps, the co­ corrected. The main limitation of such an approach here was the time
ordinates of representative points were first obtained. The spectrum of needed to quantify the spectrums as the analysis could not be fully

15
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

automated. Therefore, only one example is presented. The reconstructed composition. This information is necessary to correctly validate reactive
spectrums where created with our framework and hyperspy [52] to transport models of chloride ingress, as it allows fixing the chemistry
extract the data from the raw hypermap. With an automated analysis, it model. We also presented how to obtain information of the phases below
would be possible to only average spectrums in the regions found by the the resolution of the EDS analysis. Finally, the effect of noisy maps was
segmentation of the composite image, providing a means to obtain high- discussed. Due to our ratio plots approach, good phase separation can
quality quantitative data from lower-end detectors. still be obtained even if absolute values are not reliable. An additional
method to refine the signal-to-noise ratio is proposed.
5.4. Possible extension to machine learning Future work will focus on using the morphological information to
improve further the signal-to-noise ratio as well as extending the
To simplify the analysis, it is natural to think about using one of the quantitative analyses of the phase maps with additional application-
many machine learning algorithms developed to solve the classification specific analysis tools.
problem. Our method to extract the representative points can be used as
a pre-processing step for these algorithms. A manual classification CRediT authorship contribution statement
approach was preferred for the reasons highlighted in the introduction.
As outlined in the article, the challenge is even greater if we realize Fabien Georget: Conceptualization, Methodology, Software,
that mixtures with three or more phases can exist, and especially, mix­ Writing - original draft. William Wilson: Conceptualization, Method­
tures with phases that cannot be observed directly at the scale of the ology, Investigation, Writing - review & editing. Karen L. Scrivener:
SEM interaction volume (e.g. ettringite). In addition, the main phase of Writing - review & editing, Supervision, Funding acquisition.
cement paste, C-(A-)S-H, is very sensitive to the formulation. It is not
possible to define a boundary that will work for every formulation of Declaration of competing interest
cement paste. As a consequence, it means that a black-box classification
model would have to be re-calibrated for every formulation, curing The authors declare that they have no known competing financial
conditions, and even, age of the sample. However, even if the chemistry interests or personal relationships that could have appeared to influence
of cement paste is intrinsically complicated, it is now possible to get the work reported in this paper.
reasonable predictions. Therefore, by leveraging domain knowledge, it
is possible to obtain an a-priori distribution of phases. As emphasized by Acknowledgments
Munch et al. [16], this domain knowledge is important to obtain good
results with machine learning approaches. Rather than adapting The authors would like to thank the members of the Laboratory of
machine-learning to include domain knowledge, we chose to by-pass it Construction Materials who provided raw data and feedbacks during the
to only use domain knowledge. The results presented in this paper development phase of this project.
should highlight the practicality of our choice.
In the opinion of the authors, the next step to improve machine Appendix A
learning methods would be to explicitly handle mixtures and to use the
morphology information of the BSE image. For example, particle A.1. Code download and installation
detection might help define clearer phase boundaries. It is relatively
easy for large and clear particles such as the clinker grains, but it is a The code is available in the python package index (https://pypi.
much more complex challenge for the small hydrate particles. org/project/edxia/), the Anaconda distribution (https://anaconda.
Improving the phase separation is a worthy research topic. As an open- org/specmicp/edxia), zenodo [30], or openly available git repository
source Python framework, our tool is a good start as it can be easily (https://bitbucket.org/specmicp/edxia/src/master/). Installation in­
extended with well-established machine learning framework such as structions and documentations is available with the code, online
scikit-learn [53]. (https://edxia.readthedocs.io/en/latest/) or in [31] for the user friendly
interface. The main datasets (set C in Table 2) analysed in this study are
6. Concluding remarks available [54] to test the interface.

A new framework to analyse quantified SEM-EDS hypermaps of References

cementitious materials is presented in this paper. Using superpixel
subsampling of a composite map, representative points are extracted [1] J. Rossen, K. Scrivener, Optimization of sem-eds to determine the c–a–s–h
from quantified chemical maps. Ratio plots representations are then composition in matured cement paste samples, Mater. Charact. 123 (2017)
294–306, https://doi.org/10.1016/j.matchar.2016.11.041.
used to manually classify the points and create phase masks. The manual [2] S. Meulenzyer, J. Chanussot, J.J. Chen, S. Crombez, Spectral-spatial image
approach can be used to adapt the classification to the scientific problem processing strategies for classifying multispectral sem-eds x-ray maps of
at hand. Pure chemical composition, mixture identification or quanti­ cementitious materials, in: 14th Euroseminar on Microscopy Applied to Building
Materials, 2013 (EMABM), 2013.
tative volume fractions can all be obtained. The framework is flexible [3] P.T. Durdziński, C.F. Dunant, M.B. Haha, K.L. Scrivener, A new quantification
and user-friendly, as it is implemented as a plugin to Glue, a multi- method based on sem-eds to assess fly ash composition and study the reaction of its
dimensional, linked data exploration graphical interface [20,21]. This individual components in hydrating cement paste, Cem. Concr. Res. 73 (2015)
111–122, https://doi.org/10.1016/j.cemconres.2015.02.008.
interface can be used by the user to select phases in the ratio plots and to [4] X. Feng, E. Garboczi, D. Bentz, P. Stutzman, T. Mason, Estimation of the degree of
observe directly the corresponding phase masks on the BSE image. hydration of blended cement pastes by a scanning electron microscope point-
Quantitative analysis tools are also implemented in the interface. counting procedure, Cem. Concr. Res. 34 (10) (2004) 1787–1793, https://doi.org/
10.1016/j.cemconres.2004.01.014.
Furthermore, additional analysis can be easily added to our open source [5] V. Kocaba, E. Gallucci, K.L. Scrivener, Methods for determination of degree of
Python framework. In particular, other approaches could be added to reaction of slag in blended cement pastes, Cem. Concr. Res. 42 (3) (2012) 511–525,
refine the phase separation. https://doi.org/10.1016/j.cemconres.2011.11.010.
[6] K.J. Krakowiak, W. Wilson, S. James, S. Musso, F.-J. Ulm, Inference of the phase-to-
Several demonstrations of the abilities of our framework are pre­
mechanical property link via coupled x-ray spectrometry and indentation analysis:
sented. In particular, we analysed the AFm phases. We obtain their application to cement-based materials, Cem. Concr. Res. 67 (2015) 271–285,
composition before and after exposure to chloride. We also measured https://doi.org/10.1016/j.cemconres.2014.09.001.
their volume, and mass fractions, as well as their point to point corre­ [7] W. Wilson, L. Sorelli, A. Tagnit-Hamou, Automated coupling of nanoindentation
and quantitative energy-dispersive spectroscopy (ni-qeds): a comprehensive
lation functions. This analysis can also be done along the depth of a method to disclose the micro-chemo-mechanical properties of cement pastes, Cem.
sample to obtain profiles of phase content and/or phase chemical Concr. Res. 103 (2018) 49–65, https://doi.org/10.1016/j.cemconres.2017.08.016.

16
F. Georget et al. Cement and Concrete Research 141 (2021) 106327

[8] D. P. Bentz, P. E. Stutzman, SEM Analysis and Computer Modelling of Hydration of [29] S. van der Walt, J. L. Schönberger, J. Nunez-Iglesias, F. Boulogne, J. D. Warner, N.
Portland Cement Particles Petrography of Cementitious Materials, ASTM Int. doi: Yager, E. Gouillart, T. Yu, the scikit-image contributors, scikit-image: image
https://doi.org/10.1520/STP12645S. processing in Python, PeerJ 2 (2014) e453. doi:https://doi.org/10.7717/peerj.453.
[9] U.H. Jakobsen, K.D. Weerdt, M.R. Geiker, Elemental zonation in marine concrete, [30] F. Georget, W. Wilson, edxia (Jan. 2020). doi:https://doi.org/10.5281/zenodo.
Cem. Concr. Res. 85 (2016) 12–27, https://doi.org/10.1016/j. 3492220.
cemconres.2016.02.006. [31] F. Georget, W. Wilson, K. Scrivener, edxia through the glueviz interface (Feb.
[10] F. Georget, W. Soja, K.L. Scrivener, Characteristic lengths of the carbonation front 2020). doi:https://doi.org/10.5281/zenodo.3686068.
in naturally carbonated cement pastes: implications for reactive transport models, [32] A. Chambolle, An algorithm for total variation minimization and applications,
Cem. Concr. Res. 134 (2020) 106080, https://doi.org/10.1016/j. J. Math. Imaging Vis. 20 (1) (2004) 89–97, https://doi.org/10.1023/B:
cemconres.2020.106080. JMIV.0000011325.36760.1e.
[11] P. E. Stutzman, J. W. Bullard, P. Feng, Quantitative imaging of clinker and cement [33] E. Eisemann, F. Durand, Flash photography enhancement via intrinsic relighting,
microstructure, Tech. Rep. 1877 (April 2015). doi:https://doi.org/10.6028/NIST. ACM Trans. Graph. 23 (3) (2004) 673–678, https://doi.org/10.1145/
TN.1877. 1015706.1015778.
[12] S. Meulenzyer, B. Eric, L. Didier, R. Barbarulo, Sem supervised image analysis [34] G. Petschnigg, R. Szeliski, M. Agrawala, M. Cohen, H. Hoppe, K. Toyama, Digital
contribution on understanding sulfate attack on blended cement mortars, in: 15th photography with flash and no-flash image pairs, ACM Trans. Graph. 23 (3) (2004)
Euroseminar on Microscopy Applied to Building Materials, 2015 (EMABM), 2015. 664–672, https://doi.org/10.1145/1015706.1015777.
[13] D. P. Bentz, P. E. Stutzman, C. J. Haecker, S. Remond, Sem/x-ray imaging of [35] H. Taylor, Cement Chemistry, Thomas Telford, 1997.
cement-based materials, Proceedings of the 7th Euroseminar on Microscopy [36] Harrisson A. M., Winter N. B., Taylor H. F. W., X-ray microanalysis of microporous
Applied to building Materials (1999) 457–466. materials, Journal of Materials Science Letters 6 (11) (1987) 1339–1340. doi:https
[14] Q. Ding, M. Colpan, Decision tree induction on hyper-spectral cement images, Int. ://doi.org/10.1007/BF01794611.
J. Comput. Intell. 2 (3) (2006). [37] W. Wilson, J. Rivera-Torres, L. Sorelli, A. Durán-Herrera, A. Tagnit-Hamou, The
[15] L. Drumetz, M.D. Mura, S. Meulenyzer, S. Lombard, J. Chanussot, Semiautomatic micromechanical signature of high-volume natural pozzolan concrete by combined
classification of cementitious materials using scanning electron microscope images, statistical nanoindentation and SEM-EDS analyses, Cem. Concr. Res. 91 (2017)
J. Electron. Imaging 24 (6) (2015) 1–19, https://doi.org/10.1117/1. 1–12, https://doi.org/10.1016/j.cemconres.2016.10.004.
JEI.24.6.061109. [38] D. Damidot, B. Lothenbach, D. Herfort, F. Glasser, Thermodynamics and cement
[16] B. Münch, L. Martin, A. Leeman, Segmentation of elemental eds maps by means of science, Cement and Concrete Research 41 (7) (2011) 679–695, special Issue: 13th
multiple clustering combined with phase identification, J. Microsc. 260 (3) (2015) International Congress on the Chemistry of Cement. doi:https://doi.org/10.1016/j.
411–426, https://doi.org/10.1111/jmi.12309. cemconres.2011.03.018.
[17] W. Rasband, Imagej, https://imagej.nih.gov/ij/ (1997–2018). [39] R. Achanta, A. Shaji, K. Smith, A. Lucchi, P. Fua, S. Süsstrunk, Slic superpixels
[18] A. Schneider Caroline, S. Rasband Wayne, W. Eliceiri Kevin, Nih image to imagej: compared to state-of-the-art superpixel methods, IEEE Trans. Pattern Anal. Mach.
25 years of image analysis, Nat. Methods 9 (7) (2012) 671–675, https://doi.org/ Intell. 34 (11) (2012) 2274–2282, https://doi.org/10.1109/TPAMI.2012.120.
10.1038/nmeth.2089. [40] B. Lothenbach, D.A. Kulik, T. Matschei, M. Balonis, L. Baquerizo, B. Dilnesa, G.
[19] K. Scrivener, R. Snellings, B. Lothenbach, A Practical Guide to Microstructural D. Miron, R.J. Myers, Cemdata18: a chemical thermodynamic database for
Analysis of Cementitious Materials, Taylor & Francis, 2016. hydrated portland cements and alkali-activated materials, Cem. Concr. Res. 115
[20] C. Beaumont, A. Goodman, P. Greenfield, Hackable user interfaces in astronomy (2019) 472–506, https://doi.org/10.1016/j.cemconres.2018.04.018.
with glue, in: A. R. Taylor, E. Rosolowsky (Eds.), Astronomical Data Analysis [41] M. Balonis, B. Lothenbach, G.L. Saout, F.P. Glasser, Impact of chloride on the
Software an Systems XXIV (ADASS XXIV), Vol. 495 of Astronomical Society of the mineralogy of hydrated portland cement systems, Cem. Concr. Res. 40 (7) (2010)
Pacific Conference Series, 2015, p. 101. 1009–1022, https://doi.org/10.1016/j.cemconres.2010.03.002.
[21] T. Robitaille, C. Beaumont, P. Qian, M. Borkin, A. Goodman, glueviz v0.13.1: [42] A. Mesbah, C. Cau-dit Coumes, F. Frizon, F. Leroux, J. Ravaux, G. Renaudin, A new
multidimensional data exploration (Feb. 2017). doi:https://doi.org/10. investigation of the Cl-CO3 substitution in afm phases, J. Am. Ceram. Soc. 94 (6)
5281/zenodo.1237692. (2011) 1901–1910, https://doi.org/10.1111/j.1551-2916.2010.04305.x.
[22] M. Antoni, J. Rossen, F. Martirena, K. Scrivener, Cement substitution by a [43] S. Sui, W. Wilson, F. Georget, H. Maraghechi, H. Kazemi-Kamyab, W. Sun,
combination of metakaolin and limestone, Cem. Concr. Res. 42 (12) (2012) K. Scrivener, Quantification methods for chloride binding in portland cement and
1579–1589, https://doi.org/10.1016/j.cemconres.2012.09.006. limestone systems, Cem. Concr. Res. 125 (2019) 105864, https://doi.org/10.1016/
[23] D. Jansen, F. Goetz-Neunhoeffer, C. Stabler, J. Neubauer, A remastered external j.cemconres.2019.105864.
standard method applied to the quantification of early opc hydration, Cem. Concr. [44] Advanced Characterization of Chloride Binding in OPC and LC3 Pastes, TU Delft,
Res. 41 (6) (2011) 602–608, https://doi.org/10.1016/j.cemconres.2011.03.004. 2020.
[24] X. Li, R. Snellings, K.L. Scrivener, Quantification of amorphous siliceous fly ash in [45] W. Wilson, F. Georget, K. Scrivener, The mini-migration approach: a
hydrated blended cement pastes by x-ray powder diffraction, J. Appl. Crystallogr. comprehensive characterization of accelerated chloride ingress at the scale of
52 (6) (2019) 1358–1370, https://doi.org/10.1107/S1600576719013955. blended-cement pastes, (in preparation).
[25] T.E. Oliphant, Python for scientific computing, Comput. Sci. Eng. 9 (3) (2007) [46] K.L. Scrivener, H.H. Patel, P.L. Pratt, L.J. Parrott, Analysis of phases in cement
10–20, https://doi.org/10.1109/MCSE.2007.58. paste using backscattered electron images, methanol adsorption and
[26] E. Jones, T. Oliphant, P. Peterson, et al., SciPy: open source scientific tools for thermogravimetric analysis, MRS Proc. 85 (1986) 67, https://doi.org/10.1557/
Python, [online; accessed 11-06-2019] (2001–). URL http://www.scipy.org/. PROC-85-67.
[27] Virtanen Pauli, Gommers Ralf, Oliphant Travis E., Haberland Matt, Reddy Tyler, [47] M. Balonis, F. Glasser, The density of cement phases, Cem. Concr. Res. 39 (9)
Cournapeau David, Burovski Evgeni, Peterson Pearu, Weckesser Warren, Bright (2009) 733–739, https://doi.org/10.1016/j.cemconres.2009.06.005.
Jonathan, W. S. van der J., Brett Matthew, Wilson Joshua, Millman K. Jarrod, [48] S. Torquato, Random Heterogeneous Materials: Microstructure and Macroscopic
Mayorov Nikolay, Nelson Andrew R. J., Jones Eric, Kern Robert, Larson Eric, Carey Properties, Springer, New York, 2002.
C. J., Polat İlhan, Feng Yu, Moore Eric W., VanderPlas Jake, Laxalde Denis, [49] F. Zunino, K. Scrivener, Assessing the effect of alkanolamine grinding aids in
Perktold Josef, Cimrman Robert, Henriksen Ian, Quintero E. A., Harris Charles R., limestone calcined clay cements hydration, Constr. Build. Mater. (2020), 121293,
Archibald Anne M., Ribeiro Antônio H., Pedregosa Fabian, M. van Paul, https://doi.org/10.1016/j.conbuildmat.2020.121293.
Vijaykumar Aditya, Bardelli Alessandro Pietro, Rothberg Alex, Hilboll Andreas, [50] E. Samson, J. Marchand, K.A. Snyder, Calculation of ionic diffusion coefficients on
Kloeckner Andreas, Scopatz Anthony, Lee Antony, Rokem Ariel, Woods C. Nathan, the basis of migration test results, Mater. Struct. 36 (3) (2003) 156–165, https://
Fulton Chad, Masson Charles, Häggström Christian, Fitzgerald Clark, Nicholson doi.org/10.1007/BF02479554.
David A., Hagen David R., Pasechnik Dmitrii V., Olivetti Emanuele, Martin Eric, [51] G.F. Bastin, H.J.M. Heijligers, Quantitative electron probe microanalysis of
Wieser Eric, Silva Fabrice, Lenders Felix, Wilhelm Florian, Young G., Price Gavin ultralight elements (boron-oxygen), Scanning 12 (4) (1990) 225–236, https://doi.
A., Ingold Gert-Ludwig, Allen Gregory E., Lee Gregory R., Audren Hervé, Probst org/10.1002/sca.4950120408.
Irvin, Dietrich Jörg P., Silterra Jacob, Webber James T., Slavič Janko, Nothman [52] F. de la Peña, E. Prestat, V. T. Fauske, P. Burdet, P. Jokubauskas, M. Nord, T.
Joel, Buchner Johannes, Kulick Johannes, Schönberger Johannes L., de Vinicius, Ostasevicius, K. E. MacArthur, M. Sarahan, D. N. Johnstone, J. Taillon, J.
Miranda Cardoso José, Reimer Joscha, Harrington Joseph, Rodriguez Juan Luis Lähnemann, V. Migunov, A. Eljarrat, J. Caron, T. Aarholt, S. Mazzucco, M. Walls,
Cano, Nunez-Iglesias Juan, Kuczynski Justin, Tritz Kevin, Thoma Martin, Newville T. Slater, F. Winkler, pquinn dls, B. Martineau, G. Donval, R. McLeod, E. R.
Matthew, Kümmerer Matthias, Bolingbroke Maximilian, Tartre Michael, Pak Hoglund, I. Alxneit, D. Lundeby, T. Henninen, L. F. Zagonel, A. Garmannslund,
Mikhail, Smith Nathaniel J., Nowaczyk Nikolai, Shebanov Nikolay, Pavlyk hyperspy/hyperspy: Hyperspy v1.5.2 (Sep. 2019). doi:https://doi.org/10.
Oleksandr, Brodtkorb Per A., Lee Perry, McGibbon Robert T., Feldbauer Roman, 5281/zenodo.3396791.
Lewis Sam, Tygier Sam, Sievert Scott, Vigna Sebastiano, Peterson Stefan, More [53] F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel,
Surhud, Pudlik Tadeusz, Oshima Takuya, Pingel Thomas J., Robitaille Thomas P., M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos,
Spura Thomas, Jones Thouis R., Cera Tim, Leslie Tim, Zito Tiziano, Krauss Tom, D. Cournapeau, M. Brucher, M. Perrot, E. Duchesnay, Scikit-learn: machine
Upadhyay Utkarsh, Halchenko Yaroslav O., Vázquez-Baeza Yoshiki, S. 1.0 learning in Python, J. Mach. Learn. Res. 12 (2011) 2825–2830.
Contributors, Scipy 1.0: fundamental algorithms for scientific computing in [54] F. Georget, W. Wilson, K. Scrivener, edxia examples (May 2020). doi:https://doi.
python, Nature Methods (2020). doi:https://doi.org/10.1038/s41592-019-0686-2. org/10.5281/zenodo.3862958.
[28] J.D. Hunter, Matplotlib: a 2d graphics environment, Comput. Sci. Eng. 9 (3) (2007)
90–95, https://doi.org/10.1109/MCSE.2007.55.

17