Powder Technology 320 (2017) 161–178

Contents lists available at ScienceDirect

Powder Technology

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

Quantity and shape modification for random-fractal-based 3D

concrete meso-simulation
Xu Yang a, Fenglai Wang a,b,c,⁎, Xincong Yang a, Fei Zhu a, Bin Chi a
a
School of Civil Engineering, Harbin Institute of Technology, Harbin 150090, China
b
Key Lab of Structures Dynamic Behavior and Control of the Ministry of Education, Harbin Institute of Technology, Harbin 150090, China
c
Key Lab of Smart Prevention and Mitigation of Civil Engineering Disasters of the Ministry of Industry and Information Technology, Harbin Institute of Technology, Harbin 150090, China

a r t i c l e i n f o a b s t r a c t

Article history: The importance of the meso-scale analysis of concrete has been a controversial topic. The accuracy and efficiency
Received 16 February 2017 of numerical modelling directly influence the results and computational cost. In this study, we propose a versatile
Received in revised form 30 June 2017 fractal-geometry-based modelling method that can generate an aggregate model with a known gradation curve.
Accepted 7 July 2017
Two numerical meso-model groups generated by the random fractal method are measured and analysed
Available online 13 July 2017
considering the gradation curve, coarse aggregate content, and fractal dimension as evaluation indices. It is
Keywords:
verified that the evaluation indices of the numerical models are consistent with the actual ones. Based on
Random fractal method sampling statistical analysis, the geometrical homogenization of the meso-model generated by the random
Aggregate fractal method is found to be satisfactory for meso-scale applications. Furthermore, the random fractal method
Meso-simulation is shown to have low algorithm complexity.
Algorithm complexity © 2017 Elsevier B.V. All rights reserved.

1. Introduction experience, creation of the numerical model is often a time-consuming

Understanding the mechanical behaviour of concrete directly
influences construction cost and building safety, which are of great
significance.
Concrete meso-mechanics can explain the macro-scale mechanical
behaviour of concrete. Therefore, it has become the primary focus area
in the field of concrete research, yielding important results [1]. There
are two main research approaches within the framework of this theory:
experimentation and numerical simulation [2]. Compared with experi-
mentation, numerical simulation offers the advantages of low cost, con-
venient, and accurate data collection, and freedom from experimental
restrictions [2]; hence, it has attracted considerable attention.
The numerical meso-model of concrete is a complex system. In
general, analysis results of meso-level concrete simulation are
influenced by many factors. Herein, we consider two major influencing
factors: the geometry and material properties of each component. In
this paper, the geometry of concrete and its simulation are discussed
in detail.
A numerical simulation usually involves three major steps:
(i) creating the aggregate and cement models, (ii) meshing, and
(iii) finite element analysis/discrete element analysis [2]. The quality
of the concrete mechanical solution depends on the modelling accuracy,
i.e., on the first step of the simulation. However, depending on the user's
⁎ Corresponding author at: School of Civil Engineering, Harbin Institute of Technology,
Harbin 150090, China.
E-mail address: fl-wang@hit.edu.cn (F. Wang).
task [3]. Therefore, it is important to study and discuss the modelling
method.
Many researchers have introduced a number of ways to generate the
concrete geometrical model. The modelling method typically includes
two key procedures. The first procedure is generation of the aggregates,
which includes polyhedron generation using either standard Voronoï/
Delaunay tessellation [4–7] or a random spatial polygon growth [8,9].
This procedure should follow the principle of keeping each aggregate
convex, which is rather difficult. The second procedure is allocation of
the aggregate models in the concrete model, which includes the 3D
random particle allocation method introduced by Leite [10], the layering
disposition method for random aggregate arrangement introduced by
Tang [11], and the grid search method based on the background grid,
which is combined with meshing [3,9]. This procedure should follow
the principles of making the generated concrete numerical model fit
actual situations and avoiding the overlapping of aggregates.
The two key procedures described above are combined to obtain a
complete concrete modelling algorithm. The main aspects of the
modelling algorithm are computational efficiency and model quality.
This paper classifies modelling algorithms into three categories. In the
first category, random aggregate models are first generated and then
allocated in the concrete model sequentially, e.g., the method intro-
duced in [10], which is a traditional method. Based on this method,
it is necessary to judge whether the current particle is completely
inside the specimen and overlaps previously placed ones before each
particle allocation. The complexity of this method is relatively high;
the computation time increases with the number of allocated

http://dx.doi.org/10.1016/j.powtec.2017.07.031
0032-5910/© 2017 Elsevier B.V. All rights reserved.
162 X. Yang et al. / Powder Technology 320 (2017) 161–178
aggregates. Algorithms derived from this method include those intro-
duced in [11,12]. In these methods, the judgment of whether the
current particle overlaps previous ones is omitted; this improves the
computational efficiency considerably but increases the space between
aggregates. The first category profoundly influenced concrete meso-
modelling methods; for instance, the NIST model was developed to
solve the issue of hydration under seals in [13] and the HCSS to simulate
percolation [14]. In the second category, the aggregate models are
generated and allocated simultaneously, e.g., the method introduced
in [9]. Furthermore, some methods combined with meshing do not
require the overlapping of aggregates to be judged, thereby saving a
significant amount of time and facilitating follow-up work. Qin [3]
showed that it is easy to generate irregularly shaped aggregate models
using this type of algorithm, and there is usually a difference between
the gradations of the numerical model and actual conditions. In the
third category, the location and boundaries of the aggregates are first
determined and the aggregates are then generated, e.g., the method
introduced in [8] and an algorithm based on random fractal theory
introduced by us in [17]. According to the random fractal method
(RFM), the boundaries of aggregates around an aggregate can be
defined simultaneously with the generation of this aggregate. As with
the second category of generating algorithms described above, these
algorithms save time but provide a lower coarse aggregate content.
Each of the three types of algorithms described above has its
advantages and disadvantages. A major advantage of the RFM is that it
involves the application of an index (i.e., the fractal dimension of
concrete) that can describe the geometrical character of concrete quan-
titatively. Such quantitative characterization is useful for extending this
algorithm, as it can provide a theoretical basis for follow-up meso-
mechanical studies. Therefore, in this paper, we present the theory
and implementation of quantity and shape modification for three-
dimensional (3D) concrete meso-simulation based on the RFM. With
the objective of ensuring low algorithm complexity, the region of
applicability is extended and an approach for simulating graded aggre-
gates is proposed.
2. Method
2.1. Principal hypotheses and basic algorithm
2.1.1. Principal hypotheses and assumptions
To establish a systematic and complete theory system, and to
simplify the simulation methods, three basic hypotheses are introduced
in this paper.
Hypothesis 1. The set of aggregates and cement is a fractal.
Hypothesis 2. All the simulated aggregates are convex.
Hypothesis 3. There is always a space between any two aggregates.
Further details with respect to these hypotheses and some
corollaries are presented below.
Based on Hypothesis 1, the theories within the theoretical frame-
work of fractal geometry can be applied. Some conclusions concerned
with this study are presented in a later section. The possible availability
of the fractal effect of aggregates and cement enables us to use a
mathematical expression with the fractal dimension to describe the
spatial geometry property of the aggregate condition. Furthermore,
this hypothesis makes it possible to generate the aggregate simulation
model using the fractal method.
Hypothesis 2 helps to accomplish the following two purposes. (i) It
limits the change in the aggregate shape in the simulation process and
simplifies the simulation process. (ii) It facilitates element meshing
in the finite element analysis stage, thereby preventing abnormal
element generation as well as stress concentration. Hypothesis 3 pre-
vents the occurrence of aggregate overlap and reduces the amount of
computation. This rule is widely used in the meso-simulation of con-
crete. The three above-mentioned hypotheses influence the methods
of theory and simulation in later sections.
2.1.2. Basic generation algorithm
This corollary, obtained from Hypothesis 1, provides the principal
rationale for generating the concrete numerical model using RFM.
Corollary 1. One approach to fractal geometry is to regard two sets as ‘the
same’ if there is a bi-Lipschitz mapping between them.
In the theory of fractal geometry, the fractal dimension, which is
invariant under bi-Lipschitz transformations, is set up to distinguish
between sets that are not homeomorphic. Therefore, if two sets have
the same fractal dimension, they are considered to be the same [15].
Based on this corollary, the RFM can be expressed essentially as
follows: a geometrical pattern to be generated with the same fractal
dimension as that of actual concrete, using the iteration-based model-
ling algorithm, can be considered as the aggregate numerical model.
The fractal dimension of gradation can be estimated using the
definition of Hausdorff dimensions, which we briefly introduce as
follows. Considering the ratio of each sieving size xi to the maximum
aggregate size xmax as the x-coordinate and the value of the mass
retained P(xi) corresponding to the size of each xi as the y-coordinate,
several points can be plotted in a log-log plot. Then, the slope of the
straight line fitted to these points can be determined as 3 minus the
fractal dimension [17]. The method for ensuring the correspondence
between the fractal dimension of the numerical model and that of actual
concrete is detailed in Section 2.4.1.
The basic iterate algorithm for generating the fractal pattern can be
represented as follows [17]:
1. Set a cell: consider an initial model as a cell. According to the specific
purpose, the cell can be a two-dimensional (2D) or 3D pattern. A cell
is considered as an auxiliary pattern. An auxiliary pattern is used for
convenience and is virtually non-existent. It is drawn temporarily to
facilitate the construction of the model; it neither is shown after the
process of modelling nor participates in the calculation.
2. Divide the cell: divide the cell into several sub-auxiliary patterns in
accordance with a certain rule, ensuring that these sub-auxiliary
patterns are closed and convex.
3. Generate the aggregate(s): change the sub-auxiliary graph(s) at the
specified location to a real graph; the real graph is treated as the
model of an aggregate.
4. Iteration: treat each remaining sub-auxiliary pattern as a new cell,
and repeat steps 2 and 3 until the program ends.
In this study, a single performance of steps 2 and 3 is considered as
one iteration. A 2D cell, for instance, is treated by the process shown
in Fig. 1, in which a cell is closed and convex. The process shown in
Fig. 1 is a mathematical and qualitative representation that describes
how rock be crushed into aggregate productions. Hence, a possible
application of fractal theory may be the simulation of aggregate
production.
When the concrete is simulated, a fractal pattern in which a new
segment is scaled by a constant ratio is generated. This characteristic
resembles the characteristics of a series of sieve sizes that can be
approximated as a geometric progression with a common ratio.
However, in practice, a single cell cannot meet the demands of a variety
of model sizes for two reasons. First, the actual number of aggregates in
each sieving level cannot be an exponential function that depends on
the sieving size. Second, the maximum grain size in the model
generated by the basic algorithm is the edge length of a cell divided by
a fixed value.
The first problem can be solved by first calculating the probability
that the aggregate appears in the ideal location and then modifying
the quantity of aggregates, as described in Section 2.3. The second
 X. Yang et al. / Powder Technology 320 (2017) 161–178 163

Fig. 1. Schematic diagram of basic generation algorithm for generating a 2D pattern.

problem can be solved by setting an array of cells, and the group of
aggregates generated in all the cells is considered as the meso-model,
as described in Section 2.4.1.
2.2. Shape modification
This sub-section introduces different models generated by the basic
generation algorithm in practical applications. Shape modification of
the aggregate model is implemented by choosing a suitable modelling
method.
than or equal to B3. 4. Treat each remaining sub-cube as a new cell, and
repeat steps 2 and 3 until the program ends.
All the other models are improved and modified on the basis of
3D-TYPE-1; for example, Random-Cuboid-Model is generated in this
manner. A cubic cell is randomly divided into B parts so that B3 cuboids
of varying sizes are generated; then, the generation and iteration are
performed.
Sphere-Model-1 can be inscribed in the cube model. All types of
model generation methods based on the RFM are shown in Fig. 2. The
generation method of modified models is introduced in the next sub-
section.

2.2.1. Ideal model 2.2.2. Modified model

The ideal model shape, which does not coincide with the actual
condition, is introduced here to propose the concept for generating
the simulation as well as to test the algorithm. The ideal model can be
improved and modified to obtain the modified model, which would
be directly applicable.
In the case of the most basic model, namely 3D-TYPE-1, the genera-
tion method is as follows. 1. Set a cube with side S as a cell. 2. Divide each
side into B equal parts so that a cell would be divided into B3 equal sub-
cubes. 3. Change M sub-cubes into a real pattern, where M must be less
The shape of the ideal aggregate model should be modified to be
closer to the actual condition. Such shape modification can be imple-
mented in the following three ways.
The first way (OP1) is to change the manner of dividing the cell.
Random-Cuboid-Model is generated in this way. In the case of 3D-
TYPE-2, an arbitrary convex hexahedron can be set as a cell, which is
divided into B3 sub-convex hexahedrons of varying sizes; then, the
generation and iteration are performed. Fig. 3 shows the procedure of
generation with B = 3 and M= 1.

Fig. 2. All the generation methods in the framework of RFM.

164 X. Yang et al. / Powder Technology 320 (2017) 161–178

Fig. 3. Modified model (3D-TYPE-2) generated by changing cell division manner.

The second way (OP2) is to first generate a model that is treated as a
boundary and then generate a new inscribed model as the aggregate.
For example, Sphere-Model-1 introduced in the previous sub-section
is generated in this manner. Ellipsoid-Model and Sphere-Model-2
are generated by inscription in Random-Cuboid-Model. Convex-
Polyhedron-1 and Convex-Polyhedron-2 are inscribed in Sphere-
Model-1 and Sphere-Model-2, respectively (Fig. 4).
The third way (OP3) is to first generate a polyhedron model and
then cut one or more corners of this model to generate a modified
aggregate model. Convex-Polyhedron-3 is generated in this way on
the basis of 3D-TYPE-2.
After modification, these models reflect actual conditions. Ellipsoid-
Model and Sphere-Model-2 are recommended for the simulation of
cobbles, which are rocks with a particle size of 64–256 mm that are
smooth – often ellipsoid or spherical – owing to water transport. The
Convex-Polyhedron-1, 2, 3 and 3D-TYPE-2 models are recommended
for the simulation of crushed stone, which is an artificial aggregate
produced by crushers and has approximately polyhedral shapes.
Within the framework of RFM, the generation method could keep
each aggregate model convex. We present detailed descriptions of the
aggregate model generation algorithm in Appendix A, and discuss the
differences between some similar models.
2.3. Quantity modification
2.3.1. Aggregate number calculation
In general, the gradation of aggregates applied in engineering
follows the Fuller distribution. Based on the theory introduced in [16],
the grain mass cumulative distribution function P(x) of the Fuller
gradation can be expressed as

3−D
MðxÞ x
P ðxÞ ¼ ¼ ð1Þ
M0 xmax
where x is the sieve opening diameter or the grain size; M0 is the total
mass of all the aggregates; xmax is the maximum diameter of all the
aggregates; and M(x) is the total mass of aggregates with a grain size
smaller than x, which is a function of x; D is the fractal dimension of
concrete with graded aggregates; for the 3D model, D ∈ {2, 3}. By
comparing Eq. (1) with the Fuller equation P = (xi/xmax)n, we can

Fig. 4. Modified model (Ellipsoid-Model) generated by inscription.

 X. Yang et al. / Powder Technology 320 (2017) 161–178 165

demonstrate that the value of (3 − D) is equal to the exponent of the where V0 is the total volume of all the aggregates. The percent mass
Fuller Distribution, n [16]. Furthermore, the fractal dimension cannot retained on each sieve level in the numerical model generated by RFM
only indicate the gradation of concrete, but also represents the geomet- is given by
ric properties of concrete [15,16]. Thus, in this paper, we select D to 8
distinguish between different aggregates. > D  D−3 
< A −1 ; i¼1
Let ρ denote the apparent density of the aggregates; N(x), the total PRFractal ¼ D−3   ð12Þ
i
>
: D D−3 ð 1−i Þ ð 3−D Þ 3−D
number of aggregates with a grain size smaller than x; and V(x), the A B 1−B ; 1bi ≤ n
D−3
volume of aggregates with grain size x. Considering ρ to be constant,
we get
Further, the percent mass retained on each sieve level in the Fuller
MðxÞ ¼ ρV ðxÞNðxÞ ð2Þ gradation is given by

Assuming kv as the volume form factor, we get V(x) = kvx3. 8


>
> x1 3−D
Furthermore, based on the fractal geometry theory, we get N(x) = >
> 1− ¼ 1−AD−3 ; i¼1
>
> xmax
N1(x/xmax)−D [13,15], where N1 is a function of xmax. Thus, >
>


< x 3−D
xi 3−D  
PRFuller ¼
i−1
− ¼ AD−3 Bð1−iÞð3−DÞ B3−D −1 ; 1 b i ≤ n−1
i
>
> xmax xmax !
dMðxÞ ¼ ρðdV ðxÞ  NðxÞ þ V ðxÞ  dNðxÞÞ >
>
>

ð3Þ >
> xn−1 3−D
¼ ρð3−DÞkv N1 xDmax x2−D dx >
: 1− 1− x ¼ AD−3 Bð2−nÞð3−DÞ ; i¼n
max

Then, ð13Þ

1 ρkv xDmax  N 1
dPðxÞ ¼ dM ðxÞ ¼ ð3−DÞx2−D dx ð4Þ A correction factor αi is set for the percent mass retained on each
M0 M0
sieve level:
Performing an integral transformation, 8
>
> 3−D
< ; 1 ≤ i ≤ n−1
ρkv xDmax  N1 3−D PRFuller
i D
P ðxÞ ¼ x ð5Þ ai ¼ ¼ 3−D ð14Þ
PRFractal >
> D−3 B
M0 i : ; i¼n
D 1−B3−D
According to the significance of P(x), i.e.,
Therefore, it can be considered that when the aggregate volume in

P ðxmin Þ ¼ 0 each sieve level equals αiV Modified
i , the gradation follows the Fuller distri-
ð6Þ
P ðxmax Þ ¼ 1 bution. Based on the aggregate volumes, the number of aggregates in
each level N Modified
i can be approximately calculated, and it is equal to
we get the total aggregate volume of the current sieve level divided by the
volume of the average aggregate size.
M0
N1 ¼ ð7Þ
ρkv x3max
α i V Modified
NModified
i ¼ i 
In the differential interval of [x, x + dx], the increment of the particle V x3i
number dN(x) follows the condition 8V  
>
>
0
1−AD−3 x−3
>
> kv i ; i¼1
>
<V  
M 0 D−3 −1−D 0 D−3 ð1−iÞð3−DÞ −3
dNðxÞ ¼ −D x x dx ð8Þ ¼ A B B 3−D
−1 xi ; 1 b i ≤ n−1 ð15Þ
ρkv max > kv
>
>
>
>
: V 0 AD−3 Bð2−nÞð3−DÞ x−3 ;
n i¼n
Thus, the number of particles within the range [x1, x2] should be kv
Z x2
M0 D−3 −1−D M 0 D−3  −D 
NðxÞjxx21 ¼ −D x x dx ¼ x x −x−D ð9Þ where xi is the average aggregate size of sieve level i. It can be obtained
x1 ρkv max ρkv max 1 2
in the following manner:
The particle size x of the generated aggregate is considered to follow
In this article, for convenience of narration, each sieving size is a normal distribution with mean xi and variance σ
represented as xi, x1 is the maximum sieving size, and the remaining 2
. It should be noted that aggregates with particle size less than xi are
sizes are represented as x2, x3 and so on. Approximately, xi = Bxi + 1 in counted as one in level i + 1. Therefore, it can be said that the aggregates
the model generated by RFM. Further, let xmax = Ax1. Then, the actual generated by the ith iteration and belonging to level i are only distribut-
volume of each level should be ed on one side of the normal distribution, which is the part of x+ i
8 Z xmax marked in Fig. 5. On the other hand, the part of x− i in Fig. 5 represents
>
> V ðxÞdNðxÞ; i¼1 the aggregates generated by the ith iteration and belonging to level
<
V Modified ¼ NðxÞV i ðxÞ ¼ Zx1xi−1 ð10Þ i + 1.
i >
>
: V ðxÞdNðxÞ; 1bi ≤ n By calculating the limitless generalized integral of the probability
xi density function of normal distribution, the average size of aggregates
generated by the ith iteration and belonging to levels i and i + 1 can
Substituting Eqs. (8) and (9) into Eq. (10) and performing an be respectively formulated as
integral transformation,
8 Z þ∞
8   > þ 1 −
ðx−μ Þ2 σ
> D >
< x ¼ pffiffiffiffiffiffi x  e 2σ 2 dx ¼ μ þ pffiffiffiffiffiffi
< V 0 AD−3 −1 ; i¼1 2πσ Zμ 2 2π
V Modified ¼ D−3   ð11Þ μ ð16Þ
> > 1
> x− ¼ pffiffiffiffiffiffi ðx−μ Þ2 σ
: D V 0 AD−3 Bð1−iÞð3−DÞ 1−B3−D ;
i −
1bi ≤ n : x  e 2σ 2 dx ¼ μ− pffiffiffiffiffiffi
D−3 2πσ −∞ 2 2π
166 X. Yang et al. / Powder Technology 320 (2017) 161–178

2.3.2. Aggregate generation probability coefficient

If the number of aggregates of 3D-TYPE-1 is not modified, the
ideal number of aggregates generated by the ith iteration, Nidealized
i ,
would be
 i−1
Nidealized
i ¼ M  B3 −M ð22Þ

After n iterations, the total number of aggregates in one cell, Nidealized

n ,
would be
 n
X
n  i−1 1− B3 −M
Nidealized
n ¼ 3
M  B −M ¼M ð23Þ
i¼1 1−B3 þ M

Considering the actual number of aggregates, NModified

Fig. 5. Probability density curve of a normal distribution with the schematic diagram
aggregate distribution.
i , the number
of aggregates in the model can be modified by defining the ratio of the
actual vs. the ideal number of aggregates in each level as the aggregate
generation probability coefficient (AGPC), which is represented by χi
and can be defined as

NModified
χi ¼ i
ð24Þ
where the value of σ2 can be obtained by the three-sigma rule of normal Nidealized
i  Ncell
distribution.
where Ncell is the number of cell sets. The manner of calculating Ncell and
8
 its effect on χi are discussed in Section 2.4.1.
>
> 1 1 1 In the modelling process, the ideal aggregate number Nidealized
< 6σ ¼ Axi − xi ; σ¼ Axi − xi ; i¼1 n
B 6
B  is influenced by the values of B, M, and the number of iterations n.
ð17Þ
>
> 1 1 1 The values of these three parameters and the number of cells Ncell
: 6σ ¼ Bxi − xi ; σ¼ Bxi − xi ; 1bi ≤ n
B 6 B should be valued reasonably to ensure that the total number of
ideal aggregates is greater than or equal to NModified
i so that each
AGPC χi would be less than or equal to 1. In this case, the physical
Substituting Eq. (17) into Eq. (16) gives meaning of χi is the probability of occurrence of an aggregate in the
8
 ideal location. Before each generation of an aggregate, a computer-
>
> 1 1 generated random number is compared with AGPC; the result of the
pffiffiffiffiffiffi Axi − xi ; i ¼ 1
< xi 
12 2π
B 
x
i ¼ ð18Þ comparison tells the computer whether to generate this model. Thus,
>
> 1 1
: xi  pffiffiffiffiffiffi Bxi − xi ; 1 b i ≤ n the gradation of the generated model will agree with the actual
12 2π B
condition.
Additionally, the values of B and M can be manually selected. Note
The average particle size of each level is related to the number of that these two values influence not only the calculated AGPC results,
aggregates, which can be written as but also the coarse aggregate content, which we discuss in Appendix B.

2.4. Simulation method

N0 ðxi Þ  xþ 0 −
i þ N ðxi−1 Þ  xi−1
xi ¼ 0 0 ; 1bi ≤ n ð19Þ
N ðxi Þ þ N ðxi−1 Þ The main operation processes for simulating a concrete sample are
shown in Fig. 6.
where N′(xi) is the number of aggregates generated by the ith iteration The three most important steps of this flowchart are set cell,
and belonging to level i. It can be considered that N′(xi) is equal to aggregate generation, and edge modification, which are detailed in the
approximately half of the value of N Modified
i . following subsections.

2.4.1. Set cell

V 0 D−3 ð1−iÞð3−DÞ  3−D 
Based on Hypothesis 1, the following corollary can be obtained [15]:
A B B −1 x−3
N0 ðxi Þ k i
¼ v   ¼ BD ð20Þ Corollary 1. If F1 , F2 , ⋯, Fn is a countable sequence of sets and each Fi
N0 ðxi−1 Þ V 0 D−3 ð2−iÞð3−DÞ 3−D −3
A B B −1 xi−1 has the same dimension, then dim∪ni =1Fi = dim Fi
kv
Thus, the fractal dimension of the model combined with several cells
would be equal to D as long as the fractal dimension of each cell is equal
Then, the average particle size of each level can be obtained.
to D. This is the theoretical basis for simulating concrete using arrayed
8 þ cells.
< xi ; i¼1 The set cell operation should follow the principle of ensuring that
xi ¼ BD  xþ −
i þ xi−1 ð21Þ the cells can completely cover the concrete sample.
: D
; 1bi ≤ n
B þ1 Based on this principle, the number of cells Ncell can be calculated as
follows. Let Ci , i = 1 , 2 , 3, denote the lengths of the concrete sample
Thus, based on the values of sieving, the actual number of graded in three directions. If the sample is irregularly shaped, Ci should be the
aggregates can be obtained. edge length of the smallest cuboid that can hold the sample. Therefore,
 X. Yang et al. / Powder Technology 320 (2017) 161–178 167

Fig. 6. Flowchart of simulation method.

the number of cells that can ensure that the entire sample can be
covered is given by
Y
3
C respectively. There are two important requirements for the aggregate
Ncell ¼
i  ð25Þ
i¼1 ceil
min ðC i Þ
S actual aggregates, and avoiding the overlapping phenomenon of the
Substituting Eqs. (15), (23), and (25) into Eq. (24), the AGPC of each
level can be obtained. It can be assumed that the value of χ1 does not
change with the length of the cell side without regard for the ceiling
operation; only the fractal dimension, volume of aggregates, and
sequence of sieving sizes are relevant. Further, according to the previous
formula, the AGPC is the only function of the fractal dimension that can
directly influence the generated model. Therefore, it can be said that the
application of AGPC can ensure that the fractal dimension of the simu-
lated model equals that of the actual condition. Based on Corollary 1, if
the numerical model has the same fractal dimension as the concrete
sample, these two geometrical patterns are considered ‘the same’. In
other words, AGPC directly influences the effect of simulation.
2.4.2. Aggregate generation
After the set cell operation, aggregates can be generated in each cell.
The aggregate generation method and the manner of modifying the
number of aggregates have been introduced in Sections 2.2 and 2.3,
generation: choosing a suitable modelling method to simulate the
aggregates.
The models generated by RFM do not actually cause the overlapping
phenomenon, whereas the phenomenon of sharing vertices by multiple
aggregates might occur more frequently. This phenomenon is influ-
enced by the manner of cell division and location of aggregate genera-
tion in each iteration. For example, Fig. 7 schematically shows some
possible situations.
In fact, if the coarse aggregate content of the concrete sample to be
simulated is high, it would probably involve the sharing of vertices.
Thus, this paper proposes the generation of a modified model using
the second or third approach introduced in Section 2.2.2 in the case of
simulating a concrete sample with higher coarse aggregate content.
This modelling approach can avoid the generation of aggregate models
that share a common vertex.
168 X. Yang et al. / Powder Technology 320 (2017) 161–178
Fig. 7. Schematic diagram of situations in which vertices are shared.
2.4.3. Edge modification
Because the arrayed cells should completely cover the concrete
sample, there would be some aggregate models or several parts of
an aggregate model beyond the sample. These models should not be
calculated or displayed in the final numerical model. Thus, it would be
necessary to treat these aggregates. In this paper, this operation is
referred to as edge modification.
The theoretical basis and the feasibility of edge modification are
discussed as follows. Based on Hypothesis 1, a converse proposition of
Corollary 2 can be obtained [15]:
Corollary 2. Any part selected from concrete is fractal, where the set of
aggregates and the set of cement have the same fractal dimension as the
original concrete.
Based on this corollary, the simulated model can be cut in a random
shape, while the geometric properties and fractal dimension remain
constant.
During edge modification, all the aggregate models located
outside the sample should be discarded; then, the models with
parts located outside the sample and other parts located inside the
sample should be treated in the following two ways. 1. Using the
Boolean intersection, keep each inside part and discard the other part,
as shown in Fig. 8(a). 2. Discard all these models directly, as shown in
Fig. 8(c).
In effect, these two edge modification approaches do not affect the
stress or strain distribution in the simulated model; however, the stress
or strain near the surface of the simulated model would appear to have a
greater effect. Therefore, if the analysis result near the surface is impor-
tant, the approach should be chosen carefully on the basis of the actual
condition. In particular, concrete products that have a surface with
aggregates revealed or those formed by cutting should be simulated
using the first edge modification approach.
Additionally, edge modification influences the application of the
simulation, especially for samples with thin cross-sections. Further
discussion of edge modification is noted in Appendix C.
2.4.4. Cement paste generation
In a meso-mechanics framework, concrete can be simplified into a
three-phase mixture, including aggregates, cement paste and the inter-
face transition zone (ITZ) between the previous two components.
Cement paste is considered a continuous material. Its numerical
model is usually generated such that after aggregate generation, the
remaining space in the numerical concrete sample is directly treated
as cement [7]. Normally it is not necessary to generate the cement
paste specially; its generation is usually performed with the numerical
model mesh [10].
Moreover, note that depending on the meso-mechanical theoretical
model selected, air-voids might exist in the cement paste. Some
researchers also assume air-voids in the cement paste as a fourth
phase of concrete, which affects computation [18].
The material properties of cement paste and ITZ affect the mechan-
ical behaviour of numerical concrete models. Nevertheless, our study
mainly involves the geometric properties and generation processes of
aggregates, so we limit our discussion of cement paste generation in
this paper.
3. Simulation and statistical analysis
Two simulation examples are presented in this section. One is the
simulation of a concrete hollow block for the purpose of explaining
the simulation method for a complex concrete production; the other is
the generation of a group of concrete cube samples to be measured
and analysed.
3.1. Simulation model
3.1.1. Simulation of concrete block
The simulation process for a hollow concrete block is shown in Fig. 9.
This type of block is produced by crushed stone aggregates; the
coarse aggregate content is 40%. The sieving result of the aggregates is
summarized in Table 1.
 X. Yang et al. / Powder Technology 320 (2017) 161–178
Fig. 8. Schematic diagram edge modification.
By calculating the value of D as in Section 2.1.2, we can plot the curve
shown in Fig. 10. The slope of the trend line is 0.3675 and the fractal
dimension of the aggregates is 2.6325.
With x1 equal to 10 mm, the cell was set as a cube with S =20 mm
on each side. The other parameters were set as A = 1.2 and B = 2; the
value of M can be determined to be 3 based on known conditions, as
detailed in Appendix B. The location of the generated aggregates was
random, i.e., in any one of the eight sub-cells.
Table 1 represents an experimental sieving test. Particles whose size
is no greater than 1.25 mm can be considered part of the cement paste.
Therefore, the number of iterations was set to 3, and χ1, χ2, and χ3 were
obtained by calculation as 0.370, 0.905, and 0.817. Convex-Polyhedron-
3 was chosen to simulate this sample.
After preparing the parameters, the arrayed cells were set. It would
be feasible to set an array of 10× 10 × 10 to completely cover the entire
sample, and the extra parts would be discarded eventually. However,
more than 30% of the aggregate models would be useless. Therefore,
to reduce the computation burden, the arrayed cells were set as
shown in Fig. 11(a), i.e., 640 cells were set. Then, the aggregates were
Fig. 9. Dimensions of concrete hollow block.
169
generated in each cell, as shown in Fig. 11(b). It could be measured
that the coarse aggregate content of this simulation model at this
stage was 42.82%. Finally, edge modification was performed. Because
of the moulding technology for concrete blocks, no aggregates would
be exposed. The discarding method for edge modification was chosen.
The overhead view of the final model is shown in Fig. 11(c). The coarse
aggregate content at this stage, i.e. after the edge modification, became
33.27%. The relative error between the values before and after this
operation was more than 20%. It could be observed that the edge
modification influenced a fluctuation of the coarse aggregate content
in this model, which would be discussed further in Appendix C.
The generation process introduced here is explains the entire
simulation method. However, it was not applied to statistically analysed
sampling because of its complex shape, which we do not further
analyse. The model introduced in Section 3.1.2 is measured and
analysed quantitatively.
3.1.2. Simulation of concrete cube sample
The models generated here are analysed as the overall sample in the
following section. The dimension of the models was determined by
referring to the sampling method presented by Wang et al. [19,20]:
the size of the overall sample should be at least 15 times larger than
that of the maximum aggregate size. Thus, 100 cube models with
edges equalling 500 mm were generated according to the gradation
shown in Table 1. Fifty of the simulation models were generated by
Ellipsoid-Model, as shown in Fig. 12, while the remaining models
were generated by Convex-Polyhedron-3, as shown in Fig. 13. It could
be measured that the coarse aggregate contents of the models shown
in Figs. 12 and 13 were respectively 41.16% and 40.39%. More statistical
results were shown and discussed in Section 4.2.
3.2. Sample statistical analysis method
To verify the modelling effect, the group of concrete meso-models
generated in Section 3.1.2 was subjected to sample statistical analysis.
3.2.1. Sampling method
Based on the method introduced in [19,20], the size of the sub-
sample should be at least seven times that of the maximum aggregate
size. Thus, a randomly located cube with an edge length of 150 mm
was taken from each overall sample as a sub-sample, and each overall
170 X. Yang et al. / Powder Technology 320 (2017) 161–178

Table 1
Result of sieving test.

Total mass IS sieve size Separated sieve residue Separated mass retained Cumulated sieve residue Mass retained Mass retained P(x)
(kg) (mm) (kg) (%) (%) (kg) (%)

1 10 0.065 0.065 0.065 0.935 0.935

5 0.210 0.210 0.275 0.725 0.725
2.5 0.163 0.163 0.438 0.562 0.562
1.25 0.126 0.126 0.565 0.436 0.436
0.63 0.097 0.097 0.661 0.339 0.339
0.315 0.076 0.076 0.738 0.262 0.262
0.16 0.058 0.058 0.795 0.205 0.205
PAN 0.204 0.204 0.999 0.001 0.001

sample was collected four times. A schematic diagram of the sampling
method is shown in Fig. 13.
3.2.2. Meso-structure indicator
After sampling, the gradation curves of the overall sample and the
sub-sample could be plotted. Further, two indices were selected as the
meso-structure indicators: coarse aggregate content and fractal dimen-
sion of sample. These indices are determined as follows.
The coarse aggregate content is 40% in this model. The fractal dimen-
sion of the numerical model can be obtained using the box-counting
method [15]. In general, the fractal dimension can be defined in several
ways, including via the Hausdorff or box-counting dimensions. In fact,
the mathematical meaning of the Hausdorff dimension is different
from that of the box-counting dimension, and each has its advantages
and disadvantages. The Box-counting dimension can be measured
easily; however, its application is restricted. In contrast, the Hausdorff
dimension is defined for any set, and is mathematically convenient as
it is based on measures that are relatively easy to manipulate. Neverthe-
less, a major disadvantage is that in many cases, it is difficult to calculate
or estimate computationally.
For the properties of the set and fractal dimension studied in this
article, these two types of fractal dimension are the same. Thus, we
selected the box-counting dimension to represent and measure the
fractal dimension of the numerical model, and the Hausdorff dimension
to represent and measure the fractal dimension of the actual gradation
of aggregates.
3.2.3. Geometrical homogenization
One of the purposes of using meso-mechanics to study materials is
to replace a heterogeneous material by an equivalent homogenous
material and then determine an effective macro-scale constitutive
relation [21]. Therefore, to a certain extent, the meso-mechanics is to
be studied at two levels: macro-scale and meso-scale. In the macro-
scale, the object analysed is considered as a continuum model that
consists of many material points. A material point is relative to a
meso-space, which is named as a representative volume element. As a
basic concept in meso-mechanics, the representative volume element
has duality of scale. On the one hand, in the macro-scale, the size of
the element should be sufficiently small so that the stress and strain in
the macro-scale can be considered homogeneous. On the other hand,
in the meso-scale, the size of the element should be sufficiently large
so that an element can contain sufficient meso-elements and
meso-information. Therefore, a representative volume element should
reflect the statistical properties of a part of the continuum, and the
sample taken from the meso-model should follow the geometrical
homogenization.
A method to determine the geometrical homogenization of a meso-
structure has been introduced in [20]. Two conditions are specified.
First, the meso-structure indicators should not be influenced by the
randomness of the aggregate sample. Second, the meso-structure
indicators of the sub-sample should have statistical similarity with the
overall sample.
These two conditions can be statistically analysed by quantitatively
using the coefficient of variation and relative tolerance. In this article,
the coarse aggregate content and fractal dimension of concrete are
selected as the meso-structure indicators. Based on the measured
method in [20], the coefficient of variation and relative tolerance was
set as 5%.
To verify whether the sample fits the first condition of geometrical
homogenization, i.e., discreteness, the coefficient of variation should
be less than the allowed value of 5%. The function to calculate the
value of the coefficient of variation (C.V.) is given by

C:V: ¼ ðSD=MNÞ  100% ð26Þ

where SD is the standard deviation and MN is the mean. If the coefficient

of variation is less than the allowed value of 5%, it implies that the
random error of the sample is stable, which fits the first condition.
To verify whether the sample fits the second condition of geometri-
cal homogenization, i.e., similarity, a criterion is introduced as follows:

P s −P t
ε¼ b jε j ð27Þ
Pt

where ε is the relative error between each sub-sample and the

overall sample; Ps and Pt represent the meso-structure indicator in
Section 3.2.2, i.e., the coarse aggregate content and fractal dimension
of the sub-sample and overall sample, respectively; and | ε | is the
Fig. 10. P(x) and (x/xmax) drawn in a log-log plot to calculate the value of D. allowed value of the relative error, which is set as 5% in this article.
 X. Yang et al. / Powder Technology 320 (2017) 161–178
Fig. 11. Schematic diagram of modelling process.
If all the measured values are less than the allowed 5%, it could be
considered that the spatial distribution properties of the sub-sample
are similar to those of the overall sample, proving that the numerical
model fits the second condition.
4. Results
The measured values and statistical analyses are presented in this
section.
4.1. Gradation curve
The density of each aggregate can be considered to be the same.
Based on the number of aggregates obtained using a computer, the
gradation curves of the simulation models generated by Convex-
Polyhedron-3 and Ellipsoid-Model are plotted in Fig. 14(a) and (b),
respectively, and the gradation curve of the referenced object is also
plotted in the same figure. Further, two of the simulation models
generated by different types of model is randomly selected and treated
Fig. 12. Concrete simulation model generated by Ellipsoid-Model.
171
as the overall sample. The gradation curves of the overall sample and
sub-samples are respectively plotted in Fig. 14(c) and (d).
These curves show a similar form overall and there are some small
differences influenced by random parameters. This indicates that the
numerical model of concrete generated by the two methods fits the
aggregate gradation.
The gradation curve is used to represent the aggregate relative
content of each sieving level, which reflects the components of the
aggregates. Further, the meso-structure of concrete can influence
other physical quantities. Thus, the coarse aggregate content and fractal
dimension are analysed in the following sub-section.
4.2. Statistical result of other measurement values and analysis of
geometrical homogenization
The statistical results of measurement of all of the models are
summarized in Table 2. Based on the statistical method introduced in
Section 3.2, the coarse aggregate content and Box-counting dimension
are analysed.
We selected two of the models generated by Convex-Polyhedron-3
and Ellipsoid-Model; the measurements of the overall sample and four
sub-samples are listed in Table 3 and Table 4. The results show that
each index remains stable when the sub-sample changes, and the
difference between the measurements of the sub-sample and the over-
all sample is small. The measurements of other models are similar to
this one, and they are not represented here. Additional discussion of
the coarse aggregate content of the RFM is introduced in Appendix B.
5. Discussion
5.1. Effect of modelling
Based on the measurement result, both the actual concrete and its
numerical model have fractal effects that could be represented by the
fractal dimension. The measured fractal dimension of the numerical
model Dbox is the same as that of the aggregate gradation. Based on
Corollary 1, it could be considered that the set of simulated aggregates
is the same as the set of real aggregates.
Patterns with the same fractal dimension have the same geometrical
properties. In addition, based on gradation curve and coarse aggregate
content, numerical models with the same fractal dimension have the
same gradation characteristics, as shown in Fig. 14.
The value of D can indicate the geometric properties of the concrete,
reflecting its scaling properties and self-similarity. Considering all the
aggregates as one geometric figure, a larger D indicates more complex
graphics [15]. Moreover, D is directly related to the Fuller exponent,
which decreases as the value of D increases. A larger value of D, i.e., a
172 X. Yang et al. / Powder Technology 320 (2017) 161–178
Fig. 13. Concrete simulation model generated by Convex-Polyhedron-3 and sampling method.
smaller Fuller exponent, represents a gradation with fewer large
aggregates than small ones. In conclusion, changes in the value of D
can reflect the geometry and gradation of concrete.
Thus, it is reasonable to consider the fractal dimension as an
evaluation criterion to determine whether two groups of aggregates
are the same.
The discussion in this section only focuses on how the simulation
results fit actual conditions with a specific sieve status. Additional
applications are discussed in Appendix B.
5.2. Geometrical homogenization
The measurement values and statistical results demonstrate that the
model generated by RFM has a stable random error. The measurement
value of each sub-sample has a relative error that is less than 5%, indicat-
ing that the spatial distribution of the sub-samples is similar to that of
the overall sample. This occurs because the pattern generated by the
fractal method represents the property that each of its parts has the
same geometrical properties. Moreover, based on the gradation curve
and coarse aggregate content each sub-sample has the same gradation.
In summary, the meso-model generated by RFM fits the conditions of
geometrical homogenization; the spatial distribution is in accordance
with practical situations, and this method is applied effectively.
5.3. Algorithm complexity
The amount of time and storage required to execute algorithms can
be represented quantitatively by their complexity, which is an impor-
tant measure of the quality of an algorithm. The algorithm complexity
consists of space complexity and time complexity. The space complexity
represents the storage space required by the algorithm, which is
expressed using big S notation. The time complexity describes the
time required by the algorithm from start to end, which is expressed
using big O notation.
Algorithm complexities are classified as a function about n. For
instance, the following functions represent complexities from small to
large: F(1), F(n log n), F(n), F(n2), 2F(n).
The lower the values of these two quantities, the more efficient is the
algorithm.
In general, the meso-modelling process for concrete should be
repeated as many times as possible, which involves large quantities of
objects, such as geometric forms, nodes, and elements. Hence, the algo-
rithm complexity of the modelling method influences the runtime,
which is discussed below.
5.3.1. Space complexity
The storage space required by an algorithm consists of three parts:
the space used to store the algorithm, the space used to store the
input and output data, and the space used to store the temporary data.
The space used to store the algorithm depends on the code of the algo-
rithm. The space used to store the input and output data depends on the
problem to be solved, which cannot be influenced by the algorithm.
Further, the quantity of temporary data is determined by the algorithm.
The input data of concrete meso-modelling consist of the sample
dimension, results of sieving, etc., and the output data consist of the
geometrical parameters of each aggregate. Both the input and output
data were almost unchanged for a simulation example. Further, the
code length does not have a significant influence on the entire space
used. Thus, the major concern of space complexity for RFM algorithm
is the size of temporary data. If a recursive algorithm is used for RFM
modelling, the space complexity exhibits a linear relationship with the
input data, which can be noted as S(n). Nevertheless, using the iterative
algorithm could significantly reduce the storage required, i.e., the space
complexity, denoted as S(1) [22]. In practice, the iterative algorithm is
indeed a preferred choice.
5.3.2. Time complexity
As the RFM code is designed by a recursive algorithm, the program
would require an unacceptable exponential time, i.e., 2O(n). Moreover,
the storage of temporary data would increase linearly as the input
data increase. Thus, it can be concluded that the recursive algorithm is
accurate but hopelessly inefficient. Therefore, the modelling program
is usually coded using an iterative algorithm. In this case, RFM can be
implemented simply by using a single loop, and it would become a
linear time operation, O(n) [22]. The recursive method can be used to
develop the modelling approach; nevertheless, the iterative algorithm
should be used in practical computation.
Furthermore, limited by the fractal effect limited dimension [17], the
number of iterations is usually less than a constant. In this instance, the
number of operations would be stable, which is an optimal situation,
denoted by O(1).
5.3.3. Comparison with other algorithms
Theoretically, the effects of RFM modelling and other modelling
methods do not differ significantly, which can be demonstrated by the
gradation curve and statistical result of other measurement values, as
shown in Section 4. By contrast, the algorithm complexities of these
methods differ considerably. The first and second types of modelling
methods introduced in Section 1 are essentially a type of comparison
sort.
 X. Yang et al. / Powder Technology 320 (2017) 161–178 173

Fig. 14. Gradation curves.

Table 2 The space complexity of the first and second types of methods
Statistical analysis of the results of measurements of concrete meso-simulation.
generally does not change, i.e., S(1). Thus, the RFM coded using an iter-
pv (%) Dbox ative algorithm has the same space complexity as these two types of
M0 M1 M2 M0 M1 M2
methods.

Mean 39.898 40.033 39.763 2.503 2.502 2.503

Sd 1.632 1.554 1.711 0.010 0.010 0.010 Table 3
Max 43.715 43.715 43.075 2.527 2.527 2.522 Sampling measurement of models generated by Convex-Polyhedron-3.
Min 36.516 36.687 36.516 2.480 2.483 2.480
Med 39.993 39.978 39.997 2.503 2.501 2.504 pv Dbox
CV 4.089% 3.882% 4.303% 0.398% 0.392% 0.408%
val (%) ε (%) val ε (%)
Note: M0 represents all the 100 models; M1 represents the 50 models generated by
G0 (Pt) 40.394 2.521
Convex-Polyhedron-3; M2 represents the 50 models generated by Ellipsoid-Model; pv
Sub1 (Ps) 42.657 5.60 2.499 0.88
represents the coarse aggregate content; Dbox represents the dimension calculated by
Sub2 (Ps) 40.003 0.97 2.500 0.84
Box-counting; Sd represents the standard deviation; Max and Min represent the maxi-
Sub3 (Ps) 39.953 1.09 2.504 0.69
mum and minimum values, respectively; Med represents the median; CV represents the
Sub4 (Ps) 38.830 3.87 2.518 0.11
coefficient of variation.
174 X. Yang et al. / Powder Technology 320 (2017) 161–178
Table 4
Sampling measurement of models generated by Ellipsoid-Model.
pv Dbox
val (%) ε (%) val
G0 (Pt) 41.159 2.492
Sub1 (Ps) 40.914 0.60 2.515
Sub2 (Ps) 39.856 3.17 2.499
Sub3 (Ps) 42.351 2.89 2.503
Sub4 (Ps) 38.227 7.12 2.506
Note: val represents the measurement value; ε is the relative error between each sub-
sample and overall sample; G0 represents the overall sample; Sub1–Sub4 represent the
four sub-samples.
As for the time complexity, there are fundamental limits on the
performance of comparison sorts. A comparison sort must have an
average-case lower bound of comparison operations, which is known
as linearithmic time [23], e.g., the fastest possible comparison sort,
shell sort (O(n log n)) [24,25]. Others, such as bubble sort (O(n)) [26]
and insertion sort (O(n2)) [27] need more running time. In effect, even
the most efficient method, i.e., shell sort, needs quadratic time (O(n2))
to handle the worst-case situation. The average time complexity of
comparison sorts could be treated as quadratic time (O(n2)) [28].
6. Conclusion and future work
This paper introduced a versatile and efficient concrete meso-
modelling method. The following conclusions could be drawn.
a) In the framework of fractal geometry, the theoretical basis of RFM
was introduced, according to which methods for modifying the
aggregate model shape and the quantity of aggregates were
introduced. Furthermore, a complete modelling method could be
developed.
b) Taking a concrete hollow block as an example, we detailed the
implementation of RFM for simulating a complex object, i.e., a
concrete block. Further, we generated two groups of concrete cube
samples with aggregates of cobblestone and crushed stone using
RFM.
c) The numerical cube models were statistically analysed. The grada-
tion curves of the models had forms similar to the actual sieving
results. Further, the coarse aggregate content and fractal dimension
of the numerical models were identical to those of the actual
situation.
d) A comparison was made between the RFM and other modelling
methods in terms of algorithm complexity. With regard to the
iterative algorithm implemented, the storage used by RFM and the
traditional modelling methods are basically the same. The operation
efficiency of RFM is relatively stable, which requires linear time. The
time complexity of other modelling methods would be lower than
that of RFM in the best of circumstances, which is not common;
the average time complexity of the traditional methods would be
considered as quadratic time, which shows that the RFM modelling
process is faster.
It was reported that the applied effect and processing efficiency of
RFM meet the needs of concrete meso-modelling.
The significance of this research exists in providing a method of
expanding the sample size of numerical concrete while ensuring that
all samples have a stable geometry, which can help researchers to
focus on the influence of each component's material properties. RFM
could potentially be applied to concrete non-destructive testing with
the help of CT technology, concrete mechanical behaviour, research on
representative volume elements or concrete homogenization.
A precise numerical model is the basis for future concrete meso-
scale studies on the mechanical behaviour, chlorine diffusion, etc.,
which are related to many other physical properties apart from the
geometrical characteristics. Thus, the following discussions should be
included in future work: using the known gradation curve of concrete,
the material properties of the aggregate, cement paste, and the ITZ
ε (%)
between them should be considered. Furthermore, numerical simula-
tion of the failure process and prediction of the capacity of concrete
should be studied.
0.89
0.25
0.42 Acknowledgment
0.55
This study was supported by the National Key R&D Plan (Grant No.
2016YFC0701502-2). Thanks are due to Dr. WEI Chun of Huazhong Uni-
versity of Science and Technology and Prof. LIANG Yongshun of Nanjing
University of Science & Technology for the valuable suggestions about
the theoretical framework of fractal dimension.
Appendix A. Aggregate generation method
A.1. Shape modification
The operations introduced in Section 2.2.2, i.e., OP1, OP2 and OP3,
are detailed in this appendix.
OP1 can be divided into two categories: generation of the Random-
Cuboid-Model and 3D-TYPE-2. Flowcharts describing these two
procedures are presented in Fig. A.1.
OP2 can be separated by object into four categories: generating a
sphere in a cube, generating a random sphere in a random cuboid,
generating an ellipsoid in a random cuboid and generating a polyhedron
in a sphere. These models can be generated in the following way:
1. The diameter of the sphere inscribed in a cube equals the edge of the
cube, and the spherical centre is located in the centre of the cube;
2. The diameter of the sphere inscribed in a random cuboid equals the
minimum edge of the cuboid, and the location of the sphere can be
selected randomly as long as the entire sphere could be located
inside the cuboid;
3. The three axes of the ellipsoid respectively equal the edge of the
cuboid in three directions, and the ellipsoidal centre is located in
the centre of the cuboid;
4. A polyhedron generated in a sphere can refer to the methods intro-
duced in [5,30].
The object on which OP3 works is one or some of the polyhedral
angles, and may be performed as follows. First, generate a sphere
whose centre is located on the vertex to be treated and its radius equals
a positive real number which should be less than the shortest length of
the edges connected to this vertex. This spherical surface intersects with
at least three polyhedral edges. Therefore, at least three intersecting
points can be obtained. Then, randomly select three intersecting points
to define a spatial plane. This plane is used to cut this corner.
Shape modification for aggregate models can be implemented using
OP1, OP2, or OP3 as required. These operations can be applied not only
to RFM, but also to other modelling methods.
A.2. Differences between models
In this section, we compare similar aggregate models, including their
advantages, disadvantages and applications.
The spherical aggregate model can be generated by Sphere-Mode-1
or Sphere-Model-2, which are applied to simulate cobbles. There would
rarely be any stress concentration or variation on the surface of spheri-
cal aggregate or in the cement near it. For this reason, a sphere can also
be used to simulate crushed stone, which is applicable to a variety of
situations [29].
There are some differences between these two models: first, all
spheres of the same iteration generated by Sphere-Model-1 have the
same radius and a stable location. This kind of simulation model can
 X. Yang et al. / Powder Technology 320 (2017) 161–178 175

a) Flowchart for generating the Random-Cuboid-Model

b) Flowchart for generating 3D-TYPE-2

Fig. A.1. Flowchart of OP1.
176 X. Yang et al. / Powder Technology 320 (2017) 161–178
be generated quickly, which could be used in trials to obtain a qualita-
tive or fast result; second, the models obtained by Sphere-Model-2 are
more random and approximate real-world situations.
Convex-Polyhedron-1, Convex-Polyhedron-2, and Convex-
Polyhedron-3 can all be implemented to generate convex polyhedral
models for simulating crushed stone. Polyhedral aggregate models
might lead to enormous stress or strain variations in the cement near
an aggregate. However, there is a lower probability of this phenomenon
in the models generated by these three methods than those generated
by 3D-TYPE-1, Random-Cuboid-Method or 3D-Type-2. The numerical
model generated by Convex-Polyhedron-1, Convex-Polyhedron-2 and
Convex-Polyhedron-3 can be applied to simulate concrete.
There are some differences between these three kinds of models.
The aggregate models generated by Convex-Polyhedron-1 have a stable
location, can be generated quickly, and are usually applied to obtain
some qualitative result. The aggregate models generated by Convex-
Polyhedron-2 have a less stable size than the other two models because
of the corner cutting procedure. In Convex-Polyhedron-1 and Convex-
Polyhedron-2, aggregates are simulated by polyhedral growth, and
frequently uses a search algorithm, requiring significant calculation.
Appendix B. Discussion of volume fraction
B.1. Relation between modelling method and coarse aggregate content
In the RFM simulation process, cell division and aggregate genera-
tion are limited by conditions which are mathematically described as
follows.
A cell in Euclidean space E, where E = 2,3, each edge is divided into
B parts, and M aggregates are generated in each iteration, where
M ≤BE. Then, the number of each level is M ⋅ (BE − M)i−1, and the fractal
dimension is
 
ln BE −M
D¼ ; ðB:1Þ
lnB
where when E = 2 (i.e., a 2-D model is formed), D ∈ [1, 2], and when
E = 3 (i.e., a 3-D model is formed), D ∈ [2, 3].
Defining the measure ϕ of each aggregate as volume (3-D) or area
(2-D), according to the iterative method, the measure of one aggregate
is (S ⋅ B−i)E, so the sum of the measures in level i is
 i−1  E
ϕi ¼ M B −ME
 kform S  B−i ; ðB:2Þ
where kform is the volume or area factor for grain size. After n iterations,
the total measure of aggregates in each cell is
X
n  i−1  E
ϕTotal ¼ M BE −M  kform S  B−i
i¼1   n  ðB:3Þ
¼ kform SE 1− 1−B−E M
Note that depending on the actual situation, the coarse aggregate
content pv, i.e., the proportion of the measure (volume or area) of
aggregates in the whole measure (volume or area) of concrete, should
satisfy the following relation:
  n 
pvnumerical ¼ kform 1− 1−B−E M
  ðB:4Þ
¼ kform 1−BnðD−EÞ ≥pvactual ;
Eq. (B.4) limits the aggregates' number in each iterative level and
divides iteratively to generate the aggregates. Additionally, M must be
an integer; when the other parameters are known, its value can be
calculated.
Eq. (B.4) represents how the coarse aggregate content of the model
generated by RFM is limited by the method of division. Specifically, in
the case of n = 1, M = 1, coarse aggregate content is minimized,
i.e., kformB−E; in the case of M = BE, not considering the issue of sharing
vertices, coarse aggregate content can be maximized, i.e., kform. There-
fore, the RFM can meet most concrete simulation demands.
In addition, depending on the meso-mechanical model selected, the
volume of cement paste in a numerical concrete sample is influenced
not only by the total volume of aggregates, but also the thickness of
the ITZ.
B.2. Supplemental simulation
As a supplement to this method, we simulate and analyse additional
concrete samples, to discuss the effect of different conditions on our
simulation.
The coarse aggregates' content pv and fractal dimension D were
selected as independent variables, which could determine the state of
the aggregate models. Setting the values of pv to 20%, 30%, 40%, … 80%
and the values of D to 2.1, 2.2, 2.3, … 2.9 gives us 63 different situations
to simulate.
Other parameters are set as follows: considering the gradation
following the Fuller distribution, Fuller exponent n is equal to 3 − D,
x1 = 10 mm, A = 1.2, B = 2, and the number of iterations was set to
three.
Each situation was simulated 30 times, and each model was
measured and analysed by plotting the sieving result of each numerical
model using a curve fitting power function similar to the Fuller
Equation, which can be written as

α m
m x i; j
f i; j ðxÞ ¼ ; in the case of Di and pv j ðB:5Þ
xmax
where i and j, respectively, indicate the order of D and pv; m indicates
the model order.
Then, calculate the relative difference ε m m
i , j between α i , j and the
corresponding ni. The mean of these 30 models εi,j is the relative differ-
ence for each concrete situation of concrete. That is, the relative of each
situation could be written as in Eq. (B.6), which can be considered a
function about D and pv:
 
i; j ¼ g Di ; pv j :
ε i; j ¼ εm ðB:6Þ
Then, set D, pv and εi , j as the x-, y- and z-axes, respectively of a 3D
surface that can be plotted as shown in Fig. B.1.
Observe that the relative difference increases as D decreases.
Because smaller values of D represent a less complex geometry graph,
in the model, all aggregate sizes are stable. Once an aggregate model
with an abnormal randomly generated size occurs, the geometry
graph becomes more complex, and an obvious change in gradation
appears.
Meanwhile, ε increases as coarse aggregate content decreases.
Because there are fewer aggregates for lower values of pv, once an
aggregate model generated has had an abnormal size, there would be
a remarkable change in the gradation.
Appendix C. Edge modification
In this section, we quantitatively discuss the influence of edge
modification on coarse aggregate content. The total volume of aggre-
gates before edge modification would be represented by the function
in Eq. (C.1).
V 0 ¼ Ncell  pv; ðC:1Þ
 X. Yang et al. / Powder Technology 320 (2017) 161–178
Fig. B.1. Surface of relative difference about D and pv.
where pv represents the statistical average of each cell, which could be
considered approximately equal to the pv of the sample.
Note that the total volume of aggregates after edge modification V1
is related to the number of cells treated. V1 can be obtained from
Eq. (C.2).
V 1 ¼ Ncell cell
0  pv þ N T  pvT ;
Fig. C.1. Schematic diagram of cell location types.
177
where Ncell cell
0 represents the number of cells not treated; NT represents
the number of cells that need to be treated; pvT represents the
statistical average of coarse aggregate content of each cell after edge
modification.
The value of pvT is relevant to many factors: the values of B and M,
location type of the cell (as shown in Fig. C.1), edge modification
method, ratio between the volume of remaining part and the volume
of the entire cell kR and pv. Especially, kR and pv have a remarkable
effect on pvT . Smaller values of kR or pv would probably cause low
values of pvT , and may even cause pvT to equal zero; meanwhile, larger
values for kR or pv could stabilize pvT .
The influence of edge modification to coarse aggregate content can
be represented quantitatively by the relative difference between V1
and V0:
jV 1 −V 0 j Ncell pv
εV ¼ ¼ Tcell  T −1 ðC:3Þ
V0 N pv
where εV represents the relative difference between V1 and V0.
By Eq. (C.3), assuming that the ratio between pvT and pv is constant
in each model, we can demonstrate that a higher ratio between
Ncell
T and Ncell causes a large change in coarse aggregate content after
edge modification. Thus, edge modification might significantly
influence a simulation model which has many cells to treat, especially
for samples with thin cross-sections.
References
[1] C. Zhang, X. Tang, Y. Zhou, H. Jiang, State-of-the-art literature review on concrete
ðC:2Þ meso-scale mechanics, J. Hydroelectr. Eng. 34 (12) (2015) 1–17.
178 X. Yang et al. / Powder Technology 320 (2017) 161–178
[2] J. Feng, W. Sun, Z. Liu, C. Cui, X. Wang, An armour-piercing projectile penetration in a
double-layered target of ultra-high-performance fiber reinforced concrete and
armour steel: experimental and numerical analyses, Mater. Des. 102 (2016)
131–141.
[3] C. Qin, C.Q. Guo, C.H. Zhang, A pre-processing scheme based on background grid
approach for meso-concrete mechanics, J. Hydraul. Eng. 42 (2011) 99–106.
[4] R. Klein, Concrete and Abstract Voronoi Diagrams, Springer-Verlag, Berlin, 1989.
[5] A. Caballero, C.M. López, I. Carol, 3D meso-structural analysis of concrete specimens
under uniaxial tension, Comput. Methods Appl. Mech. Eng. 195 (52) (2006)
7182–7195.
[6] Z.L. Wang, F. Lin, X.L. Gu, Numerical simulation of failure process of concrete under
compression based on mesoscopic discrete element model, Tsinghua Sci. Technol.
S1 (2008) 19–25.
[7] L. Wang, Meso-scale numerical modeling of the mechanical behavior of reinforced
concrete members, Int. J. Emerg. Technol. Learn. (2013) 680–684.
[8] Z.M. Wang, A.K.H. Kwan, H.C. Chan, Mesoscopic study of concrete I generation of
random aggregate structure and finite element mesh, Comput. Struct. (1999)
533–544.
[9] Z. Wu, B. Chen, N. Liu, Fabrication and compressive properties of expanded polysty-
rene foamed concrete: experimental research and modeling, J. Shanghai Jiaotong
Univ. 18 (1) (2013) 61–69.
[10] L.P.B. Leite, V. Slowik, H. Mihashi, Computer simulation of fracture processes of con-
crete using mesolevel models of lattice structures, Cem. Concr. Res. 34 (6) (2004)
1025–1033.
[11] X. Tang, C. Zhang, J. Shi, A multiphase mesostructure mechanics approach to the
study of the fracture-damage behavior of concrete, Sci. China Ser. E Technol. Sci.
51 (S2) (2009) 8–24.
[12] A.H.J.M. Vervuurt, A. Luttikholt, Failure behavior of segmented tunnel linings: nu-
merical modeling validated by full scale tests, Eccomas Thematic Conference on
Computational Methods in Tunneling, Vienna, Austria 27–29 August 2007, pp. 1–13.
[13] D.P. Bentz, Three-dimensional computer simulation of Portland cement hydration
and microstructure development, J. Am. Cer. Soc. 80 (1) (1997) 3–21.
[14] D.N. Winslow, M.D. Cohen, D.P. Bentz, K.A. Snyder, E.J. Garboczi, Percolation and
pore structure in mortars and concrete, Cem. Concr. Res. 24 (1) (1994) 25–37.
[15] K.J. Falconer, Fractal Geometry: Mathematical Foundations and Applications, Wiley,
Chichester, UK, 1990.
[16] X. Yang, F. Wang, X. Yang, Q. Zhou, Fractal-dimension in concrete and implementa-
tion for meso-simulation, Constr. Build. Mater. 143 (2017) 464–472.
[17] X. Yang, F. Wang, Random-fractal-method-based generation of meso-model for
concrete aggregates, Powder Technol. 284 (2015) 63–77.
[18] L.T. Mo, M. Huurman, S.P. Wu, A.A.A. Molenaar, 2D and 3D meso-scale finite element
models for ravelling analysis of porous asphalt concrete, Finite Elem. Anal. Des. 44
(4) (2008) 186–196.
[19] J. Wang, Q. Li, L. Qing, J. Guan, Studies on representative volume element size of
concrete based on meso-structure statistics, Eng. Mech. 29 (12) (2012) 1–6.
[20] J. Wang, Study on Simulation and Application Based on Meso-mechanics, China
Architecture and Building Press, Beijing, 2015.
[21] M. Stroeven, H. Askes, L.J. Sluys, Numerical determination of representative volumes
for granular materials, Comput. Methods Appl. Mech. Eng. 193 (30) (2004)
3221–3238.
[22] S. Dasgupta, C. Papadimitriou, Algorithms, fourth ed. Addison Wesley, Boston, 2011.
[23] G. Ausiello, P. Crescenzi, G. Gambosi, V. Kann, A. Marchetti-Spaccamela, M. Protasi,
Complexity and Approximation: Combinatorial Optimization Problems and Their
Approximability Properties, Springer-Verlag, Berlin, 2012.
[24] D.L. Shell, A high-speed sorting procedure, Commun. ACM 2 (7) (1959) 30–32.
[25] T. Jiang, M. Li, P. Vitányi, A lower bound on the average-case complexity of Shellsort,
J. ACM 47 (5) (2000) 905–911.
[26] O. Astrachan, Bubble sort: an archaeological algorithmic analysis, ACM SIGCSE Bull.
35 (1) (2003) 1–5.
[27] M.A. Bender, M. Farach-Colton, M.A. Mosteiro, Insertion sort is O(n log n), Theory
Comput. Syst. 39 (3) (2006) 391–397.
[28] J. Bentley, Programming Pearl, second ed. Pearson Education, India, 2000.
[29] X. Gu, L. Hong, Z. Wang, F. Lin, Experimental study and application of mechanical
properties for the interface between cobblestone aggregate and mortar in concrete,
Constr. Build. Mater. (2013) 156–166.
[30] C. Du, L. Sun, S. Jiang, Z. Ying, Numerical simulation of aggregate shapes of three-
dimensional concrete and its applications, J. Aerosp. Eng. 26 (3) (2007) 172–178.