Engineering Applications of Artificial Intelligence 115 (2022) 105255

Contents lists available at ScienceDirect

Engineering Applications of Artificial Intelligence

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

Short communication

Determining ultimate bearing capacity of shallow foundations by using multi

expression programming (MEP)
Ruiliang Zhang, Xinhua Xue ∗
State Key Laboratory of Hydraulics and Mountain River Engineering, College of Water Resource and Hydropower, Sichuan University, Chengdu 610065, PR China

ARTICLE INFO ABSTRACT

Keywords: This study presents an artificial intelligence approach, namely multi expression programming (MEP), for
Ultimate bearing capacity determining ultimate bearing capacity of shallow foundations on cohesionless soils. Five governing parameters
Shallow foundations (i.e., internal friction angle, soil unit weight, the length to width ratio of foundation, foundation depth and
Multi expression programming
foundation width) were used as input variables to develop the MEP model. Through the determination of
Cohesionless soil
the optimal parameter setting of MEP, a group of expressions were proposed. Then, the MEP model was
compared with linear multiple regression, non-linear multiple regression and several previous models, and
three statistical indices (i.e., coefficient of determination (𝑅2 ), root mean squared error (RMSE) and mean
absolute error (MAE)) were employed to evaluate the prediction accuracy of these models. The results show
that the proposed model has higher prediction precision than the other models, with higher 𝑅2 value and lower
RMSE and MAE values. Additionally, a monotonicity analysis was performed to verify the correct relationship
between ultimate bearing capacity and various factors. From the monotonicity analysis, the ultimate bearing
capacity increases with the increase of internal friction angle (𝜑), soil unit weight (𝛾), foundation width (B)
and foundation depth (D), whereas it decreases with the increase of the length to width ratio of foundation
(L/B). Then, a sensitivity analysis was performed. Through the sensitivity analysis, the effect rank of the five
input parameters on ultimate bearing capacity is 𝜑> B > D > 𝛾> L/B. Finally, a graphical user interface (GUI)
of the MEP model is developed for practical application.

1. Introduction for estimating ultimate bearing capacity of shallow foundations, which

was applied to a strip footing with vertical central load, as shown in
A foundation can transmit loads directly to the underlying soil. If Eq. (1).
the soil near the surface can adequately support the structural loads, 1
it is possible to use shallow foundations (Craig, 2004). The ultimate 𝑃𝑢 = 𝑐𝑁𝑐 + 𝛾𝐷𝑁𝑞 + 𝛾𝐵𝑁𝛾 (1)
2
bearing capacity refers to the maximum carrying potential per unit
where 𝑃𝑢 is ultimate bearing capacity, c denotes cohesion of the soil, 𝛾
area of the foundation soil prior to foundation instability, which is an
is soil unit weight, B and D are width and depth of foundation, and 𝑁𝑐 ,
important index for foundation design and foundation stability eval-
𝑁𝑞 and 𝑁𝛾 are the cohesion, surcharge and density factors, respectively.
uation. Determining ultimate bearing capacity of shallow foundations
has always been one of the most important research topics in the On the basis of Terzaghi’s equation, Meyerhof (1963), Hansen
field of geotechnical engineering. Since the bearing capacity is affected (1970) and Vesic (1973) proposed the same modified form, and made
by many factors, such as the characteristics of the soil, foundation different corrections to the bearing capacity factors, as shown in
depth and foundation shape, it is difficult and complicated to estimate Eq. (2).
ultimate bearing capacity. Therefore, this topic has attracted many
𝑃𝑢 = 𝑠𝑐 𝑑𝑐 𝑐𝑁𝑐 + 𝑠𝑞 𝑑𝑞 𝛾𝐷𝑁𝑞 + 0.5𝑠𝛾 𝑑𝛾 𝛾𝐵𝑁𝛾 (2)
researchers’ attention in the past century.
Prandtl (1921) and Reissner (1924) firstly proposed the ultimate where 𝑠𝑐 , 𝑠𝑞 and 𝑠𝛾 are the shape factors and 𝑑𝑐 , 𝑑𝑞 and 𝑑𝛾 are the depth
bearing capacity calculation theories based on the limit equilibrium factors.
method. However, they ignored the weight of the soil and the friction Most bearing capacity theories are based on the limit equilibrium
between the foundation and soil. After that, many researches were method, and the obtained bearing capacities are validated by labo-
conducted to present different calculation formulas based on diverse as- ratory studies by numerous researchers. However, the experimental
sumptions. Terzaghi (1943) was the first to propose a generic equation researches are generally conducted on smaller sized models, which

∗ Corresponding author.
E-mail address: [email protected] (X. Xue).

https://doi.org/10.1016/j.engappai.2022.105255
Received 18 February 2022; Received in revised form 13 July 2022; Accepted 19 July 2022
Available online 2 August 2022
0952-1976/© 2022 Elsevier Ltd. All rights reserved.
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

( )
are highly scaled down models compared to real footings (Padmini ⎧1 + 0.1 tan 45 + 𝜑 ⋅ 𝐷 , 𝜑 > 10
⎪ 2 𝐵
et al., 2008). The size of the foundation has great influence on ultimate 𝑑𝛾 = 𝑑𝑞 = ⎨ (8)
bearing capacity. Therefore, small models of footings prepared in a ⎪1, 𝜑=0
⎩
laboratory are different from real size footings with regard to behavior
and stress distribution, called the scale effect (Kalinli et al., 2011). Hansen (1970) calculated the shape and depth factors using the
A simple method to solve the scale effect is to use a larger footing, expressions in Eqs. (9)–(12).
however, it is expensive, time consuming and experimentally difficult.
𝐵
Therefore, there is an increasing tendency to develop some artificial 𝑠𝑞 = 1 + sin 𝜑 (9)
𝐿
intelligence (AI) models to accurately predict ultimate bearing capacity 𝐵
of shallow foundations. These AI techniques include artificial neural 𝑠𝛾 = 1 − 0.4 (10)
𝐿
network (Padmini et al., 2008; Kalinli et al., 2011), random forest ⎧1 + 2 tan 𝜑 (1 − sin 𝜑)2 ⋅ 𝐷 𝐷
for ≤1
(Kohestani et al., 2017), genetic programming (Sadrossadat et al., 2014; ⎪ 𝐵 𝐵
𝑑𝑞 = ⎨ 𝐷 𝐷 (11)
Shahnazari and Tutunchian, 2012; Tsai et al., 2012), gene expression ⎪1 + 2 tan 𝜑 (1 − sin 𝜑)2 arctan for >1
programming (Tahmasebi et al., 2015), hybrid least squares support ⎩ 𝐵 𝐵
vector machine (Xue and Chen, 2019) and M5’ model tree (Khor- 𝑑𝛾 = 1 (12)
rami et al., 2020). However, the ANN model may face some issues,
such as slow convergence rates and convergence to local minima. In Vesic (1973) developed a new expression for determining 𝑠𝑞 , as
addition, most existing models use insufficient data. Similar to ge- shown in Eq. (13). The other forms of the shape and depth factors are
netic programming and gene expression programming, multi expression the same as Hansen’s equations.
programming (MEP) has proven to be an alternative and efficient
technique that uses a linear representation of chromosomes, which can 𝐵
𝑠𝑞 = 1 + tan 𝜑 (13)
encode computationally complex problems (Oltean and Dumitrescu, 𝐿
2002). The high prediction accuracy of genetic programming and gene Shahnazari and Tutunchian (2012) proposed a mathematical for-
expression programming in this field proves the feasibility of genetic mula based on multigene genetic programming, a soft computing
algorithms. Thus, with larger data scale, multi expression programming method, to determine ultimate bearing capacity of shallow foundations
can be chosen as a proper method for predicting ultimate bearing
on granular soils, as shown in Eq. (14).
capacity.
In recent years, the MEP technique has been applied widely in the ( )2 10−8 𝐷2 𝜑6 (𝐵 − 𝛾)2
𝐿
𝑃𝑢 = 2 × 10−12 𝜑7 (𝐵 + 𝐷) +𝜑 + (14)
field of geotechnical engineering, such as soil compaction parameters 𝐵 𝐿
𝐵
prediction (Wang and Yin, 2020), consolidation assessment (Sharifi
et al., 2020), soil deformation modulus estimation (Alavi et al., 2012) Tsai et al. (2012) presented three genetic programming techniques,
and soil classification (Alavi et al., 2009). The objective of this study i.e., genetic programming (GP), weighted genetic programming (WGP),
is to present a multi expression programming technique for estimating and soft-computing polynomials (SCP), to predict ultimate bearing
ultimate bearing capacity of shallow foundations on cohesionless soils. capacity of shallow foundations. According to the results, several em-
A large and reliable database was employed to develop the MEP
pirical formulae can be obtained. The most accurate one is presented
model. The MEP model was compared with several previous models
in the following expression:
(e.g., classical theories and AI formulae), linear multiple regression and
( )
non-linear multiple regression, and three statistical indices were used 2.88 − 𝜑
𝑃𝑢 = e0.134𝜑 + + (log 𝐷 + 𝐵𝜑)log(6.77+𝛾) (15)
to evaluate the performance of these models. Then, a monotonicity 𝐷𝛾
and sensitivity analysis was conducted to verify the prediction model.
Finally, a graphical user interface (GUI) was developed for practical Sadrossadat et al. (2014) developed a new design equation for
application. prediction of ultimate bearing capacity based on linear genetic pro-
gramming (LGP) approach, as shown in Eq. (16).
2. Existing empirical equations for ultimate bearing capacity (( ) )
3.95 (𝜑 − 35)2 2
𝑃𝑢 = 𝜑 𝛾 +𝜑+ 𝐿
𝐷 + 2.5 (𝐵 (𝜑 − 35) + 1) . (16)
Numerous explicit expressions have been proposed for determin- 𝐵
ing ultimate bearing capacity. As stated above, the classical modi-
fied equation was developed by Meyerhof (1963), Hansen (1970) and Tahmasebi et al. (2015) used gene expression programming (GEP)
Vesic (1973). They determined the surcharge factor 𝑁𝑞 using the same technique to determine ultimate bearing capacity of shallow founda-
following equation, which was presented by (Prandtl, 1921). tions on cohesionless soils. Two modified expressions, namely GEP-
( 𝜑 ) 𝜋 tan 𝜑 Meyerhof and GEP-Vesic, were developed, as presented in Eqs. (17) and
𝑁𝑞 = tan2 45 + 𝑒 (3) (18).
2
where 𝜑 is internal friction angle. ( )( )
𝑃𝑢 = −3.047149𝑁𝑞 𝐷 + 𝐵 −9.230316𝐷𝐹𝑠2 + 𝛾
However, they presented various expressions for the density factor ( ( ( )))
𝑁𝛾 , as shown in Eqs. (4)–(6). + (sin (sin (𝐵)))2 𝛾𝑁𝛾 + 𝛾 2 − 𝑁𝛾
( ) ( ) ( ( ))
𝑁𝛾 = 𝑁𝑞 − 1 tan (1.4𝜑) (4) +𝐹𝑑 tan tan (tan (𝐷𝐹 ))3 + 𝑁𝛾 + 9.230316 tan tan cos 𝐷 − 𝑁𝛾
( ) (17)
𝑁𝛾 = 1.5 𝑁𝑞 − 1 tan 𝜑 (5) ( )
( )
( ) 2 𝐿( ) ( )2 𝐷 𝑁𝛾 + 9.9964
𝑁𝛾 = 2 𝑁𝑞 + 1 tan 𝜑 (6) 𝑃𝑢 = 𝑠𝑞 + 𝐷 𝐵 𝑁𝛾 − 8.1403 + 𝐷 8.1178𝑑𝛾 𝐿
𝐵
𝐵
Meyerhof (1963) presented the shape and depth factors in the
following expressions: + 𝑁𝛾 − 9.0498𝐵 4 𝑑𝑞4 − 19.8444 (18)
( )
⎧1 + 0.1 tan2 45 + 𝜑 ⋅ 𝐵 , 𝜑 > 10 Khorrami et al. (2020) developed new expressions to predict ulti-
⎪ 2 𝐿
𝑆𝛾 = 𝑆𝑞 = ⎨ (7) mate bearing capacity using M5’ model tree based on a more compre-
⎪1, 𝜑=0 hensive database in contrast to previous studies. These expressions are
⎩

2
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Fig. 1. Box plots of input parameters.

as follows: Prior to any modeling procedure, it is necessary to assess original

( )0.37
𝐵 𝐵 database in terms of presence or absence of outliers (Faradonbeh and
For ≤ 0.15, 𝑃𝑢 = 0.0017 (𝐷 + 1)5.31 𝛾 4.78 𝜑3.54
𝐿+1 𝐿+1 Monjezi, 2017; Faradonbeh and Taheri, 2019). The presence of outliers
𝐵 negatively affects the accuracy and reliability of the proposed model.
For > 0.15 and 𝜑 ≤ 0.7,
𝐿+1 In this regard, the boxplot was employed to evaluate the existence of
( )0.2
𝐵 outliers. The box plot is a common method to display the distribution of
𝑃𝑢 = 78.18 (𝐷 + 1)3.05 𝛾 0.82 𝜑2.02 (19)
𝐿+1 data based on minimum, first quartile (Q1), median (Q2), third quartile
𝐵 (Q3), and maximum values (Faradonbeh
For
𝐿+1
> 0.15 and 𝜑 > 0.7, ( ( et al.,) 2022). (In the boxplot,
))
( )0.6 the data outside the range of 𝑄1 − 3 𝑄3 − 𝑄1 , 𝑄3 + 3 𝑄3 − 𝑄1 are
𝐵
𝑃𝑢 = 3133.79 (𝐷 + 1)3.31 𝛾 0.18 𝜑4.68 extreme outliers. The boxplots of all input parameters are shown in
𝐿+1
Fig. 1. As can be seen from Fig. 1, foundation width (B), foundation
depth (D) and ultimate bearing capacity (𝑃𝑢 ) have some outliers. With
3. Database
the exception of D and B, it will cover a small range. Therefore, only the
outliers of ultimate bearing capacity are removed in this study. Based
A large and reliable database collected by Khorrami et al. (2020) on Fig. 1, the median is not in the center of boxes, indicating that the
was employed to conduct an MEP-based model for predicting ultimate input parameters do not have a symmetric distribution.
bearing capacity of shallow foundations on cohesionless soils, which Fig. 2 depicts the frequency of each input variable. From Fig. 2, the
was published by Golder et al. (1941), Eastwood (1951), Subrah- internal friction angle ranges from 0.55 rad to 0.80 rad and the total
manyam (1967), Muhs et al. (1969), Muhs and Weiz (1971, 1973), specimens have soil unit weight from 9.85 kN∕m3 to 20.8 kN∕m3 . In
Weiz (1970), Briaud and Gibbens (1997, 1999), Gandhi (2003), Cerato addition, the length to width ratio of foundation ranges from 1 to 6 and
and Lutenegger (2007) and Akbas and Kulhawy (2009). The data in- the majority of the specimens have foundation width from 0 m to1 m.
clude load test data on real sized foundations and the corresponding in- For foundation depth, it mainly locates in the range between 0 m and
formation regarding the footing and soil. Totally, 65 data are obtained 0.3 m. Fig. 2 also shows the probability model of each variable. Among
through the small-scale experiments and 104 data from the large- them, the internal friction angle approximately follows a normal dis-
scale experiments (Khorrami et al., 2020). In the case of large-scale tribution, and the probability model of soil unit weight is Weibull
experiments, the ultimate load is defined as the load corresponding to distribution. The length to width ratio of foundation, foundation width
the point where the slope of the load settlement curve is a minimum. and foundation depth all follow a lognormal distribution.
For small-scale experiments, it is defined as the load corresponding Moreover, the correlation between inputs is important. To pre-
vent multicollinearity between the independent variables, a Pearson
to the point of break of the load settlement curve in a log–log plot
analysis can be conducted to obtain the correlation matrix (Faradon-
(Padmini et al., 2008).
beh et al., 2020). Table 1 lists the Pearson’s correlation coefficients
In this study, five governing parameters, i.e., internal friction angle
between different parameters. According to this table, there is no
(𝜑), foundation width (B), foundation depth (D), length to width ratio significant correlation between independent variables, and thus the
of foundation (L/B) and soil unit weight (𝛾), were selected as inputs, database was suitable for subsequent analysis. Additionally, principal
while ultimate bearing capacity (𝑃𝑢 ) was the output. That is because component analysis (PCA) was employed to study the multicollinear-
internal friction angle, soil unit weight and the foundation geometry ity phenomenon. Fig. 3 shows the Pareto chart of PCs against the
are considered as effective factors for determining ultimate bearing eigenvalues of them. As observed from Fig. 3, 84% variability of the
capacity in many studies (Padmini et al., 2008; Kalinli et al., 2011; database can be explained by three first PCs. The scatter plots of PC1–
Kohestani et al., 2017; Tahmasebi et al., 2015; Xue and Chen, 2019; PC2 and PC1–PC3 are shown in Fig. 4. According to the above results,
Khorrami et al., 2020). the dataset can be used to develop a model.

3
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Fig. 2. Histograms of the input parameters.

In the present study, totally 158 data are divided into two datasets (Deng et al., 2013). To solve this problem, graph-based and linear-based
of training (106 data) and testing (52 data). Note that the training and GPs are proposed. As a representative linear-based GP method, multi
testing sets should have similar statistical descriptions. Table 2 shows expression programming (MEP) has higher accuracy and efficiency in
the maximum, minimum, average and median values of training and comparison with other GPs. A significant MEP feature is the ability of
testing data. encoding multiple solutions of a problem into a chromosome. The best
solution is selected by the fitness assessment.
4. Multi expression programming (MEP) model There are several major steps to develop a best solution for a
program, which are summarized as follows (Wang and Yin, 2020;
4.1. Overview of multi expression programming Sharifi et al., 2020; Oltean, 2006):
(1) A population is set up with randomly generated programs.
Genetic programming (GP) is one of the branches of evolutionary al- (2) Two parents are chosen using a binary tournament.
gorithms. Early GP is achieved with tree-based structures, and operated (3) Two offspring are obtained by crossover and mutation over the
directly on the tree, which has the low efficiency of function mining parents.

4
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Table 1
Pearson’s correlation coefficients between different parameters.
𝑃𝑢 𝜑 𝛾 L/B D B
𝑃𝑢 1 – – – – –
𝜑 0.179* 1 – – – –
𝛾 0.171* 0.430** 1 – – –
L/B −0.309** −0.068 −0.117 1 – –
D 0.420** −0.403** −0.286** 0.029 1 –
B 0.697** −0.288** 0.001 −0.342** 0.605** 1

Note:
*Indicates significant correlation at the 0.05 level.
**Indicates significant correlation at the 0.01 level.

Fig. 3. Pareto chart of PC.

Table 2
Statistical description of the input data.
Parameter Minimum Maximum Average Median
Train Test Train Test Train Test Train Test
𝜑 (rad) 0.56 0.56 0.80 0.80 0.68 0.69 0.68 0.70
𝛾 (kN/m3 ) 9.85 10.2 20.8 20.8 15.46 16.31 16.1 16.65
L/B 1 1 6 6 2.22 2.28 1 1
D (m) 0 0 0.89 0.76 0.10 0.08 0 0
B (m) 0.03 0.04 3.02 3.00 0.53 0.49 0.5 0.31 Fig. 4. Scatter plots of PC1–PC2 and PC1–PC3..
𝑃𝑢 (kPa) 14 59 1277.78 1158 388.5 383.8 290.5 287.8

respectively. Gene 2 indicates the operation + on the operands at the

locations 0 and 1 of the chromosome. Therefore, gene 2 encodes the
(4) The worst individual is replaced if the offspring is better than expression 𝐸2 = 𝑥1 + 𝑥2 . Gene 4 indicates the operation × on the
the worst individual in the tournament. operands at the locations 2 and( 3 of the ) chromosome. Thus, gene
(5) The above steps are repeated until convergence. 4 encodes the expression 𝐸4 = 𝑥1 + 𝑥2 × 𝑥3 . Gene 6 indicates the
The representation of the MEP solutions is similar to the procedure operation/ on the operands at the locations 4 ((
and 5 of the
) chromosome.
)
in which C and Pascal compliers convert expressions into machine code Hence, gene 6 encodes the expression 𝐸6 = 𝑥1 + 𝑥2 × 𝑥3 ∕𝑥4 . Each
chromosome of MEP encodes a number of expressions, which are equal
(Gandomi et al., 2015). The MEP genes are represented by substrings.
to the chromosome length. Due to its multi expression representation,
The number of genes defines the chromosome length (code length),
each chromosome can be described as a forest of trees (Fig. 5). The best
which is unchanged during the calculation. Each gene is composed of
expression is selected by evaluating the fitness of all expressions in a
a terminal (an element in the terminal set T ) or a function symbol
MEP chromosome.
(an element in the function set F ). A gene that encodes a function
includes pointers towards the function arguments. Function arguments
4.2. MEP model development
always have indices of lower values than the position of the function
itself in the chromosome (Oltean, 2006). Note that the first gene of a According to previous study (Oltean, 2006), several code parame-
chromosome must be a terminal to ensure the validity of a program. ters must be set for a problem, namely, number of generation, popula-
An example of a MEP chromosome is illustrated below by using the
{ } tion size, code length, crossover probability, and mutation probability.
terminal set 𝑇 = 𝑥1 , 𝑥2 , 𝑥3 , 𝑥4 and the function set 𝐹 = {+, ×, ∕}: The number of generation is the calculation times before the run
0: 𝑥1 of a program terminates, whereas the population size defines the
1: 𝑥2 number of programs (Wang and Yin, 2020). As mentioned above,
2: +0, 1 the code length defines the number of genes in a chromosome. The
3: 𝑥3 crossover/recombination (e.g., one-point recombination, two-point re-
4: ×2, 3 combination and uniform recombination) and mutation are two main
5: 𝑥4 search operators of MEP algorithm. The offspring genes are randomly
6: /4, 5 taken from one parent to another by uniform recombination. More
By parsing the chromosome top-down, the MEP genes are translated details of search operators can be found in Ref. (Oltean, 2006).
into computer programs. In this example, genes 0, 1, 3, and 5 encode To determine the optimal code parameters of the MEP model, the
simple terminal symbols as: 𝐸0 = 𝑥1 , 𝐸1 = 𝑥2 , 𝐸3 = 𝑥3 , 𝐸5 = 𝑥4 , trial and error strategy was employed in the present study. Firstly,

5
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Fig. 5. Expressions of a MEP chromosome represented by gene trees.

Table 3 explicit expressions with the code length of 50 have only half number
Parameter settings for MEP.
of the expressions compared to the case with the code length as 100
Parameter Setting (Wang and Yin, 2020). Therefore, the code length of 100 can be chosen
Number of generation 500, 1000, 2000 when focusing on precision; on the contrary, the code length of 50 can
Population size 500, 1000, 2000
be chosen to meet the simplicity of the model.
Chromosome length 50, 100
Crossover probability 0.1, 0.5, 0.7, 0.9 In this study, in combined consideration of the complexity and the
Crossover type Uniform precision of the prediction expressions, the code length of 50 was cho-
Mutation probability 0.01, 0.1, 0.5, 0.9 sen to develop the model. According to the selected code parameters,
Function set +, −, ∗, /, pow10, lg the optimal expressions can be obtained as follows:
( ) 2
⎛ 𝜑2 ⎞
Table 4 ⎜ − 𝜑 𝐴1 𝐴2 lg 𝐴4 − 𝐴3 𝐵 − 𝐴1 ⎟
𝐵 2𝜑 𝐿 𝜑2
Optimal parameter setting for MEP. 𝑃𝑢 = ⎜ ⎟ + 10 𝐴3 𝐵 − 𝐵 + 𝐵
⎜ 𝜑 ⎟
Parameter Setting ⎝ ⎠
Number of generation 2000
( )
𝐴 102𝜑 𝐴 𝐵
Population size 2000 + 𝛾 − 𝐴1 − 102𝜑 𝜑5 𝐴2 𝐴4 + 1
+ 𝐴1 + 𝐴3 − 3 𝐷 (21)
Chromosome length 50
𝐴4 𝜑 𝐴1
Crossover probability 0.9 in which 𝐴1 , 𝐴2 , 𝐴3 and 𝐴4 are parameters as:
Crossover type Uniform
Mutation probability 0.01 10𝜑 lg(10𝜑 𝜑2 )𝜑2
Function set +, −, ∗, /, pow10, lg
𝐴1 = 𝜑 (22)
𝐵
−1
𝐴2 = 10𝜑 𝜑2 + 𝐵 (23)
the initial code parameters of MEP (Table 3) were chosen according
to previous optimal values in other fields of geotechnical engineering 𝜑2
𝐴3 = + 102𝜑 (24)
(Wang and Yin, 2020; Sharifi et al., 2020; Alavi et al., 2012, 2009; 𝐵
Gandomi et al., 2015; Heshmati et al., 2008; Hossein et al., 2012; 𝐴4 = lg(10𝜑 𝜑2 ) (25)
Mohammadzadeh et al., 2014; Shahnazari et al., 2010). From the trial
and error calculations, the initial optimal code parameters with the
5. Results and discussion
lowest mean absolute error (MAE) value were selected, as shown in
Table 4. The MAE can be calculated as
5.1. Comparison between the MEP model and other models
1 ∑|
𝑛
𝑀𝐴𝐸 = 𝑂 − 𝑃𝑖 || (20)
𝑛 𝑖=1 | 𝑖 In the present study, three statistical indices, i.e., the coefficient
of determination (𝑅2 ), root mean squared error (RMSE), and mean
where 𝑂𝑖 and 𝑃𝑖 are the actual and predicted values for the 𝑖th data of
absolute error (MAE), were used to evaluate the prediction accuracy
n; 𝑂𝑖 is the average of the actual value. The MAE value is close to 0,
of the proposed MEP model. The expressions of 𝑅2 and RMSE can be
indicating a high accuracy.
written as
Then, a code parametric study was conducted. During this process, ∑𝑛 ( )2
the desired independent code parameter should be varied within a 𝑂𝑖 − 𝑃𝑖
𝑅2 = 1 − 𝑖=1( )2 (26)
reasonable range, while the other code parameters are set as the values ∑𝑛
𝑂 𝑖 − 𝑂𝑖
in Table 4. Fig. 6 shows the effect of each code parameter on the 𝑖=1
√
prediction accuracy of ultimate bearing capacity. Generally, the MAE ∑𝑛 ( )
𝑖=1 𝑂𝑖 − 𝑃𝑖
value decreases with the increase of number of generation, population 𝑅𝑀𝑆𝐸 = (27)
size, code length and crossover probability, indicating the prediction 𝑛
accuracy increases. The prediction accuracy decreases with the increase For a group of data, a higher value of 𝑅2 and lower values of RMSE
of mutation probability. and MAE indicate a higher precision of the model. The aforementioned
According to the code parametric study, the parameter setting in statistical values for the training and testing data are presented in
Table 4 is verified to have the highest accuracy of the proposed model. Fig. 7. As seen from Fig. 7, for the training and testing data of the
However, with the increase in the code length, the MEP algorithm will MEP model, the 𝑅2 values are 0.780 and 0.767, indicating the predicted
provide a more complex model with higher accuracy. This will affect results using the proposed model are in reasonable agreement with
the applicability of the proposed model in practice. Furthermore, the the actual values. Therefore, the proposed MEP model can reasonably

6
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Fig. 6. Effects of code parameters on the prediction accuracy.

𝐿
predict ultimate bearing capacity of shallow foundations on cohesion- 𝑞u = 4073.9 − 17754.2𝜙 + 154𝛾 − 108.2 + 3389.2𝐷 − 24.4𝐵
𝐵
less soils. ( )2
𝐿
+17459.3𝜙2 + 4.6𝛾 2 + 32.7 +
The linear multiple regression (MLR) is a common method to fit 𝐵
a linear equation between input and output parameters (Armaghani 𝐿
428.9𝐷2 − 21.6𝐵 2 − 327.6𝜙𝛾 + 247.9𝜙 (29)
et al., 2018). The non-linear multiple regression (MNLR) method can 𝐵
solve many non-linear problems. For comparison purpose, the two 𝐿
−4051.6𝜙𝐷 + 2961.6𝜙𝐵 − 17.3𝛾 +
methods were also employed in this study. The proposed equations are 𝐵
𝐿
shown as follows: 22.0𝛾𝐷 − 82.7𝛾𝐵 + 14.5 𝐷 − 59.5𝐿 − 501.1𝐷𝐵
𝐵
𝐿
𝑞u = −1426.2 + 2291.5𝜙 + 2.9𝛾 − 6.4 + 306.6𝐷 + 364.6𝐵 (28)
𝐵

7
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Table 5
Sensitivity analysis of the proposed model.
Inputs MAE (kPa)
𝜑, 𝛾, B, D and L/B 75.76
𝛾, B, D and L/B 113.70
𝜑, B, D and L/B 77.01
𝜑, 𝛾, D and L/B 111.9
𝜑, 𝛾, B and L/B 95.03
𝜑, 𝛾, B and D 76.95

5.2. Monotonicity analysis

To investigate whether the proposed MEP model can capture the

inherent relationship between ultimate bearing capacity and various
factors, a monotonicity analysis was conducted in the present study. As
such, one independent variable varies, while the other variables remain
constant and equal to their average values. Therefore, the values of
the input variables are considered as follows: 𝜑 = 0.68 rad, 𝛾 = 15.74
kN/m3 , L/𝐵 = 2.24, 𝐷 = 0.09 m, 𝐵 = 0.52 m. Through the monotonicity
analysis, the variation of the predicted ultimate bearing capacity versus
each input variable is plotted in Fig. 9.
As observed from Fig. 9, the ultimate bearing capacity increases
with the increase of internal friction angle (𝜑), soil unit weight (𝛾),
foundation width (B) and foundation depth (D), whereas it decreases
with the increase of the length to width ratio of foundation (L/B). Note
that the variation of ultimate bearing capacity is obvious with the in-
crease internal friction angle (𝜑), foundation width (B) and foundation
depth (D), while the impact of soil unit weight (𝛾) and the increase
of the length to width ratio of foundation (L/B) are low. However,
soil unit weight impacts effective friction angles, which strongly impact
Fig. 7. Comparison of the actual and predicted results for the MEP model: (a) training the ultimate bearing capacity. This may be because the proposed MEP
data; (b) testing data.
model is a data-driven model, and it depends on the sufficiency and
quality of data. Nevertheless, these monotonic variations can reflect
Additionally, previous approaches (e.g., Meyerhof (1963), Hansen the correct variation trend, indicating the proposed model can correctly
(1970), Vesic (1973), Shahnazari and Tutunchian (2012), Sadrossa- reflect the relationship between ultimate bearing capacity and various
dat et al. (2014), Tahmasebi et al. (2015), Khorrami et al. (2020)) factors.
for determining ultimate bearing capacity are taken into comparison
with the MEP model. Note that the model of Tsai et al. (2012) is 5.3. Sensitivity analysis
not applicable when the foundation depth is 0 m, and thus it is not
taken into comparison with other models. Fig. 8 depicts the prediction To investigate the significance of each input variable (i.e., 𝜑, 𝛾, B,
performance of previous models and the present models. D and L/B) on ultimate bearing capacity, a sensitivity analysis was
As seen from Fig. 8, the 𝑅2 values for Meyerhof (1963) and Tah- performed. This investigation is conducted by some sensitivity tests.
masebi et al. (2015) are smaller than 0.6, whereas the error values of As such, one of the five input parameters is removed, and the rest are
RMSE and MAE are large, indicating that the prediction accuracy is used in modeling (Zhang and Xue, 2021). Then, MAE is employed to
poor. The 𝑅2 values for Hansen (1970), Shahnazari and Tutunchian evaluate the new model. The Sensitivity analysis results are shown in
(2012) and MLR range from 0.6 to 0.7. Moreover, the 𝑅2 values Table 5.
for Vesic (1973), Sadrossadat et al. (2014), Khorrami et al. (2020), As can be seen from Table 5, the MEP model without 𝜑 has the
MNLR and MEP model are larger than 0.7, showing the prediction highest MAE values, indicating that the ability of predicting ultimate
precision is reasonable. Among all models, the MEP model presents bearing capacity is significantly degraded when 𝜑 is excluded. There-
the highest accuracy, with the 𝑅2 , RMSE and MAE values of 0.776, fore, it is reasonable to find that 𝜑 has great effect on ultimate bearing
142.52 and 75.76, whereas the values for Vesic (1973), Sadrossadat capacity. The MEP model without B has the second highest MAE value.
et al. (2014), Khorrami et al. (2020), and MNLR are 0.711, 167.73 In addition, the impact of 𝛾 and L/B on modeling is not significant. The
and 105.66, 0.745, 155.09 and 90.89, 0.747, 158.38 and 91.73 and results are similar to Khorrami et al. (2020), and it is in agreement with
0.771, 136.21, 90.58, respectively. Note that the MEP model and MNLR the results of monotonicity analysis. Generally, the effect rank of the
model have similar accuracy, indicating the MEP model can reflect five input parameters on ultimate bearing capacity is 𝜑 > B > D > 𝛾 >
the non-linear relationship between ultimate bearing capacity and the L/B.
inputs. Moreover, the precision of the MEP model can be improved by
increasing the code length. Therefore, it is necessary to go for more 5.4. Graphical user interface (GUI)
complex models such as MEP. Generally, the prediction performance
rank of all these approaches is MEP > MNLR > Khorrami et al. (2020) For practical application, a small software was developed based on
> Sadrossadat et al. (2014) > Vesic (1973) > MLR > Shahnazari and Visual Basic 6.0 software, and the graphical user interface (GUI) is
Tutunchian (2012) > Hansen (1970) > Tahmasebi et al. (2015) > shown in Fig. 10. The users can download the software for the practical
Meyerhof (1963). use.

8
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Fig. 8. The performance of all models.

6. Conclusions unit weight, the length to width ratio of foundation, foundation depth
and foundation width.
The optimal parameters for the MEP algorithm were determined
In the present study, multi expression programming (MEP) ap- firstly. Using the optimal parameter setting, a group of expressions
proach was employed to determine ultimate bearing capacity of shal- were developed to predict ultimate bearing capacity. The prediction
low foundations on cohesionless soils. The proposed model was devel- accuracy of the MEP model was evaluated by three statistical indices
oped by five governing parameters, namely, internal friction angle, soil (i.e., 𝑅2 , RMSE and MAE) and taken into comparison with several

9
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Fig. 8. (continued).

previous models in the literature, linear multiple regression and non- analysis, the ultimate bearing capacity increases with the increase of
linear multiple regression. According to the results, the MEP model has internal friction angle (𝜑), soil unit weight (𝛾), foundation width (B)
the highest prediction precision, with the 𝑅2 , RMSE and MAE values of and foundation depth (D), whereas it decreases with the increase of the
0.776, 142.52 and 75.76, indicating the proposed model can reasonably length to width ratio of foundation (L/B). Then, a sensitivity analysis
predict ultimate bearing capacity. was performed. Through the sensitivity analysis, the effect rank of the
A monotonicity analysis was conducted to investigate whether the five input parameters on ultimate bearing capacity is 𝜑 > B > D >
proposed MEP model can capture the inherent relationship between 𝛾 > L/B. Finally, a graphical user interface (GUI) of the MEP model
ultimate bearing capacity and various factors. From the monotonicity is developed for practical application.

10
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Fig. 9. Monotonicity analysis of the effect of each input variable on the output variable.

CRediT authorship contribution statement Declaration of competing interest

Ruiliang Zhang: Data acquisition, Writing – original draft, Soft- The authors declare that they have no known competing finan-
ware. Xinhua Xue: Conceptualization, Methodology, Writing – review cial interests or personal relationships that could have appeared to
& editing. influence the work reported in this paper.

11
R. Zhang and X. Xue Engineering Applications of Artificial Intelligence 115 (2022) 105255

Fig. 10. Graphical user interface (GUI).

Data availability statement Khorrami, R., Derakhshani, A., Moayedi, H., 2020. New explicit formulation for ultimate
bearing capacity of shallow foundations on granular soil using M5’ model tree.
Measurement. p. 163.
Some or all data, models, or code that support the findings of this Kohestani, V.R., Vosoughi, M., Hassanlourad, M., 2017. Bearing capacity of shallow
study are available from the corresponding author upon reasonable foundations on cohesionless soils: a random forest based approach. Civ. Eng.
request. Infrastruct. J.-Ceij 50, 35–49.
Meyerhof, G.G., 1963. Some recent research on the bearing capacity of foundations.
Can. Geotech. J. 1, 16–26.
References Mohammadzadeh, S.D., Bolouri, B.J., Alavi, A.H., 2014. An evolutionary computational
approach for formulation of compression index of fine-grained soils. Eng. Appl.
Artif. Intell. 33, 58–68.
Akbas, S.O., Kulhawy, F.H., 2009. Axial compression of footings in cohesionless soils.
Muhs, H., Elmiger, R., Weiz, K., 1969. Sohlreibung Und GrenztragfäHigkeit Unter
II: bearing capacity. J. Geotech. Geoenviron. Eng. 135, 1575–1582.
Lotrecht Und SchräG Belasteten Einzelfundamenten. Deutsche Forschungsge-
Alavi, A.H., Gandomi, A.H., Nejad, H.C., 2012. Design equations for prediction of
sellschaft für Bodenmechanik (DEGEBO), HEFT, Berlin, p. 62.
pressuremeter soil deformation moduli utilizing expression programming systems. Muhs, H., Weiz, K., 1971. Untersuchung Von GrenztragfäHigkeit Und Setzungsverhal-
Neural. Comput. Appl. 23, 1771–1786. ten Flachgegründeter Einzelfundamente Im Ungleichförmigennichtbindigen Boden.
Alavi, A.H., Gandomi, A.H., Sahab, M.G., 2009. Multi expression programming: a new Deutsche Forschungsgesellschaft für Bodenmechanik (DEGEBO), HEFT, Berlin, p.
approach to formulation of soil classification. Eng. Comput. 26, 111–118. 69.
Armaghani, D.J., Faradonbeh, R.S., Rezaei, H., Rashid, A.S.A., Amnieh, H.B., 2018. Muhs, H., Weiz, K., 1973. Inclined load tests on shallow strip footings. In: Proceedings
Settlement prediction of the rock-socketed piles through a new technique based on of the 8th International Conference on Soil Mechanism and Foundation Engineering,
gene expression programming. Neural Comput. Appl. 29, 1115–1125. II. pp. 173-179.
Briaud, J.L., Gibbens, R., 1997. Large Scale Load Tests and Data Base of Spread Footings Oltean, M., 2006. Multi Expression Programming. Technical Report, Babes-Bolyai
on Sand. Federal Highway Administration. University, Cluj-Napoca.
Briaud, J.L., Gibbens, R., 1999. Behavior of five large spread footings in sand. J. Oltean, M., Dumitrescu, D., 2002. Multi Expression Programming. Technical Report,
Geotech. Geoenviron. Eng. 125, 787–796. UBB-01-2002, Babes-Bolyai University, Cluj-Napoca.
Cerato, A.B., Lutenegger, A.J., 2007. Scale effects of shallow foundation bearing Padmini, D., Ilamparuthi, K., Sudheer, K.P., 2008. Ultimate bearing capacity prediction
capacity on granular material. J. Geotech. Geoenviron. Eng. 133, 1192–1202. of shallow foundations on cohesionless soils using neurofuzzy models. Comput.
Geotech. 35, 33–46.
Craig, R.F., 2004. Craig’s Soil Mechanics, seventh ed. Spon Press, USA and Canada.
Prandtl, L., 1921. Über die eindringungsfestigkeit (härte) plastischer baustoffe und
Deng, W., He, P., Huang, Z., 2013. Multi-expression based gene expression program-
die festigkeit von schneiden (on the penetrating strengths (hardness) of plastic
ming. In: Proceedings of 2013 Chinese Intelligent Automation Conference, pp.
construction materials and the strength of cutting edges). Z. Angew. Math. Mech.
439-448.
1, 15–20.
Eastwood, W., 1951. A comparison of the bearing power of footings on dry and Reissner, H., 1924. Zum erddruckproblem (concerning the earth-pressure problem). In:
inundated sand. Struct. Eng. 29 (332). 1st International Congress of Applied Mechanics. Delft, pp. 295–311.
Faradonbeh, R.S., Monjezi, M., 2017. Prediction and minimization of blast-induced Sadrossadat, E., Soltani, F., Mousavi, S.M., Marandi, S.M., Alavi, A.H., 2014. A new
ground vibration using two robust meta-heuristic algorithms. Eng. Comput. 33, design equation for prediction of ultimate bearing capacity of shallow foundation
835–851. on granular soils. J. Civ. Eng. Manage. 19, S78–S90.
Faradonbeh, R.S., Taheri, A., 2019. Long-term prediction of rockburst hazard in deep Shahnazari, H., Dehnavi, Y., Alavi, A.H., 2010. Numerical modeling of stress–strain
underground openings using three robust data mining techniques. Eng. Comput. behavior of sand under cyclic loading. Eng. Geol. 116, 53–72.
35, 659–675. Shahnazari, H., Tutunchian, M.A., 2012. Prediction of ultimate bearing capacity of
Faradonbeh, R.S., Taheri, A., Karakus, M., 2022. The propensity of the over-stressed shallow foundations on cohesionless soils: An evolutionary approach. KSCE J. Civ.
rock masses to different failure mechanisms based on a hybrid probabilistic Eng. 16, 950–957.
approach. Tunn. Undergr. Space Technol. 119, 104214-104226. Sharifi, S., Abrishami, S., Gandomi, A.H., 2020. Consolidation assessment using multi
expression programming. Appl. Soft. Comput. 86.
Faradonbeh, R.S., Taheri, A., Sousa, L.R.E., Karakus, M., 2020. Rockburst assessment in
Subrahmanyam, G., 1967. The effect of roughness of footings on bearing capacity. J.
deep geotechnical conditions using true-triaxial tests and data-driven approaches.
Int. Soc. Soil Mech. Found. Eng. 6, 33–45.
Int. J. Rock Mech. Min. Sci. 128, 104279-104298.
Tahmasebi, P.A., Barari, A., Behnia, M., 2015. Determination of the ultimate limit states
Gandhi, G., 2003. Study of Bearing Capacity Factors Developed from Lab. Experiments of shallow foundations using gene expression programming (GEP) approach. Soils
on Shallow Footings on Cohesionless Soils (Ph.D.). Found. 55, 650–659.
Gandomi, A.H., Faramarzifar, A., Rezaee, P.G., 2015. New design equations for elastic Terzaghi, K., 1943. Theoretical Soil Mechanics. John Wiley & Sons, New York.
modulus of concrete using multi expression programming. J. Civ. Eng. Manage. 21, Tsai, H.C., Tyan, Y.Y., Wu, Y.W., Lin, Y.H., 2012. Determining ultimate bearing capacity
761–774. of shallow foundations using a genetic programming system. Neural. Comput. Appl.
Golder, H., Fellenius, W., Kogler, F., 1941. The ultimate bearing pressure of rectangular 23, 2073–2084.
footings. J. Inst. Civ. Eng. 17, 161–174. Vesic, A.S., 1973. Analysis of ultimate loads of shallow foundations. J. Soil Mech.
Hansen, J.B., 1970. A Revised and Extended Formula for Bearing Capacity (Bulletin Found. Div. 99, 45–73.
No. 28). Danish Geotechnical Institute. Wang, H.L., Yin, Z.Y., 2020. High performance prediction of soil compaction parameters
Heshmati, A.A.R., Salehzade, H., Alavi, A.H., 2008. A multi expression programming using multi expression programming. Eng. Geol. 276.
Weiz, K., 1970. Der Einfluz Der Fundamentform Auf Die GrenztragfäHigkeit
application to high performance concrete. World Appl. Sci. J. 5, 215–223.
Flachgegründeter Fundamente. Deutsche Forschungsgesellschaft für Bodenmechanik
Hossein, A.A., Mollahasani, A., Hossein, G.A., 2012. Formulation of secant and reload-
(DEGEBO), HEFT, Berlin, p. 65.
ing soil deformation moduli using multi expression programming. Eng. Comput.
Xue, X., Chen, X., 2019. Determination of ultimate bearing capacity of shallow
29, 173–197.
foundations using lssvm algorithm. J. Civ. Eng. Manage. 25, 451–459.
Kalinli, A., Acar, M.C., Gündüz, Z., 2011. New approaches to determine the ultimate Zhang, R.L., Xue, X.H., 2021. A new model for prediction of soil thermal conductivity.
bearing capacity of shallow foundations based on artificial neural networks and Int. Commun. Heat Mass Transf. 129, 105661-105670.
ant colony optimization. Eng. Geol. 117, 29–38.

12