Optimal Design of Rigid Inclusions and Stone Columns

Using Multiobjective Optimization

Luan Nguyen1

Abstract: Ground improvement adopting stone columns or rigid inclusions has become popular in engineering practice as it can be more
economical under certain design criteria. In typical designs of either stone columns or rigid inclusions, the column diameter, column length,
and grid spacing are three main parameters that affect the costs and performance of the improved ground. This work is concerned with
achieving an optimal design for these two ground improvement methods. The design of the three main design parameters is subject to
optimization using a multiobjective optimization algorithm with respect to minimizing the construction cost of the soil improvement system.
We explain how the semiautomated design optimization works in connection with the considered calculation method used for the design
and how execution constraints, if required, can be incorporated in the design workflow. Through the use of Priebe’s method and
(ASIRI) recommendations for the calculation of stone columns and rigid inclusions, respectively, the usefulness of the presented design
optimization approach for each of the two ground improvement methods is demonstrated. DOI: 10.1061/(ASCE)GM.1943-5622.0002623.
© 2022 American Society of Civil Engineers.

Introduction the weak natural soil is not sufficient to act against lateral bulging
of the aggregate column.
The construction of heavy structures, such as buildings, liquid or RIs, also called controlled modulus columns, are different from
gas tanks, or embankments, on soft soils requires either deep foun- SCs in their relatively high stiffness elements. They are similar to
dations, such as piles, to transfer loads to deeper bearable soil de- unreinforced piles, regarding functionality. End bearing and skin
posits or improvement of the weak soil layers to increase their friction of the inclusions are to carry a large portion of the surcharge
bearing capacity. The latter family of methods, known as soil im- load, which is transferred to the inclusions via the arching effect oc-
provement techniques, is of concern in the presented work. One curring within the load transfer platform (LTP) above the column
commonly used method for soil improvement involves placing re- heads. The remaining surcharge load is carried by the soil itself.
inforcement in the soft soil layers to increase the bearing capacity The existence of a well-compacted granular LTP layer instead of
of the ground and, at the same time, reduce settlement. These a direct structural connection between piles and the ground slab dis-
soil improvement techniques include stone columns (SCs), rigid in- tinguishes RIs from piled raft foundations. The isolation between
clusions (RIs), jet grouting, and other deep soil mixing (DSM) tech- foundation elements and the superstructure provides protection
niques, such as cutter soil mixing (CSM) and mixed-in-place (MIP) mechanisms under earthquake loads (Pecker 2004). If necessary,
methods. In general, soil improvement is preferred over deep foun- the LTP layer can be reinforced by geotextiles or geogrids to resist
dations, as there are advantages in construction costs and time. loads induced by lateral spreading of, e.g., high earth dams or em-
Vibroreplacement or vibrodisplacement with SCs is commonly bankments (Klobe 2020). RIs are preferably used where the com-
used to improve noncompactable cohesive soil. The installed verti- petent soil stratum is not at a shallow depth or where the
cal SCs not only help in increasing bearing capacity and reducing cohesive soil is very or extremely soft.
settlement but may also act as vertical drains, which accelerate con- In the presented work, two design problems of ground improve-
solidation by allowing excess pore pressure to dissipate more easily. ment for SCs and RIs are studied. A critical aspect in designing such
Allowing pore pressure dissipation can also mitigate liquefaction in ground improvement systems is to choose the right combination of
granular soils (Seed and Booker 1977). Considerations for modeling, the design parameters (e.g., column length, grid pattern and spacing,
including use of the finite-element method of SCs, are discussed in column diameter) for a design solution to meet serviceability require-
Castro (2017). Advancements in numerical modeling with such ments (e.g., a prescribed level of absolute or differential settlements)
methods as finite difference or finite elements and discrete elements and to be cost-efficient. For RIs, the ultimate-limit-state design crite-
(Yoo and Kim 2009; Indraratna et al. 2015; Gu et al. 2017; Tan et al. ria (e.g., column strength) depend on the column type, column spac-
2021) have helped to better reveal the behavior of SCs in various ing, column diameter, and external loads. This sort of optimization
soils and subject to different loading conditions. SCs can be ordinary task can be tedious for a geotechnical designer because the manual
or encased by geotextiles (Simon 2012). An encasement of SCs is optimization work involves repeated evaluations of many feasible
needed when the host soil is very soft, i.e., the shear strength of combinations of the design parameters. Moreover, a manual iterative
search for an optimal cost-efficient solution often involves a great
1
BAUER Spezialtiefbau GmbH, BAUER-St. 1, 86529 Schrobenhausen, amount of work with little improvement and, therefore, is not
Germany. ORCID: https://orcid.org/0000-0001-9005-0172. Email: luan.th often systematically carried out in engineering practice. Experience-
.nguyen@gmail.com
based design can help greatly in finding a design solution which,
Note. This manuscript was submitted on January 22, 2022; approved
on July 30, 2022; published online on <PubMonth 00, 0000>. Discussion however, is often suboptimal.
period open until <DisMonth 00, 0000>; separate discussions must be sub- The importance of optimizing the design has been raised for sys-
mitted for individual papers. This paper is part of the International Journal tems with complex soil–structure interactions, such as piled raft
of Geomechanics, © ASCE, ISSN 1532-3641. foundation, by parametric studies (Cunha et al. 2001; Reul and

© ASCE 1 Int. J. Geomech.

Randolph 2004). Although not a common practice, numerical opti- framework for design optimization, in which uncertainties in soil
mization has been demonstrated to work well for geotechnical de- properties and wind speed were considered, Ravichandran and
sign. Meier (2019) used particle swarm optimization (PSO) to Shrestha (2020) used the NSGA-II to optimize a pile group and
optimize the positions and angles of ground anchors for the design piled raft foundations for a wind turbine. Furthermore, Leung
of an excavation pit supported by ground anchors. Phutthananon et al. (2017) proposed the use of multiobjective optimization ar-
et al. (2021) adopted a genetic algorithm to optimize the design rangements of piles in a piled raft while considering the interaction
of conventional and T-shape deep cement mixing columns for between the foundation and superstructure in the finite-element
the improvement of soft soil under an embankment using a re- model using the stiffness matrix condensation method. Using a
sponse surface method to build a surrogate model based on data generative design framework supported by metamodeling, Shen
points generated by finite-element evaluations. The use of the re- et al. (2022), has shown the potential of the approach for the design
sponse surface model accelerates the optimization run time. Chen of wind turbine foundations.
et al. (2021) proposed the use of a variant of a genetic algorithm In this paper, an optimization framework is proposed to solve
for the settlement-based cost optimization of a rigid inclusion sys- the problem of finding optimal design solutions of the two soil im-
tem, called a geogrid-reinforced pile-supported foundation, under provement systems by using numerical optimization. In particular,
an embankment. To achieve quick evaluations of the settlement, a multiobjective optimization strategy is employed to form a Pareto
the authors used a kriging-based surrogate model built by fitting front, which contains the Pareto-optimal solutions, for system per-
data generated by finite-element calculations. formance and the construction cost of the considered soil improve-
Many researchers have worked on algorithm-based design opti- ment systems. The design variables for both SCs and RIs are grid
mization of pile group and piled raft foundations. The interactions spacing, column depth, and column diameter. Compared with
between soil, piles, and the foundation slab, and the finite dimen- single-objective optimization, a multiobjective strategy explores
sions of the foundation, makes modeling of the full foundation sys- the search space based on the measures of at least two objectives.
tem necessary. Such sophisticated full system modeling provides, When the objectives compete with each other, the result of multi-
moreover, the possibility to subject a whole system to optimization. objective optimization is a Pareto front, which is a front of (near)
Leung et al. (2010) used a so-called pattern function to parameter- optimal solutions. The proposed optimization design framework
ize pile lengths subject to optimization. Kim et al. (2001) optimized is concerned with serviceability limit design, given that any ulti-
the arrangement of piles in piled raft foundations, with the objective mate limits are satisfied. With regard to the design of floating
of minimizing differential settlement. Chan et al. (2009) used a SCs in very loose sands and soft clays of very low undrained
modified genetic algorithm to optimize pile groups by minimizing shear strengths, the design engineer is to carry out bearing capacity
the total concrete volume of the pile group. Design solutions are checks for bulging and punching at the toes of the columns sepa-
subject to optimization constraints that respect serviceability re- rately, preferably using detailed numerical modeling. The consider-
quirements. Nakanishi and Takewaki (2013) optimized the total ation of the bearing capacity and checks using analytical
length of piles in a piled raft foundation using a sequential linear approaches can, in principle, be incorporated in the automated op-
programming method with an adaptive step length and controlled timization workflow as constraint rules.
the resulting settlement by applying constraint rules in the optimi-
zation process. Topology optimization was demonstrated for the
design optimization of shallow foundations (Seitz and Grabe Methods
2016). However, its use in practice is not yet well adopted because
of the need for unconventional methods for construction. With respect to optimization of the system performance in the de-
The work in the aforementioned articles follows a single- sign, we are concerned with the design for serviceability to calcu-
objective approach for optimization. Using this approach, one late the settlement. Therefore, the characteristic values for loads,
can either build the objective as a function of, e.g., concrete volume soil parameters, and soil–structure interaction parameters are used
or total pile length, which is to be minimized, while serviceability in the calculation without applying any safety factors.
limits, such as the total and differential settlements, are set as opti-
mization constraints. It is noted that the role of system performance
and construction cost can be exchanged: system performance (total Unit Cell Model
or differential settlement) as objective and construction cost (con- For both SCs and RIs, the calculation for a unit cell is considered for
crete volume or total pile length) as design constraint, or vice the evaluation of the effects of soil replacement and reinforcement to
versa. Since the relation between construction cost and system per- the improved soil layers. In the calculation domain, an attributable
formance is competitive by nature, it is more natural and advanta- volume is considered that includes the reinforcement and the soil sur-
geous to formulate the design optimization problem simultaneously rounding it. In the vertical direction, the model can extend beyond the
for a number of objectives. Kinzler et al. (2007) introduced, per- base of the reinforcement, as far as a considered limit depth. In the
haps for the first time, in detail, the use of multiobjective optimiza- horizontal direction, the model is extended over the column radius
tion in geotechnical design with an application for pile group and is limited at the boundaries of the attributable area, which de-
design. Deb and Dhar (2011) used a nondominated sorting genetic pends on the grid spacing of the reinforcements. It is assumed that,
algorithm for the optimization of SC design, and chose to optimize at the model’s vertical boundaries, the lateral displacement and
the maximal and differential settlements of the design by consider- shear stresses are zero. The limit depth is considered for the calcula-
ing the SCs’ elastic modulus and ultimate shear strength as design tion of settlement for both improved and unimproved soil layers
variables. Juang and Wang (2013) proposed a reliability-based con- under surcharge loading. The limit depth is set at the depth where
cept for geotechnical design, termed robust geotechnical design, the ratio between the increased stress caused by surcharge load,
and used a nondominated sorting genetic algorithm (NSGA-II) to Δσ − σ ′v , and the effective vertical stress caused by overburden and
drive a biobjective optimization. With an illustration for the design surcharge, Δσ + σ ′v , has been reduced by at least 10%. For the unit
of a spread foundation, the proposed framework was used to obtain cell model, the applied load is equivalent to the surcharge Δσ being
Pareto-optimal design solutions by minimizing both the construc- infinitely spread on the ground surface and, therefore, the whole ap-
tion cost and the probability of failure. In a similar robust plied surcharge load is decreased within the unit cell along the depth.

© ASCE 2 Int. J. Geomech.

For example, if a surface load Δσ = 50 kPa is acting on a submerged achieve a relatively high constrained modulus and angle of internal
soil with a unit weight under buoyancy γ′ = 10 kN/m3, using the friction. The arching effect developed within the LTP distributes
condition (Δσ − σ ′v )/(Δσ + σ ′v ) = 0.1 will result in a limit depth the structural load to both the RIs and the surrounding soil. The
Llimit = (Δσ − 0.1Δσ)/1.1γ′ = (50 − 0.1 × 50)/(1.1 × 10) ≈ 4 m. fraction of the load transmitted to the surrounding soil is captured
over the depth by the RIs through the development of negative skin
friction. The load distribution ratio depends on the ground and LTP
Design of SCs properties as well as the column diameter and spacing.
In engineering practice, the calculation method proposed by Priebe The analytical method presented in ASIRI (IREX 2012) is
(1995) is well accepted for the calculation of improvement factors, followed to calculate the behavior of the RIs by considering an axi-
owing to the installation of SCs in loose granular or soft cohesive symmetric unit cell model, which consists of a single rigid column
soil layers. According to Priebe (1995), the basic improvement fac- and the surrounding soil over a half grid spacing. This calculation
tor n0 is calculated from the reciprocal of the area replacement ratio method can be used to predict the mechanism of load transferring
A/Ac, where A and Ac are, respectively, the unit cell area and the from the structure, through the LTP, into the underground reinforced
area of a single SC in the unit cell, and the friction angle of a by rigid columns. Interactions at the interface between the soil and the
SC, φc. The effect of column compressibility is considered in the RIs, including the development of negative and positive skin frictions,
improvement factor n1 by adding to the ratio A/Ac and the incre- and between the inclusion base and the load-bearing soil are also con-
ment Δ(A/Ac). The value of Δ(A/Ac) depends on the ratio of stiff- sidered. Moreover, checking for the limit stress allowed in the LTP
nesses Esc /Ess , where Esc and Ess are, respectively, the constrained volume above the column head is part of the RI design in ASIRI.
modulus of the SC and that of the host soil. Finally, the contributing The load efficiency of a system of RIs is defined as the ratio
effect of earth pressure, which increases with depth, accounts for the between the load carried by the column heads and the total load. Con-
final improvement factor, n2 = fd × n1. The depth factor fd accounts sidering RIs under a concrete slab, the volume between the column
for the increase in the lateral support on the column with increasing head and the bottom of the slab determines the thickness of the LTP.
overburden pressure. It is noted that the depth factor fd must satisfy Insufficient thickness of the LTP prevents the arching effect from
the condition 1 ≤ fd ≤ (Esc /Ess )/(pc /ps ) (Priebe 1995), where pc and fully developing. Stress concentration in the LTP is, therefore, limited
ps are pressures caused by the column self-weight and the natural soil by the LTP’s thickness (Chevalier et al. 2011; Blanc et al. 2013).
at any considered improvement depth. Elastic settlement can then be Negative skin friction is developed where soft soil settles more than
calculated for the improved soil based on the constrained modulus the inclusion. The activation of negative skin friction through the rela-
improved by a factor of n2 for each of the considered soil layers. tive downward displacement of the soil with respect to the column
For an unlimited spread of the surcharge load, the loading stress is transfers load to the column. Below the depth at which the settlements
not reduced with increasing depth. Considering a large area of of the column and its surrounding soils are the same, the RI settles
ground improved with SCs, bearing capacity failure should be of more than the soil; this results in positive skin friction acting against
concern only at the edge of the surcharge loading. This issue is to the axial load on the column. At the inclusion base, the column settles
be separately checked by the design engineer. and pushes against the soil, activating a base resistance force acting up-
ward against the column. As a result, the mobilized skin friction and
based resistance are governed by how much the rigid column settles.
Design of RIs
A typical RI system consists of rigid vertical elements arranged in a
Multiobjective Optimization
grid with certain spacing to a depth required for load-bearing ca-
pacity and a LTP between the structure and the rigid members. The design of an engineering system such as a geotechnical struc-
The LTP is commonly a layer of densely compacted sand, to ture is a multiobjective decision making problem (Chankong and

Fig. 1. Foundation design workflow assisted by a multiobjective optimization algorithm.

Table 1. Soil strata and soil parameters

Parameter Soft clay Loose sand Soft clay Medium dense sand
Layer top below ground surface (m) 0 7 11 15
Dry unit weight γunsat (kN/m3) 15 16 15 17
Saturated unit weight γsat (kN/m3) 16.5 18 16 19
Oedometer modulus Eoed (kN/m2) 2,000 9,000 5,000 35,000
Poisson’s ratio ν (−) 0.3 0.3 0.3 0.3

© ASCE 3 Int. J. Geomech.

Haimes 2008), in which the interplay between system performance that have analogy with the biological processes of selection, muta-
and the associated construction cost must be considered. We are tion, and crossover. Based on an idea of Schaffer (1985), Srinivas
concerned with multiobjective optimization problems when more and Deb (1994) introduced nondominated sorting at the selection
than one objective is involved in an optimal decision making pro- step of the genetic algorithm to avoid convergence bias of the op-
cess. Unlike single-objective optimization, there exists no single timal solutions to one or more Pareto-optimal solutions.
best solution with respect to all objectives in a multiobjective opti- Deb et al. (2002) improved the convergence speed of the NSGA
mization problem. A best solution to a multiobjective problem is by using a fast nondominated sorting method and further intro-
available conditioned on other objectives. Although one can duced an elitist-preserving approach and a so-called parameterless
apply a scalarization rule to get a single-objective function and niching operator. These additional measures help to better spread
then use a single-objective algorithm to solve the optimization the nondominated Pareto-optimal solutions while maintaining a rel-
problem, doing so requires good understanding of the relation of atively low computational complexity of O(n2 m), where n and m
one objective to the others to set up a kind of summation function. are, respectively, the number of design variables and objectives.
This is not a trivial task as there is often no clear relation between In this work, the nondominated sorting genetic algorithm
the objectives. Even if the mentioned summation rule can be estab- NSGA-II (Deb et al. 2002) is used to perform the optimization
lished, the summation of many single objectives is likely to result in task. As it is based on a genetic algorithm, NSGA-II can handle
a highly multimodal single objective, making the optimization not only continuous real values but also discrete set values. By
using the so-called subset values, discrete design variables are read-
problem more likely to suffer more badly from local minima.
ily given to the optimization process. This is advantageous for a de-
When applied to many scientific and engineering problems for
sign variable, such as column diameter, whose values are often a set
decision making, it is advantageous that the decision maker can si-
of diameters predefined by the machine manufacturer.
multaneously observe the influence of the possibly many optimal
In the multiobjective optimization problem under consideration
solutions on each of the defined objectives separately. The genetic
in this paper, we aim to minimize the construction cost, which in-
algorithm is preferably used, as the intermediate solutions are a
cludes material cost, while minimizing settlement resulting from
mating pool of candidates who descend from one generation to the ground improvement design at the same time. Given a design
the next by the computer simulation of metaheuristic procedures of SCs or RIs, let us define f1() and f2() as the cost and settlement,
respectively. The key parameters for the design of SCs and RIs,
hereafter called design variables, as they are changing values dur-
ing the optimization-based design process, are subject to optimiza-
tion. The column diameter x1 = D, grid spacing x2 = a, and column
length x3 = L are taken as three design variables. Then any set of
design variables x1, x2, x3 corresponds to the resulting objectives
for settlement f1(x1, x2, x3) and construction cost f2(x1, x2, x3).
The settlement function f1 encompasses the respective calculation
procedures for SCs and RIs described previously, whereas the
cost objective function f2 is a simple mapping from the material
volume for RIs and SCs, given the unit prices for stone aggregates
and plain concrete, respectively.
The design variables can be subject to one or more constraint
rules; for example, the grid spacing a should not exceed a certain
ratio to the column diameter D. Let us name this constraint rule
g1(x1, x2).
The multiobjective optimization problem is defined as

find: (x1 , x2 , x3 )
such that: (f1 , f2 ) is minimized (1)
while being subject to: g1

Generally, there can be as many objectives, design variables,

and constraint rules as the optimization problem requires. How-
ever, to make the run time feasible in practice, the number of design
variables is usually not more than a dozen and the number of objec-
tives is even less. Although the constraint rules are not directly re-
Fig. 2. Improvement factor and unimproved or improved settlements
lated to the complexity of the multiobjective optimization problem,
for SC design solution with grid spacing a = 1.6 m, column diameter
a large number of constraint rules increases the computational ef-
D = 0.9 m, and column depth L = 16 m. Graphs are plotted as far as
fort related to evaluations of the additional intermediate solutions,
the limit depth, Lm = 17.5 m.
as design solutions that violate the constraint rules are rejected.

Table 2. Skin friction and base resistance for RIs

Parameter Soft clay Loose sand Soft clay Medium dense sand
Limit skin resistance qs (kN/m2) 50 120 80 200
Limit tip resistance qb (kN/m2) 0 0 0 2,000

© ASCE 4 Int. J. Geomech.

 (a) (b)

(c)

Fig. 3. Calculation results for RI design solution with grid spacing a = 2.0 m, RI diameter D = 0.44 m, and RI depth L = 15.6 m: (a) settlement;
(b) skin resistance; and (c) vertical force.

The optimization-based design is ideally carried out in a objectives, and constraint rules have all been defined. After the op-
computer-controlled manner. For the automated optimization pro- timization run is finished, the user can then select from a pool of
cess to be executed, the multiobjective optimization procedure is optimal solutions, which is represented as a Pareto front, a design
connected with the calculation model, for either RIs or SCs. solution with performance meeting a certain serviceability
Fig. 1 shows a typical design optimization workflow assisted requirement.
with a multiobjective optimization algorithm. Compared with a Most often the initial design solutions, which are called the ini-
conventional approach to design SCs and RIs, the conventional de- tial population in the genetic algorithm, are randomly generated,
sign steps until the calculation model being created remains un- given the defined ranges or possible set values for the design vari-
changed. Additional refinement of the design is then carried out ables. But if the designer has a set of feasible solutions in mind,
by a numerical algorithm after the design variables, optimization mostly based on settings of the design problem and engineering

© ASCE 5 Int. J. Geomech.

 (a)
Fig. 4. NSGA-II converged objective values for an optimization run of
the SC design. Marked points are the NSGA-II converged solutions at
the final generation. Gray points are intermediate results from previous
generations.

experience, these possible solutions can be fed as the initial popu-

lation for the optimization run.
At convergence, the optimization algorithm delivers a set of op-
timal design solutions with objective values competing with each
other. The quality for each of these optimal designs can be evalu-
ated by inspecting the corresponding objective values. The so-
called Pareto front is plotted on a multidimensional graph. In this
work, with construction cost and settlement as two competing de-
sign objectives, the resulting objectives can be viewed in a two-
dimensional graph. The designer can then pick a design solution
(b)
that meets the design requirements based on both optimization ob-
jectives: construction cost and settlement performance.

Optimization Examples

Design Problem
Let us consider a soil stratigraphy with the soil parameters given in
Table 1. The soil stratigraphy begins with a top soft clay, 7 m thick,
followed by loose sand, 4 m thick. The soft clay occurs again for
the next 4 m before a competent medium dense sand begins from
a depth of 15 m.
A gravity surcharge of 120 kPa acts on top of the ground improve-
ment system. For this magnitude of surface load, the influence of the
added vertical load due to surcharge becomes negligible from
17.5 m below ground surface (b.g.s). The limit depth for settlement cal- (c)
culation is therefore set at 17.5 m b.g.s. To have an impression of how
Fig. 5. Variables for the optimal design of the SC design: (a) grid spac-
effective each of the soil improvement systems are, the results of a one-
ing and column diameter; (b) grid spacing and column length; and
time calculation for a setup of SC and RI are presented in the following.
(c) column diameter and column length. Marked points are the
NSGA-II converged solutions at the final generation. Gray points are
Calculation Results for SCs intermediate results from previous generations.
Fig. 2 shows calculation results for a design of SCs with triangular
grid spacing a = 1.6 m, column diameter D = 0.9 m, and column
depth L = 16 m. The stone aggregates have an oedometer modulus
of 100 MPa and an internal friction angle of 42◦ . It can be observed installation, the surface settlement is reduced from 578 mm of the
that the improvement factor n2 increases quasilinearly with increas- natural soil to 152 mm of the treated soil.
ing depth in the upper soft clay up to a maximum at 6.0. It then
drops to 3.0 in the loose sand and remains so in the lower clay Calculation Results for RIs
layer. From 15 m b.g.s, where the medium dense sand begins, The considered RIs are full displacement plain concrete piles of
the improvement factor n2 is nearly 1.0. As a consequence of SC 0.44 m in diameter. The concrete strength of RIs is 10 MPa. The

© ASCE 6 Int. J. Geomech.

Table 3. Selected Pareto-optimal design solutions for the SC system
Selected solution Column diameter (m) Grid spacing (m) Column length (m) Settlement (mm) Cost (1,000 EUR)
Solution 1 0.8 1.62 14.95 157.10 90.31
Solution 5 0.8 2.00 14.27 220.57 55.98
Solution 17 0.8 2.35 14.32 256.84 41.01

(a)

Fig. 6. NSGA-II converged objective values for an optimization run of

the RI design. Marked points are the NSGA-II converged solutions at
the final generation. Gray points are intermediate results from previous
generations.

limit skin friction and base resistance for SCs given in Table 2 are
adopted for the calculation for RIs. Those are taken as resistance
capacities of bored piles, which are in agreement with soil param-
eters of the clayey soils and sands given in Table 1.
The LTP layer is 0.4 m well-compacted sand with a Young’s
modulus of 75 MPa and an internal friction angle of 38°. Fig. 3
shows the calculation results for a RI design with grid spacing a
= 2.0 m, RI diameter D = 0.44 m, and RI depth L = 15.6 m. It is
worth mentioning that the mobilized skin resistance at any depth
is additionally limited by the available earth pressure there. It is (b)
noted that, for the calculation in RIs, the oedometer moduli for
both RI and soil are adjusted following a stress-dependent relation.
Fig. 3 shows that the soil settles more than the RIs do above the
neutral line at 5.3 m b.g.s. (negative skin resistance), while the
RI axial forces increase at the same time. Below the neutral line,
the RI axial force is transferred gradually to the soil through posi-
tive skin friction. The surface settlement as a result of soil improve-
ment with RI is only 30 mm.

Design Optimization
For both of the optimization examples, a simple cost model is used
to calculate the monetary cost. It is assumed that the construction
cost is only dependent on the stone aggregates for SCs, and con-
crete volume for RIs. In particular, stone aggregates and concrete
volume cost 30 and 520 EUR/m3, respectively. This can be a fair (c)
assumption because personnel and operation cost can be thought
Fig. 7. Variables for the optimal RI design: (a) grid spacing and RI di-
of as proportional to the reinforcement volume to be constructed,
ameter; (b) grid spacing and RI length; and (c) RI diameter and RI
while the cost for mobilizing the construction machines can be con-
length. Marked points are the NSGA-II converged solutions at the
sidered as a fixed cost. The lateral extent for the calculation of cost
final generation. Gray points are intermediate results from previous
is a circular surface with a diameter of 34 m, representing the size
generations.
of a typical foundation for a circular liquid tank.

© ASCE 7 Int. J. Geomech.

Table 4. Selected Pareto-optimal design solutions for the RI system
Selected solution RI diameter (m) Grid spacing (m) RI length (m) Settlement (mm) Cost (1,000 EUR)
Solution 3 0.44 2.25 16.38 33.57 232.37
Solution 10 0.62 1.86 16.96 14.60 698.58
Solution 14 0.44 1.88 16.77 22.21 340.45

Optimization of SC Design magnitude less settlement but also an order of magnitude more in
Of the design parameters for a SC system, the length of column construction costs.
L, column diameter D, and grid spacing a are considered the As can be seen in Fig. 7, the resulting optimal designs span all RI
most influential parameters that decide the system’s performance diameters, but grid spacing is concentrated between 1.8 and 2.3 m.
and cost. The column diameter D [m] is a set value variable, The RI length is at least 15 m, where medium dense sand occurs.
which can have one of the values in the set {0.8, 0.9, 1.2}. The col- The embedment of rigid RIs in medium dense sand is set on
umn spacing a [m] is a real value variable, which has its value in the purpose because only this soil layer provides a tip resistance
range [1.6, 2.8]. The column length L [m] is also a real value var- (see Table 2). Table 4 gives the exemplary solutions marked in
iable, whose value is in the range [12, 17]. The limit depth is set Figs. 6 and 7. For all the selected design solutions, the values for
at 17.5 m for settlement calculation. RI length do not vary much. The system’s construction cost and
The results of optimization are given in Figs. 4 and 5 for the ob- settlement are the results of various combinations of grid spacing
jectives and design variables, respectively. It can be seen in Fig. 4 and RI diameter. If the optimal combinations are to be determined
that the 20 Pareto-optimal solutions have converged at the end of by a geotechnical engineer, he or she must take great effort to come
50 generations of the NSGA-II run. The Pareto-optimal results pro- close to a solution that lies in the Pareto front converged on by an
vide solutions for a range of settlement from 16 to 33 cm. The op- NSGA-II run.
timal solution that delivers the smallest settlement costs EUR 90
thousand and that for the largest settlement costs only EUR 24
thousand.
Each Pareto-optimal point shown in Fig. 4 has a corresponding Conclusions
set of values for the design variables shown in Fig. 5. Three repre-
It has been shown in this work that multiobjective optimization
sentative design solutions corresponding to the 1st, 5th, and 17th of
20 solutions are marked in Figs. 4 and 5. The numerical values of works well for the design of soil improvement grids, with examples
the selected optimal solutions are given in Table 3. for the design of SCs and RIs. The optimization method, based on a
Within the presented optimization settings, it can be observed in genetic algorithm, not only provides an optimal design solution that
Fig. 5 that smaller-diameter SCs are preferred over larger-diameter meets certain cost and settlement requirement but also a set of other
SCs. The resulting optimal diameter of SC is 0.8 m for all optimal possible optimal designs. This is beneficial for engineering judg-
design solutions. As column diameters are the same, differences in ments for other feasible design solutions, mainly for trading con-
construction cost and settlement performance depend on the other struction cost for system performance and vice versa.
two design variables: the grid spacing and column length. While We applied a numerical method to solve the multiobjective opti-
grid spacing is well spread over the design space, column length mization problem in which the calculations for SCs and RIs are an-
is mainly concentrated at a depth between 12 and 15 m. Because alytically modeled considering the unit cell model. The related
column length also does not vary over a large range, the system’s calculation time is not of concern. If more demanding computational
construction cost and performance greatly depend on the grid spac- design analyses, such as time-dependent consolidation and three-
ing of SCs, which is directly related to the replacement ratio. The dimensional models, are involved, the optimization approach is
selected optimization results given in Table 3 clearly reveal the likely to suffer from computational burden. To mitigate this issue,
importance of grid spacing on settlement and cost. a response surface construction method can be used to build
fast-to-evaluate surrogate models for the computationally heavy
model before the optimization step. If the full model, instead of
Optimization of RI Design
the unit cell model, is considered, it enables the use of more sophis-
With the settings very similar to the optimization of the SC design,
ticated design analyses and will also allow for more demanding op-
the same three design variables for the RIs are subject to optimiza-
timization tasks, such as the minimization of differential settlement.
tion. The RI diameter D [m] is a set value variable, which can have
Under the pressure of CO2 reduction and increasing sustainability
one of the values in the set {0.44, 0.51, 0.62}. The RI spacing a [m]
in construction, such design optimization using numerical methods is
is a real value variable, which has its value in the range [1.6, 2.8].
to be widely practiced in the design of foundations in general and
The RI length L [m] is a real value variable, whose value is in the
ground improvement systems in particular in the coming years. In
range [12, 17]. As for SCs, the limit depth for the calculation of RI
line with the digital movements in construction and fast advance-
settlement is set at 17.5 m.
ments in optimization methods in such fields as applied mathematics
Fig. 6 displays the Pareto-optimal solutions in the objective
and machine learning, the developments and use of numerical opti-
space. It can be observed that the surface settlement can be as
mization in design and construction are better supported than ever.
low as 15 mm, where the construction cost is highest. The compet-
ing nature of the optimization is shown, in that the lower the con-
struction cost, the worse the performance that the system can
deliver with respect to settlement reduction. A final solution can Data Availability Statement
then be selected based on the project requirement, such as the al-
lowable settlement. Compared with the optimal design solutions All data, models, or code that support the findings of this study are
for SCs, solutions provided by RIs generally deliver an order of available from the corresponding author upon reasonable request.

© ASCE 8 Int. J. Geomech.

Acknowledgments Kinzler, S., F. Konig, and J. Grabe. 2007. “Entwurf einer pfahlgrundung
unter anwendung der mehrkriterien-optimierung.” Bauingenieur
The author thanks Dr. Techn. Chien-Hsun Chen for his kind help in 82 (9): 367–379.
proofreading and giving valuable suggestions that improved the Klobe, B. 2020. “The design of plane earthwork structures on pile founda-
quality of this manuscript. Support in implementation and valida- tions.” Geotechnik 44 (1): 24–32. https://doi.org/10.1002/gete.v44.1.
Leung, Y. F., A. Klar, and K. Soga. 2010. “Theoretical study on pile length
tion of the ASIRI recommendations and the Priebe method from
optimization of pile groups and piled rafts.” J. Geotech. Geoenviron.
Mr. Ragadeep Bojja, Mr. Pablo Forgoso, and Mr. Fadi Haddad is Eng. 136 (2): 319–330. https://doi.org/10.1061/(ASCE)GT.1943-5606
gratefully acknowledged. .0000206.
Leung, Y. F., A. Klar, K. Soga, and N. Hoult. 2017. “Superstructure-
foundation interaction in multi-objective pile group optimization con-
References siderin settlement response.” Can. Geotech. J. 54 (10): 1408–1420.
https://doi.org/10.1139/cgj-2016-0498.
Blanc, M., G. Rault, L. Thorel, and M. Almeida. 2013. “Centrifuge inves- Meier, J. 2019. “Automatisierte Optimierung von Baugrubensicherungen.”
tigation of load transfer mechanisms in a granular mattress above a rigid In Vorträge zum 26. Darmstädter Geotechnik-Kolloquium, 19–30.
inclusions network.” Geotext. Geomembr. 36: 92–105. https://doi.org Darmstadt, Germany: Technische Universität Darmstadt.
/10.1016/j.geotexmem.2012.12.001. Nakanishi, K., and I. Takewaki. 2013. “Optimum pile arrangement in piled
Castro, J. 2017. “Modeling stone columns.” Materials 10 (7): 782. https:// raft foundation by using simplified settlement analysis and adaptive
doi.org/10.3390/ma10070782. step-length algorithm.” Geomech. Eng. 5 (6): 519–540. https://doi.org
Chan, C. M., L. Zhang, and J. T. Ng. 2009. “Optimization of pile groups /10.12989/gae.2013.5.6.519.
using hybrid genetic algorithms.” J. Geotech. Geoenviron. Eng. Pecker, A. 2004. “Design and construction of the Rion Antirion bridge.” In
135 (4): 497–505. https://doi.org/10.1061/(ASCE)1090-0241(2009) Geotechnical engineering for transportation projects, edited by
135:4(497). Edward Kavazanjian and Mishac K. Yegian, 216–240. Reston, VA:
Chankong, V., and Y. Y Haimes. 2008. Multiobjective decision making: ASCE.
Theory and methodology. Mineola, NY: Courier Dover Publications. Phutthananon, C., P. Jongpradist, D. Dias, P. Jamsawang, and D. T.
Chen, C., F. Mao, G. Zhang, J. Huang, J. G. Zornberg, X. Liang, and Bergado. 2021. “Performance-based design optimization of embank-
J. Chen. 2021. “Settlement-based cost optimization of geogrid- ments resting on soft soil improved with t-shaped and conventional
reinforced pile-supported foundation.” Geosynth. Int. 28 (5): 541–557. dcm columns.” Acta Geotech. 16 (10): 3301–3326. https://doi.org/10
https://doi.org/10.1680/jgein.21.00002. .1007/s11440-021-01258-x.
Chevalier, B., P. Villard, and G. Combe. 2011. “Investigation of load- Priebe, H. J. 1995. “The design of vibro replacement.” Ground Eng.
transfer mechanisms in geotechnical earth structures with thin fill plat- 28 (10): 31.
forms reinforced by rigid inclusions.” Int. J. Geomech. 11 (3): 239–250. Ravichandran, N., and S. Shrestha. 2020. “Performance-and cost-based ro-
https://doi.org/10.1061/(ASCE)GM.1943-5622.0000083. bust design optimization procedure for typical foundations for wind tur-
Cunha, R., H. Poulos, and J. Small. 2001. “Investigation of design alterna- bine.” Int. J. Geotech. Eng. 14 (4): 395–408. https://doi.org/10.1080
tives for a piled raft case history.” J. Geotech. Geoenviron. Eng.
/19386362.2018.1428387.
127 (8): 635–641. https://doi.org/10.1061/(ASCE)1090-0241(2001)
Reul, O., and M. F. Randolph. 2004. “Design strategies for piled rafts sub-
127:8(635).
jected to nonuniform vertical loading.” J. Geotech. Geoenviron. Eng.
Deb, K., and A. Dhar. 2011. “Optimum design of stone column-improved
130 (1): 1–13. https://doi.org/10.1061/(ASCE)1090-0241(2004)130:1(1).
soft soil using multiobjective optimization technique.” Comput.
Schaffer, J. D. 1985. Some experiments in machine learning using vector
Geotech. 38 (1): 50–57. https://doi.org/10.1016/j.compgeo.2010.10.005.
evaluated genetic algorithms. Rep. No. Nashville, TN: Vanderbilt Univ.
Deb, K., A. Pratap, S. Agarwal, and T. Meyarivan. 2002. “A fast and elitist
Seed, H. B., and J. R. Booker. 1977. “Stabilization of potentially liquefiable
multiobjective genetic algorithm: NSGA-II.” IEEE Trans. Evol.
sand deposits using gravel drains.” J. Geotech. Eng. 103 (7): 757–768.
Comput. 6 (2): 182–197. https://doi.org/10.1109/4235.996017.
Seitz, K.-F., and J. Grabe. 2016. “Three-dimensional topology optimization
Gu, M., J. Han, and M. Zhao. 2017. “Three-dimensional discrete-element
method analysis of stresses and deformations of a single for geotechnical foundations in granular soil.” Comput. Geotech. 80:
geogrid-encased stone column.” Int. J. Geomech. 17 (9): 04017070. 41–48. https://doi.org/10.1016/j.compgeo.2016.06.012.
https://doi.org/10.1061/(ASCE)GM.1943-5622.0000952. Shen, Q., F. Vahdatikhaki, H. Voordijk, J. van der Gucht, and L. van der
Indraratna, B., N. T. Ngo, C. Rujikiatkamjorn, and S. W. Sloan. 2015. Meer. 2022. “Metamodel-based generative design of wind turbine foun-
“Coupled discrete element–finite difference method for analysing the dations.” Autom. Constr. 138: 104233. https://doi.org/10.1016/j.autcon
load-deformation behaviour of a single stone column in soft soil.” .2022.104233.
Comput. Geotech. 63: 267–278. https://doi.org/10.1016/j.compgeo Simon, B. 2012. “General report S5 rigid inclusions and stone columns.”
.2014.10.002. In Proc., Int. Symp. on Ground Improvement, 127–168. London:
IREX. 2012. “Recommendations for the design, construction and control of ISSMGE.
rigid inclusion ground improvements: ASIRI National Project.” Paris: Srinivas, N., and K. Deb. 1994. “Muiltiobjective optimization using nondo-
Presses des Ponts. minated sorting in genetic algorithms.” Evol. Comput. 2 (3): 221–248.
Juang, C. H., and L. Wang. 2013. “Reliability-based robust geotechnical https://doi.org/10.1162/evco.1994.2.3.221.
design of spread foundations using multi-objective genetic algorithm.” Tan, X., L. Feng, Z. Hu, and M. Zhao. 2021. “Failure modes and ultimate
Comput. Geotech. 48: 96–106. https://doi.org/10.1016/j.compgeo.2012 bearing capacity of the isolated stone column in soft soil.” Bull. Eng.
.10.003. Geol. Environ. 80 (3): 2629–2642. https://doi.org/10.1007/s10064
Kim, K. N., S. -H. Lee, K. -S. Kim, C. -K. Chung, M. M. Kim, and H. S. -020-02066-5.
Lee. 2001. “Optimal pile arrangement for minimizing differential settle- Yoo, C., and S.-B. Kim. 2009. “Numerical modeling of geosynthetic-
ments in piled raft foundations.” Comput. Geotech. 28 (4): 235–253. encased stone column-reinforced ground.” Geosynth. Int. 16 (3):
https://doi.org/10.1016/S0266-352X(01)00002-7. 116–126. https://doi.org/10.1680/gein.2009.16.3.116.

© ASCE 9 Int. J. Geomech.