Computers and Geotechnics 169 (2024) 106239

Contents lists available at ScienceDirect

Computers and Geotechnics

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

A limit analysis-based topology optimisation method for

geostructure design
Xifan Li, Xue Zhang *, Yujia Zhang
Department of Civil and Environmental Engineering, School of Engineering, University of Liverpool, Liverpool, UK

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

Keywords: Geostructures, vital for the progress of civilisation, often face inefficiencies and suboptimal performance due to
Topology optimisation the lack of optimisation in current designs. Achieving cost-efficiency in geostructure design involves optimising
Limit analysis material usage while considering practical construction aspects. While size and shape optimisations are common
Density-based method
in geostructure design, the application of topology optimisation remains underexplored. This paper addresses
Embankment
Soil foundation
this gap by introducing a novel topology optimisation method for three-dimensional geostructure design. The
method integrates mixed limit analysis and density-based topology optimisation theories, allowing for two-
material design focused on the ultimate bearing capacity of the geostructure. The innovation resides in align­
ing the applied external load in the topology optimisation process with the ultimate load that the designed
geostructure can sustain. The robustness of the proposed method is exemplified through its application to the
design of an embankment and soil foundation, showcasing its potential to enhance the efficiency and perfor­
mance of geostructures. This research contributes to the advancement of geostructure design practices, ulti­
mately promoting sustainable and resilient infrastructure development.

1. Introduction constraints. Although topology optimisation is renowned for its effi­

ciency in achieving optimal designs, its application in geostructure
Geostructures play a key role in the advancement of human civili­ design remains underutilised with limited contributions specifically
sation, providing essential support for buildings and transportation dedicated to its use in underground excavation (Ren et al., 2005,2014)
infrastructure. Despite their pivotal role, current designs of geo­ and foundation design (Kammoun et al., 2019, Pucker and Grabe, 2011,
structures often lack optimisation, resulting in inefficiencies and sub­ Seitz and Grabe, 2016). This is notably distinct from other engineering
optimal performance. Designing geostructures to be cost-efficient disciplines such as automotive engineering, aircraft engineering, and
involves optimising material usage while considering practical con­ even structural engineering, where topology optimisation is actively
struction aspects. However, achieving this balance is challenging due to employed to tailor structural designs optimally to problem-specific
the multitude of factors at play, encompassing geological, geotechnical, demands.
structural, and environmental considerations. Several topology optimisation approaches have been devoted over
In geostructure design, conventional practices often rely on size and the past few decades (Sigmund and Maute, 2013). Among them, the
shape optimisations. Size optimisation aims to identify the optimal di­ density-based topology optimisation method, particularly the so-called
mensions or scaling of structural elements while preserving the existing Solid Isotropic Material with Penalization (SIMP) method, stands out
shape, while shape optimisation focuses on finding the optimal config­ as widely adopted. The traditional SIMP method involves minimising
uration of structural elements while maintaining a consistent size. the elastic strain energy of a structure subjected to specified external
Common structural elements in geostructure design include piles, loads, a volume constraint, and assumptions of linear elastic material
shallow foundations, walls, anchors, among others. An alternative to behaviour. This approach found practical application in the field of
these two methods is the topology optimisation, which optimises the geotechnical engineering, specifically in the design of foundation
spatial distribution of material within a specified domain. This is ach­ structures, as demonstrated by Pucker and Grabe (2011). Despite the
ieved by minimising a predefined cost function while meeting specified widespread use of the traditional SIMP method in structural

* Corresponding author.
E-mail address: xue.zhang2@liverpool.ac.uk (X. Zhang).

https://doi.org/10.1016/j.compgeo.2024.106239
Received 27 November 2023; Received in revised form 11 March 2024; Accepted 12 March 2024
Available online 16 March 2024
0266-352X/© 2024 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
X. Li et al. Computers and Geotechnics 169 (2024) 106239

optimisation, designs obtained through this approach may not ensure and stress as variables, as introduced by Krabbenhøft et al. (2007), and
the feasibility of stress states in relation to material strength. In simpler operates within the framework of the generalized Hellinger Reissner
terms, the optimised design generated by the traditional SIMP method variational principle proposed by Zhang et al. (2019). Specifically, a
does not guarantee that the stress states within the structure remain mixed limit analysis formulation is expressed as a min–max optimisation
below the yield limit of the material when subjected to the considered problem:
external loads. This raises concerns about the structural integrity and ∫ ∫ ∫
performance of the designed system in real-world conditions. The stress-
T
min max α + σ T ∇T (u)dΩ − α t udΓ − bT udΩ
u ( σ ,α ) (1)
constrained topology optimisation method was proposed to address this Ω Γt Ω
subject to f (σ ) ≤ 0
issue (Duysinx and Bendsøe, 1998), wherein stress constraints are
additionally incorporated. Remarkably, linear elastic structural analysis ( )T
continues to be utilised in the traditional stress-constrained topology where σ = σ xx σyy σzz σyz σ zx σxy represents the Cauchy
optimisation, notwithstanding the added stress constraints (Holmberg stress, u denotes the displacement, b is the body force, t is the prescribed
et al., 2013). This persistence results in a conservative design, as traction, f(σ ) is the yield function, α is the collapse load factor (e.g., αt
demonstrated in the work of Zhang et al., (2023a) and Li et al. (2023). indicating the ultimate force the structure can sustain), and ∇ is the
An alternative approach, which considers stress, is the topology opti­ differential operator matrix.
misation method for stiffness and strength. Functioning as a hybrid In 3D cases, when employing a four-node tetrahedron, the approxi­
method, it seeks to minimise both the compliance of a structure and the mation of displacement (u) within an element is expressed as:
homogenisation of stresses and was recently applied in the design of u(x) ≈ N̂
u (2)
strip foundations (Hadjiloo et al., 2023).
In a recent development, Kammoun and Smaoui (2015) proposed a where ̂u comprises the displacement values at mesh nodes, and N rep­
density-based topology optimisation formulation that integrates limit resents the shape function. Accordingly, the strain is approximated by:
plasticity. This method, grounded in lower bound finite element limit
analysis, ensures the preservation of a statically admissible stress field ε ≈ ∇T (N̂
u ) = B̂
u (3)
throughout the optimisation process, thereby facilitating the determi­
nation of a structure’s ultimate plastic limit. Subsequently, Herfelt et al. with B = ∇T N being the strain–displacement matrix. Sequentially, the
(2019) introduced an alternative plasticity-based formulation that en­ min–max problem (1) is discretised as
ables both upper and lower bound finite element limit analysis using ∫ ∫
T
∫
different elements. However, their formulation produces designs solely min max α + σ T BdΩ̂
u− α t NdΓ̂
u− bT NdΩ̂
u
û (σ ,α) Ω Γt Ω (4)
in greyscale. To overcome this limitation, Zhang et al., (2023a) intro­ subject to f (σ ) ≤ 0
duced a penalty into the framework proposed by Herfelt et al. (2019),
allowing for the creation of black-and-white designs. It is also demon­ where σ represents the uniform stress in the integral domain. The min­
strated that the formulation proposed in the work of Zhang et al., imisation part of (4) can be resolved analytically, leading to a subse­
(2023a) does not require a stress-relaxation technique. Li et al. (2023) quent maximisation problem which is
further extended the work specifically for three-dimensional scenarios
involving mesh refinement techniques. Alternatively, Mourad et al. max
(σ ,α)
α
(2021) introduced a formulation for topology optimisation incorpo­ ∫ ∫ ∫
σ T BdΩ = α
T
bT NdΩ (5)
rating plastic limit analysis. Diverging from the previously mentioned subject to : t NdΓ +
Ω Γt Ω
approaches (Herfelt et al., 2019; Kammoun and Smaoui, 2015; Li et al., f (σ ) ≤ 0
2023; Zhang et al., 2023a), Mourad et al. (2021) aims to maximise the
load-bearing capacity of the structure. This is accomplished while The mixed limit analysis formulation mentioned above has been inte­
considering material strength properties and adhering to a material grated into the finite element limit analysis commercial software
volume constraint. OptumG2/G3. In the implementation, the finite elements utilised as­
This paper aims to develop a tailored topology optimisation method sume a quadratic displacement field and a linear stress field. It is note­
for designing three-dimensional geostructures, with a emphasise on worthy that the occurrence of volumetric locking is evident when
achieving the design’s limit state. To fulfill this goal, the devised method employing elements with linear shape functions, for instance three-node
integrates both mixed limit analysis theory and density-based topology triangular elements in 2D cases and four-node tetrahedral elements in
optimisation theory. It accommodates two materials, such as soils and 3D cases. To tackle this issue, this study incorporates nodal integration
foundations/reinforcements, and incorporates gravity effects into the based on smoothing domains, also referred to as cells, following the
topology optimisation process. The ultimate formulation seeks to work of Zhang et al. (2023b). The construction of cells involves dividing
minimise the volume of material used for foundations or reinforcement, each tetrahedron into four hexahedrons (see Fig. 1(a)), ensuring equal
considering yielding states of both soil and foundations/reinforcements. volumes, by connecting the centroid of the tetrahedron to the centroids
To validate the proposed method, the geostructure resulting from the of the four surface triangles and then the centroid of each surface tri­
developed method is extracted for bearing capacity analysis. Addition­ angle to the three mid-edge nodes of the triangle. The cell associated
ally, a comparative analysis between the performance of designs derived with a node comprises all hexahedrons adjacent to that specific node.
from our approach and those using traditional geotechnical practices Fig. 1(b) illustrates a cubic domain discretised using cells. For further
was conducted to illustrate the robustness and efficiency of our method. insights into the construction of smoothing domains/cells, interested
The rest of the paper is structured as follows. Section 2 introduces the readers are directed to the work of Li et al. (2023) and Zhang et al.
theory of mixed limit analysis. Section 3 presents the plasticity-based (2023b). It has been demonstrated that the volumetric locking issue
topology optimisation formulation for two-material design. Numerical associated with linear elements can be effectively overcome by
examples are illustrated in Section 4 with conclusions drawn in Section employing this approach (Meng et al., 2020; Zhang et al., 2023b).
5. Following nodal integration, maximisation problem (5) is written as

2. Mixed limit analysis formulation

In this section, we provide a summary of the formulation for mixed

limit analysis. The mixed formulation incorporates both displacement

2
X. Li et al. Computers and Geotechnics 169 (2024) 106239

Fig. 1. Illustrations of 3D cell construction: (a) A tetrahedron divided into four hexahedrons, and (b) the cell generated based on tetrahedrons for a cube (Zhang
et al., 2023b).

max α 3. Topology optimisation formulation for two-material

structures
(σ ,α)
⎧
⎨ BT ̂ σ − Fb = Ft (6)
subject to:
⎩ f i (̂ 3.1. Single-material structure design
σ ) ≤ 0 i = 1, 2, ⋯, NN
Li et al. (2023) indicated that the topology optimisation with plas­
in which the notation ( • )i denotes the value of ( • ) at the ith node ticity can be formulated as a volume minimisation problem through the
unless otherwise specified, NN represents the total number of nodes introduction of a novel design variable, ‘density’ (denoted as ρ ∈ [0, 1]),
(equal to the total number of cells, NC). The vector ̂
σ comprises stress in problem (6). The modified optimisation problem takes the form:
components at all mesh nodes and can be viewed as a weighted average
stress of the tetrahedrons adjacent to each node. min LT ̂
ρ
(σ̂ ,ρ̂)
The global matrix B is computed by
T ⎧
⎨ BT ̂ σ − Fb ̂ρ = Ft (10)
( ))
∫ ∑ ne ( subject to:
NN
1 ∑ 1 T j ⎩ f i (̂
(7)
T
B = T
B dΩ = B V σ , ̂ρ ) ≤ 0 i = 1, 2, ⋯, NN
Ω i=1
V i j=1 4 ij

∑ where ̂
ρ = [ρ1 , ρ2 , ⋯, ρNN ]T represents a vector containing the ‘density’
where ( • ) denotes the standard finite element assembly operator, V i
(ρ) at all nodes, and L = [V1 , V2 , ⋯, VNC ]T is a vector comprising the
is the volume of the ith cell, V j is the volume of the jth tetrahedron, ne is volumes of all cells. In the work of Li et al. (2023), only one material is
the total number of tetrahedron elements adjacent to the ith node, and considered in the optimal design. Thereby, ρi = 0 in problem (10) im­
Bij is the strain–displacement operator at the ith node contributed by the plies that the ith cell is a void. In this study, the von Mises model is
( )
jth adjacent tetrahedron. Hence, V1i ne
∑ 1 T j employed as the plasticity model for clay. The yield condition of von
j=1 4Bij V represents the
Mises provides a smooth approximation of the Tresca yield condition,
weighted average of the strain–displacement operator at the ith node. which has been widely embraced for the analysis of clay in undrained
The surface force and body force are Ft and Fb , respectively, whose conditions (Dunne and Martin, 2017, Walker and Yu, 2010, Zhang et al.,
estimates based on nodal integration are 2023b). When using the von Mises yield criterion, the inequality
∫ ∑
NN ( ) constraint in problem (10) is expressed as:
(8)
T i
Ft = α t NdΓ = α βiΓ ̃t Γi √̅̅̅̅̅̅̅
Γt i=1 f i = 3J2 − ρi fy ≤ 0 (11)

and where J2 = 16(σx − σ y )2 +16(σy − σ z )2 +16(σz − σ x )2 +σ 2xy +σ2yz +σ 2zx is the

∫ ∑NN ( ) second invariant of the deviatoric stress, and fy is the yield stress of the
Fb = bT NdΩ = i=1
bi V i (9) material. In the case of single material topology design, the modified
Ω
yield criterion (11) is supplemented with an additional inequality con­
in which Γi is the nodal contour area, and βiΓ is a factor that equals one if dition:
the node belongs to a Neumann boundary and zero otherwise. ⃒ ⃒
fspi = ⃒σx + σ y + σ z ⃒ − kfy ρi ≤ 0 (12)
It is worth highlighting that, even in the absence of an upper/lower
bound feature in the abovementioned mixed limit analysis, the solution
to prevent a spherical stress state (e.g., σx = σy = σz ) when ρi = 0.
derived from such analysis often closely approximates to the exact
bearing capacity of a structure, as indicated by Nguyen (2023).

3
X. Li et al. Computers and Geotechnics 169 (2024) 106239

min LT ̂
ρ
(σ̂ ,ρ̂)
⎧
⎪
⎪
⎪
⎪ BT ̂
σ− F ̂b̂
ρ = Ft + F ̃b
⎪
⎪
⎨ √ ̅̅̅̅̅̅
̅
subject to: f i (̂
σ , ̂ρ ) = 3J2 − ρi fy1 − (1 − ρi )fy2 ≤ 0
⎪
⎪
⎪
⎪
⎪
⎪ 0 ≤ ρi ≤ 1 i = 1, 2, ⋯, NN
⎩

⎧
⎪
⎪
⎪
⎪
⎪ ( ))
⎪
⎪
⎪ ∑NN 1 ∑ne (1
⎪
⎪
⎪
⎪
T
B = B V T j
(15)
⎪ i=1 V i j=1 4 j
⎪
⎪
⎪
⎪
⎪
⎪ ∑NN ( i
⎪
⎪ )
⎪
⎪ ̂b =
F b1 V i − bi2 V i
⎪
⎨ i=1

where
⎪
⎪
⎪
⎪
⎪
⎪
⎪ ∑NN ( i )
⎪
⎪ Ft = α i=1 βiΓ ̃tn Γi
⎪
⎪
⎪
⎪
⎪
⎪
⎪
⎪ ∑NN ( )
⎪
⎪ ̃b =
F bi2 V i
⎪
⎪
⎪
⎪ i=1
⎪
⎩

The optimisation formulation presented in (15) results in a grayscale

solution. To achieve a black-and-white design, it is necessary to penalise

Fig. 2. A plate under shear stress.

3.2. Two-material structure design

In this section, the scope of problem (10) will be generalised to

incorporate two-material topology optimisation. In other words, the
space can be either occupied by clays or foundations/reinforcements. To
this end, the Von Mises yield criterion at the ith node/cell thereby should
be modified as:
√̅̅̅̅̅̅̅
f i = 3J2 − ρi fy1 − (1 − ρi )fy2 ≤ 0 (13)

where fy1 and fy2 represent the yield stress of foundations/re­

inforcements and clays, respectively. Specifically, the ith cell is occupied
by foundations/reinforcements with yield stress fy1 if ρi = 1, and by
clays with yield stress fy2 if ρi = 0. When fy2 = 0, inequality (13) sim­
plifies to the optimal design scenario for a single material structure, for
instance, the condition outlined in (11). Given the binary occupation of
space by either foundations/reinforcements or clays in a two-material
structure design, the inclusion of the additional inequality constraint
(12) is unnecessary in this study.
Additionally, the term related to the body force should be adjusted.
For two-material structure design, the gravitational component should
be expressed as
∫ ∑NN
[ρi b1 + (1 − ρi )b2 ]NdΩ = [ρ b1 + (1 − ρi )b2 ]V i
i=1 i
Ω
∑NN ( ) ∑NN ( i )
= i=1
bi1 V i − bi2 V i ρi + i=1
b2 V i (14)

in which, b1 and b2 denote the body forces of clays and foundations/

reinforcements considered in the design.
Consequently, the topology optimisation formulation for two-
material structure design is presented as
Fig. 3. The evolution of material distributions with iteration number (a) iter =
0; (b) iter = 2; (c) iter = 4; and (d) iter = 6, where ‘density’ ρ = 0 represents
MA, and ρ = 1 represents MB, reflecting the optimisation process.

4
X. Li et al. Computers and Geotechnics 169 (2024) 106239

the objective function. In line with the approach detailed in previous ( )2

works of Kammoun et al. (2019), Zhang et al., (2023a), and Li et al. 1 ‖xj − xi ‖
( ) − σd
(19)
2

(2023), the vector L is substituted with L̃ = [c1 V1 , c2 V2 , ⋯, cNC VNC ]T , w xj = e

where the penalty factor for the ith node, ci , is determined by
in which xi denotes the coordinates of the ith node undergoing density
*
ci = ep(1− ρi ) (16) filtering, and xj represents the coordinates of the jth nodes within the
filtering region. The effectiveness of the aforementioned weighting
in which ρ*i denotes the ‘density’ at the ith node obtained in the previous function in plasticity-based topology optimisation has been demon­
iteration, and p is a constant set to 5 as suggested by Li et al. (2023), if strated in prior studies for single-material design in both 2D (Zhang
not otherwise specified. The effectiveness of this exponential penalty et al., 2023a) and 3D (Li et al., 2023) scenarios. In this study, the same
function has been investigated in previous studies conducted by Smaoui weighting function is utilised for two-material topology optimisation.
and Kammoun (2022) and Zhang et al., (2023a). The parameter σ d in Eq. (19) is set to the value of R/2, where R is 1.5
Consequentially, the plasticity-based topology optimisation for two- times the mesh size, aligning with the recommendation provided by
material structures can be expressed as: Zhang et al. (2023a).
Given that the objective function incorporates an exponential pen­
min ̃T ̂
L ρ
(σ̂ ,ρ̂) alty function of the ‘density’, achieving a converged solution necessi­
⎧ tates an iterative approach as outlined by the subsequent steps:
⎪
⎪
⎪
⎪ BT ̂
σ− F ̂b̂ ̃b
ρ = Ft + F
⎪
⎪
⎨ √ ̅̅̅̅̅̅
̅ (i) Initialisation: Start by assuming an initial ‘density’, ρ = 1, for all
subject to: f i (̂
σ , ̂ρ ) = 3J2 − ρi fy1 − (1 − ρi )fy2 ≤ 0 nodes in the domain.
⎪
⎪
⎪
⎪
⎪ (ii) Penalty Factor Calculation: Compute the exponential penalty
⎪ 0 ≤ ρi ≤ 1 i = 1, 2, ⋯, NN
⎩ factor, ci , for each node using Eq. (16).
(iii) Optimisation Problem Solving: Utilise the MOSEK solver to
⎧
⎪ solve the optimisation problem (17) which couples the limit
⎪
⎪
⎪
⎪
⎪ analysis and density-based topology optimisation, thereby
⎪
⎪
⎪
⎪
̃T = [c1 V1 , c2 V2 , ⋯, cNC VNC ]
L obtaining the density field and stress field.
⎪
⎪
⎪ (iv) Density Filtering: Apply density filtering (e.g., using Eq. (18))
⎪
⎪
⎪
(
∑NN 1 ∑ne (1 )) (17)
⎪
⎪
⎪ T T j and consider the filtered value as the updated density for each
⎪ B = B V
⎪
⎪
⎪
⎪
i=1 V i j=1 4 j node.
⎪
⎪
⎪
⎪ (v) Convergence Check: Evaluate the convergence criterion.
⎪ ∑NN ( i )
⎨
F̂b = b1 V i − bi2 V i Terminate the iteration if the criterion is met; for example, if the
where i=1
⎪
⎪
⎪ change in the objective function, Obj = L ̃T ̂
ρ , between two
⎪
⎪
⎪ ‖Objn+1 − Objn ‖
⎪
⎪ consecutive iterations satisfies ≤ tolerance value. If not,
⎪
⎪
⎪ ∑NN ( i ) Objn+1
⎪
⎪
⎪
⎪
Ft = α i=1 βiΓ ̃tn Γi proceed to step (ii) and repeat the entire process until the
⎪
⎪
⎪
⎪ convergence criterion is satisfied.
⎪
⎪ ∑NN ( i )
⎪
⎪ ̃b =
⎪
⎪ F b2 V i
⎪
⎪
⎪
i=1
Notably, both the objective function and the constrains in (15) are
⎪
⎩ linear. Consequently, the optimisation problem can be categorised as
mathematically convex. Therefore, it can generally be anticipated that
the optimisation result is a global solution. Nonetheless, with the
The optimisation problem (17) can be redefined as a conventional
introduction of the penalty factor ci , which is contingent upon the
standard second-order cone programming (SOCP) problem following
density from the preceding iteration step, the resultant binary design
the work of Meng et al. (2020) and Zhang et al. (2023a). The resolution
from the above iterative procedure may not necessarily represent a
of this SOCP problem is accomplished by MOSEK (ApS, 2019), a
global minimum.
comprehensive optimisation software designed for addressing linear,
quadratic, conic, semidefinite, and mixed integer problems, among
4. Numerical examples
others.
A common numerical instability encountered in topology optimisa­
In this section, three examples are provided to validate the proposed
tion is the checkerboard phenomenon. A standard method to enhance
method and demonstrate its robustness in geostructure design. All
numerical stability involves the use of the checkerboard-filtering
simulations were conducted on a DELL PC with a 2.20 GHz CPU and
method. In this study, we employ the density filtering method sug­
32.0 GB memory running Microsoft Windows Server (Version 11.0). The
gested by Sigmund (2007) that the filtered density at the ith point is
final SOCP problem is resolved using MOSEK (ApS, 2019), an advanced
computed as:
modern optimisation tool for addressing large-scale optimisation prob­
∑ ( )
j∈Nn w xj Vj ρj
lems, in MATLAB environment (R2022b).
ρi = ∑
̃ ( ) (18)
j∈Nn w xj Vj
4.1. Design of a short plate
Here, Vj represents the volume of the jth cell, ρj is the ‘density’ at the jth
node, and Nn denotes the total number of nodes within the filtering To validate the proposed method, we first examine the design of a
region of the ith node. The filtering region is defined as a circle with a short plate, as depicted in Fig. 2. The plate dimensions are 4 m × 8 m ×
radius of R. The weighting function, denoted as w(xi ), takes the form of a 0.01 m. The left surface of the plate is fully fixed, while the front and
Gaussian (bell shape) distribution function, as in the work of Sigmund back surfaces are fixed in the z direction. The plate is designed to
(2007): experience a vertical shear load of 227 kPa distributed over a 1.5 m
central section on its right surface.
Two materials can be used in the design of this plate: Material A (MA)
with a yield stress of 30 kPa, and Material B (MB) with a yield stress of

5
X. Li et al. Computers and Geotechnics 169 (2024) 106239

300 kPa. Given that MB is much more expensive than MA, we are ex­
pected to minimise the use of MB while ensuring that the designed plate
can sustain the given shear load.
Fig. 3 illustrates the evolution of material distribution through
various iterations. It is evident that a highly diffuse structure is achieved
when no iteration is conducted (Fig. 3(a)). Because of the application of
penalisation, an almost binary material structure emerges through
subsequent iterations (Fig. 3 (b)–(d)). As shown, with increasing itera­
tions, the plate’s domains are predominantly occupied by either MA
(ρ = 0) or MB (ρ = 1), with very few domains exhibiting 0 < ρ < 1. The
material layouts after 4 iterations (Fig. 3(c)) and 6 iterations (Fig. 3(d))
closely resemble each other, indicating a converged solution. This
convergence aligns with the trend observed in the volume ratio of MB (e.
g., the ratio of the volume of MB to the total volume of the plate) against
iteration number, as illustrated in Fig. 4. After 4 iterations, the volume
Fig. 5. The illustration of (a) the converged material layout with meshes and
ratio of MB stabilises, showing minimal change and converging to a
(b) the extracted binary-material structure.
value of 16.5 %.
Notably, this study incorporates the mesh refinement technique
introduced by Li et al. (2023) for enhancing computational efficiency.
The fundamental concept of this technique involves refining meshes in
the domain with a density greater than a specified value, such as 0.3 in
this study. The simulation begins with 6,642 nodes and 19,200 ele­
ments. As the simulation progresses, both the number of nodes and el­
ements gradually increase, reaching 10,671 nodes and 37,808 elements,
respectively. Following this point, these quantities remain constant. The
total computation time for this particular case, involving a total of 10
iterations, amounts to 243 s. This refinement results in a very thin
transition area from a density of 0 to a density of 1, as illustrated in Fig. 5
(a). For validation purposes, a binary material structure is extracted
from Fig. 5(a), as shown in Fig. 5(b). Utilising the commercial software
OptumG2, the maximum shear load that this structure can sustain is
determined to be 230 kPa, which closely matches the external shear load
value of 227 kPa applied in the initial topology optimisation simulation. Fig. 6. Schematic diagram of an embankment.
This alignment demonstrates the accuracy and validity of the proposed
method. biopolymer-treated clays, ensuring it meets the required Factor of
Safety.
4.2. Embankment design The schematic representation of the embankment is depicted in
Fig. 6, with only one-half simulated owing to its symmetrical nature. The
Earthen embankments play a crucial role in civil engineering in­ bottom of the embankment is fully fixed, while symmetric boundary
frastructures, serving to control flooding and provide support for various conditions are imposed on the left surface. The front and back surfaces
overlying transport structures. A recent focus has been on exploring are fixed in the z direction, mimicking a plane strain condition. The
biopolymer-treated soil for constructing embankments, aiming to embankment will be constructed using normal clays with a bulk unit
address challenges such as mitigating internal erosion and enhancing weight of 19.6 kN/m3 and a cohesion of 30 kPa, and biopolymer-treated
stability (Chang et al., 2020). In this section, we will apply the developed clays with a bulk unit weight of 27.5 kN/m3 and a cohesion of 230 kPa –
topology optimisation approach to design an embankment using values falling within the reported range (Chang et al., 2020). When
utilising only normal clays in the construction, the Factor of Safety (FoS)
for the constructed embankment is 2.87. In this study, we will explore
how to further incorporate the minimal biopolymer-treated clays into
the construction, ensuring a FoS of 4.0. The domain allowed to use
biopolymer-treated clays is in a shape of parallelogram (i.e., the light
blue region indicated in Fig. 6), referred to as design domain. In this
study, the yield stress fy of clays is twice the value of the cohesion,
implying that the von Mises yield surface middle circumscribes the
Tresca yield surface. Additionally, the yield stress to be used in the to­
pology optimisation process should be fy /FoS to ensure the required FoS.
In the simulation, a fine mesh with a mesh size of he = 0.1 m is employed
to discretise the region where biopolymer-treated clay can be applied. In
contrast, a relatively coarse mesh with he = 0.2 m is utilised in the
remaining area. This configuration yields a total of 10,908 nodes and
55,152 elements. Without further refinement of the meshes in the
simulation, the computational cost amounts to 335 s, allowing for the
attainment of a converged solution over 9 iterations.
Fig. 7 shows the evolution of material distribution across different
iteration steps. Like the first example, an initially diffuse distribution is
achieved without iteration. However, after undergoing 9 iterations, the
Fig. 4. The curve of volume ratio of Material B (MB) against iteration number.

6
X. Li et al. Computers and Geotechnics 169 (2024) 106239

Fig. 7. The evolution of material distributions for the embankment with iteration number (a) iter = 0; (b) iter = 2; (c) iter = 4; and (d) iter = 9, where ‘density’ ρ =
0 represents normal clays, and ρ = 1 represents biopolymer-treated clays.

material distribution transforms into a nearly binary state, yielding a topology optimisation simulation. Whilst the geometry remains constant
biopolymer-treated clay volume ratio of 6.1 %. Most biopolymer-treated for the stability analysis, it is imperative to extract interface information
clays (i.e., red part in Fig. 7(d)) are concentrated along the left and right to differentiate between the domains occupied by normal and
surfaces of the design domain (i.e., the parallelogram). A notable biopolymer-treated clays. The identification of interfaces entails
quantity of these clays is also found in the bottom-left corner of the discerning boundaries between domains with densities exceeding and
design domain. Stability analysis of the embankment design resulting falling below 0.5. This geometric and interface data is then utilised to
from the topology optimisation was conducted using the commercial construct the analysis model in OptumG2, with corresponding material
software OptumG2. In the analysis, both the normal clay and parameters, such as unit weight and cohesion, allocated to the respective
biopolymer-treated clay were treated as rigid-perfectly plastic materials. domains to facilitate the analysis. Nodes and elements are automatically
Specifically, the Drucker-Prager (DP) model in OptumG2 was employed. generated in OptumG2 utilising mesh adaptivity technique, thereby
The two strength parameters associated with the DP model are M = 0 guaranteeing the achievement of converged solutions. Fig. 8 illustrates
√̅̅̅
and k = 2c/ 3 where c is the cohesion for normal clay (30 kPa) and the stability analysis results from OptumG2. The FoS of the embankment
biopolymer-treated clay (230 kPa). Consequently, the values of with biopolymer-treated clays is 3.946, in agreement with the expected
parameter k for normal clay and biopolymer-treated clay are 35 kPa and FoS. Compared to the embankment constructed using just normal clay
266 kPa, respectively. With such settings, the DP model in OptumG2 (Fig. 8(a)), fewer soils are disturbed when the embankment with
degrades to a von Mises model with the yield surface middle circum­ biopolymer-treated clays fails (Fig. 8(b)). This is attributed to the
scribing the Tresca yield surface, ensuring that the material model used presence of the biopolymer-treated clays on the far left of the design
in the stability analysis remains consistent with that utilised in the domain, causing the shear band to propagate on its right.
It is notable that the shape of the region designated for biopolymer-
treated soil exhibits irregularity, which poses significant challenges for
practical engineering implementation. Hence, it is advisable to slightly
simplify the derived optimal topology to yield a more practical and
easily implementable soil distribution for engineering purposes. This
adaptation ensures that the insights gained from the optimisation pro­
cess are translated into feasible engineering solutions. Table 1 illustrates
the designed simplified from our topology optimisation solution and
three other designs influenced by traditional reinforcement approaches.
These conventional strategies include the integration of reinforced ma­
terials at the base, the implementation of ecological blankets, and their
synergistic combinations. All designs maintain a consistent area occu­
pied by biopolymer-treated soils, approximately 2.29 m2, reflecting a
6.1 % volume ratio of biopolymer-treated clays achieved from our to­
pology optimisation. Table 1 reveals that the design strategy inspired by
combining ecological blanket implementation and toe reinforcement
(type 3) yields the highest FoS (3.756) among the three designs
emulating conventional reinforcement strategies. Nevertheless, this FoS
is still lower than that of the embankment design simplified from our
topology optimisation solution (i.e., type 4), standing at 3.896. This
underscores the effectiveness of the proposed method for embankment
design.

4.3. Soil foundation

Fig. 8. Factor of Safety (FoS) and failure mode of the embankment (a) without
biopolymer-treated clay and (b) with biopolymer-treated clay. In the last example, we explore the design of a soil foundation as

7
X. Li et al. Computers and Geotechnics 169 (2024) 106239

Table 1
Four layout designs with the same volume ratio of the biopolymer-treated clays.
Reinforcement type Schematic diagram of the model Geometry of the reinforced soil (total area 2.29
m2)

1. Emulate ecological blanket

2. Emulate toe reinforcement

3. Emulate ecological blanket & toe

reinforcement

4. Inspired from topology optimisation solution

Fig. 9. An illustration of soil foundation design in 3D case: (a) Overall domain; and (b) top view of quarter section.

shown in Fig. 9. The scenario involves a cuboid domain measuring 12 m optimisation process.
× 6 m × 12 m. All surfaces are fixed, excepted for the top surface where The obtained material layout obtained from the developed topology
we intend to apply a pressure of 250 kPa to a circular region with a optimisation approach is shown in Fig. 10 (a). As illustrated, mesh
diameter of 2 m. When the entire domain is occupied by normal clays of refinement techniques are employed in the simulation to enhance
a bulk unit weight of 19.6 kN/m3 and a cohesion of 30 kPa, the ultimate computational efficiency, leading to a total computational cost of 927 s
pressure on the circular region it can sustain is 150 kPa. This study aims for this case. The region that should be reinforced using biopolymer-
to increase the bearing capacity up to 250 kPa using minimal treated clays is depicted in Fig. 10 (b) and (c), with a volume of 4.54
biopolymer-treated clays. The bulk unit weight and cohesion of the m3 implying a volume ratio of 2.1 %.
biopolymer-treated clays are 27.5 kN/m3 and 230 kPa, respectively. Due Although such a material layout can sustain a pressure of 250 kPa on
to symmetry, a quarter of the domain is considered in the topology the circular region, the irregularity of the reinforced clays escalates

8
X. Li et al. Computers and Geotechnics 169 (2024) 106239

Fig. 10. An illustration of the designed soil foundation from topology optimisation: (a) material layout; (b) region with biopolymer-treated soil (before rendered);
and (c) region with biopolymer-treated soil (after rendered).

Table 2
Design of soil foundations and their bearing capacity.

9
X. Li et al. Computers and Geotechnics 169 (2024) 106239

complexity in practice. We thus simplify the design according to the reproduce the results have been defined in the paper.
topology from the simulation to alleviate the construction challenges.
Three designs are illustrated in Table 2: a cylinder as the traditional Acknowledgements
circular foot, a hollow cylinder, and a design close to the shape from our
topology optimisation solution. The volume of all three cases are the This work was supported by the New Investigator Award grant of UK
same, namely 4.54 m3. Simulation analysis of the bearing capacity for Engineering and Physical Sciences Research Council (EP/V012169/1),
the three distinct designs was conducted utilising Optum G3. Although the Royal Society International Exchanges grant (IEC/NSFC/191261)
the failure mechanism of the structure is similar for these three designs, and the China Scholarship Council (CSC, No.202108420035).
the corresponding bearing capacities differ. The findings revealed that
the first scheme manifested the lowest bearing capacity of 219 kPa, References
merely achieving 87.6 % of the intended load magnitude. Following
this, the hollow cylindrical foundation scheme, being the second ApS, M., 2019. Mosek optimization toolbox for matlab. User’s Guide and Reference
Manual, Version 4.
approach, attained 92.5 % of the intended load magnitude. The third Chang, I., Lee, M., Tran, A.T.P., Lee, S., Kwon, Y.-M., Im, J., Cho, G.-C., 2020. Review on
scheme, which closely resembles our topology-optimised design solu­ biopolymer-based soil treatment (BPST) technology in geotechnical engineering
tion, comes closest to meeting the intended load magnitude, reaching practices. Transp. Geotech. 24, 100385.
Dunne, H., Martin, C., 2017. Capacity of rectangular mudmat foundations on clay under
94.8 %. This suggests that, through simple modification of the topology combined loading. Géotechnique 67 (2), 168–180.
optimisation design, we can establish a foundation that not only comes Duysinx, P., Bendsøe, M.P., 1998. Topology optimization of continuum structures with
close to or even meets the required bearing capacity but is also suitable local stress constraints. Int. J. Numer. Meth. Eng. 43 (8), 1453–1478.
Hadjiloo, E., Knutz, S., Grabe, J., 2023. Towards the saving potentials using a hybrid
for practical construction. topology optimization: Application of a coupled deterministic and empirical
approach solving a geotechnical optimization problem. Int. J. Numer. Meth. Eng.
5. Conclusions e7410.
Herfelt, M.A., Poulsen, P.N., Hoang, L.C., 2019. Strength-based topology optimisation of
plastic isotropic von Mises materials. Struct. Multidiscip. Optim. 59 (3), 893–906.
This paper presents a novel three-dimensional limit analysis-based Holmberg, E., Torstenfelt, B., Klarbring, A., 2013. Stress constrained topology
topology optimisation method tailored specifically for geostructure optimization. Struct. Multidiscip. Optim. 48, 33–47.
design. By seamlessly integrating limit analysis theory with density- Kammoun, Z., Smaoui, H., 2015. A direct method formulation for topology plastic design
of continua. In: Fuschi, P., Pisano, A.A., Weichert, D. (Eds.), Direct Methods for Limit
based topology optimisation principles, this method enables the crea­ and Shakedown Analysis of Structures: Advanced Computational Algorithms and
tion of geostructures incorporating two distinct materials, such as soils Material Modelling. Springer International Publishing, Cham, pp. 47–63.
and reinforcements/foundations. Moreover, it ensures alignment of the Kammoun, Z., Fourati, M., Smaoui, H., 2019. Direct limit analysis based topology
optimization of foundations. Soils Found. 59 (4), 1063–1072.
designed geostructure’s bearing capacity with the external load applied Krabbenhøft, K., Lyamin, A.V., Sloan, S.W., 2007. Formulation and solution of some
during the topology optimisation process. Validation confirms the effi­ plasticity problems as conic programs. Int. J. Solids Struct. 44 (5), 1533–1549.
cacy of our approach, which has been successfully applied to various Li, X., Zhang, X., Zhang, Y., 2023. Three-dimensional plasticity-based topology
optimization with smoothed finite element analysis. Comput. Mech. https://doi.org/
aspects of geostructure design, including embankments and soil 10.1007/s00466-023-02378-9.
foundations. Meng, J., Zhang, X., Huang, J., Tang, H., Mattsson, H., Laue, J., 2020. A smoothed finite
While designs resulting from the proposed topology optimisation element method using second-order cone programming. Comput. Geotech. 123,
103547.
method often exhibit complexity and pose challenges for practical en­ Mourad, L., Bleyer, J., Mesnil, R., Nseir, J., Sab, K., Raphael, W., 2021. Topology
gineering implementation, this study demonstrates the potential for optimization of load-bearing capacity. Struct. Multidiscip. Optim. 64 (3),
further simplification. By streamlining the designs from the proposed 1367–1383.
Nguyen, H.C., 2023. A mixed formulation of limit analysis for seismic slope stability.
method, we can achieve more practical and easily implementable so­
Géotech. Lett. 13 (1), 54–64.
lutions, which may not be immediate apparent. This work offers a ver­ Pucker, T., Grabe, J., 2011. Structural optimization in geotechnical engineering: basics
satile and efficient methodology for creating geostructure designs that and application. Acta Geotech. 6, 41–49.
balance material usage and structural stability. Ren, G., Smith, J.V., Tang, J.W., Xie, Y.M., 2005. Underground excavation shape
optimization using an evolutionary procedure. Comput. Geotech. 32 (2), 122–132.
Ren, G., Zuo, Z.H., Xie, Y.M., Smith, J.V., 2014. Underground excavation shape
CRediT authorship contribution statement optimization considering material nonlinearities. Comput. Geotech. 58, 81–87.
Seitz, K.-F., Grabe, J., 2016. Three-dimensional topology optimization for geotechnical
foundations in granular soil. Comput. Geotech. 80, 41–48.
Xifan Li: Software, Methodology, Writing – original draft. Xue Sigmund, O., 2007. Morphology-based black and white filters for topology optimization.
Zhang: Writing – review & editing, Supervision, Software, Methodol­ Struct. Multidiscip. Optim. 33 (4), 401–424.
ogy, Funding acquisition, Conceptualization. Yujia Zhang: Software, Sigmund, O., Maute, K., 2013. Topology optimization approaches: A comparative
review. Struct. Multidiscip. Optim. 48 (6), 1031–1055.
Methodology. Smaoui, H., Kammoun, Z., 2022. Convergence of the direct limit analysis design method
for discrete topology optimization. Optim. Eng. 23 (1), 1–24.
Declaration of competing interest Walker, J., Yu, H.-S., 2010. Analysis of the cone penetration test in layered clay.
Géotechnique 60 (12), 939–948.
Zhang, X., Oñate, E., Torres, S.A.G., Bleyer, J., Krabbenhoft, K., 2019. A unified
The authors declare that they have no known competing financial Lagrangian formulation for solid and fluid dynamics and its possibility for modelling
interests or personal relationships that could have appeared to influence submarine landslides and their consequences. Comput. Methods Appl. Mech. Eng.
343, 314–338.
the work reported in this paper.
Zhang, X., Li, X., Zhang, Y., 2023a. A framework for plasticity-based topology
optimization of continuum structures. Int. J. Numer. Meth. Eng. 124 (7), 1493–1509.
Data availability Zhang, Y., Zhang, X., Nguyen, H., Li, X., Wang, L., 2023b. An implicit 3D nodal
integration based PFEM (N-PFEM) of natural temporal stability for dynamic analysis
of granular flow and landslide problems. Comput. Geotech. 159, 105434.
The data on optimisation algorithm are available from the corre­
sponding author upon reasonable request. All the details necessary to

10