1 s2.0 S0141029621005022 Main

Engineering Structures 240 (2021) 112352
Contents lists available at ScienceDirect
Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct
Designing bending-active gridshells as falsework for concrete shells through

numerical optimization
Jef Rombouts a, b, Andrew Liew c, Geert Lombaert d, Lars De Laet b, Philippe Block c,
Mattias Schevenels a, *
a
KU Leuven, Faculty of Engineering Science, Department of Architecture, Kasteelpark Arenberg 1 - box 2431, 3001 Leuven, Belgium
b
Vrije Universiteit Brussel, Faculty of Engineering, Department of Architectural Engineering, Belgium
c
ETH Zurich, Institute of Technology in Architecture, BLOCK Research Group, Switzerland
d
KU Leuven, Faculty of Engineering Science, Department of Civil Engineering, Belgium
A R T I C L E I N F O A B S T R A C T
Keywords: Shell structures are material efficient structures capable of covering large spans with minimum weight. If con
Bending-active gridshells crete is used as the building material for the shell, it must be initially supported by a formwork. These generally
Active bending rigid formworks, and their supporting falsework structure, are time and material consuming to construct.
Numerical optimization
Recently, researchers have recommended the use of elastic gridshells as falsework system. Unfortunately, the
Flexible formwork
Concrete shells
design of such a system is complicated by large deformations of the bending elements. Moreover, since gridshells
are inherently flexible, significant displacements generally occur when applying the wet concrete.
To overcome these difficulties, this paper presents a design tool to effectively design a gridshell serving as the
falsework for a concrete shell. The gridshell is assumed to be constructed from an initially flat grid of straight
slender rods, using cables with adjustable lengths to brace the erected gridshell. An optimization algorithm is
proposed that fits the shape of the gridshell under the wet concrete loading to a given target shape by manip
ulating the lengths of the bracing cables. Meanwhile, the optimization algorithm reduces the construction effort
by minimizing the number of cables, and the diameter of the grid rods. Moreover, feasibility of the design is
ensured by constraining the axial force in the cables and limiting the displacements under increased loading. A
gradient-based optimization scheme is adopted, and an implicit dynamic relaxation approach previously
developed specifically for the simulation of bending-active structures is used to solve the nonlinear equilibrium
equations. Nonlinear effects such as buckling of the grid rods are taken into account by using co-rotational beam
elements. The resulting design tool is applied for the design of a scale model, demonstrating the feasibility of the
suggested falsework system. The optimized design fits the target better, uses less cables, and is more resistant to
additional loading compared to designs from existing design approaches. Comparisons between the numerical
and physical models show deviations of up to around 2% of the larger span for the design load and smaller
deviations for smaller loads, with differences attributed to small model scaling effects.
1. Introduction This makes on-site construction more challenging. If concrete is used as

the building material for the shell, a full and rigid formwork is usually
Shell structures are material efficient structures capable of covering required to hold the wet concrete before the structure can carry its own
large spans with minimum weight. When they are properly designed and weight. These rigid formworks, and their supporting falsework struc
built, internal bending moments are minimized or avoided altogether, ture, are time and material consuming to make, as most of the formwork
and the external loads are carried exclusively through membrane action. components have to be customized to fit the required shape. Conse
The extent to which the structure acts as a membrane rather than a plate quently, despite their inherent material efficiency, concrete shells are
is determined by the shape of the shell. This shape needs some curva usually too expensive to build in practice.
ture, preferably in all directions to attain stiffness through its geometry. With the aim of making concrete shells a more viable structural
* Corresponding author.
E-mail address: mattias.schevenels@kuleuven.be (M. Schevenels).
https://doi.org/10.1016/j.engstruct.2021.112352
Received 25 November 2020; Received in revised form 14 March 2021; Accepted 30 March 2021
Available online 3 May 2021
0141-0296/© 2021 Elsevier Ltd. All rights reserved.
J. Rombouts et al. Engineering Structures 240 (2021) 112352
solution, researchers have been working on the development of flexible complicated by the large deformation associated with the bending
formwork instead of traditional rigid formwork to reduce the time and erection process.
material spent on the construction of the shell. Flexible formwork
generally consists of a membrane on which the concrete is cast. If the Hernández [17] compared several design approaches to fit a grid
membrane is not strong enough to span the whole surface by itself, shell to a given target shape. However, because a gridshell is usually
falsework or air pressure is required to support and shape the mem allowed to relax its shape before an additional bracing layer ’fixes’ that
brane. The falsework consists of rigid elements, cables, bending-active shape by eliminating the initial shear degree of freedom, significant
elements, or a combination of these. Because this formwork is inher deviations from the intended target shape can be expected. Moreover, if
ently flexible, a common challenge is to limit the deformations of the the gridshell is to be used as the falsework for a concrete shell, the
formwork under the weight of the wet concrete. This challenge is usually weight of the wet concrete will most likely cause even larger de
addressed by applying the concrete in layers, gradually increasing formations and deviations from the target shape. The deviations due to
stiffness and weight. Moreover, because small variations in shape can relaxation can be counteracted by using the bracing to fix the shape of
cause a large drop in performance of the concrete shell [1,2], it is the gridshell before it is allowed to relax. Alternatively, Pignatelli et al.
important for the formwork to accurately match the designed shape. A [18] proposed a genetic algorithm to find the shaping forces needed to
short overview of each type of flexible formwork is given below. push or pull the grid towards the intended target after the grid has been
erected. However, to our best knowledge, design tools taking de
1. Membrane formwork supported by rigid elements already leads to a formations due to loading into account are still lacking.
high reduction in construction effort and material use compared to This paper presents a framework for the design of bending-active
rigid formwork. As a result, this type of formwork quickly became gridshells as a falsework for concrete shells. As a key part of the
very popular when it was first applied during the second world war design framework, a novel optimization method is developed to fit the
by Waller and Billig [3]. Their Ctesiphon system uses jute fabric supporting gridshell to a given target shape by accurately simulating the
stretched over a series of rigid arches. However, the rigid elements bending process, while taking the wet concrete loading into account to
can only be reused in near identical designs, which limits the make sure the gridshell fits the exact target shape in its loaded condition.
adaptability and design freedom of this system. The cross-section of the grid rods and the undeformed lengths of the
2. Pneumatic formwork has the advantage that it is light, compact, re- braces are chosen as the design variables. The design framework as
usable and fast to erect. A comprehensive overview of the history of sumes full circular cross-sections for the grid rods, and a bracing system
pneumatic formwork is given by Kromoser and Huber [4]. The main with adjustable lengths, using for example turnbuckles or lashing straps,
challenge related to this type of formwork is the large deformation to manipulate the gridshell’s shape after its relaxation. In contrast to the
caused by the weight of the wet concrete [5], which is detrimental approach proposed by Pignatelli et al. [18], our approach requires no
for the structural performance of the final structure [6]. Moreover, additional external struts or ties to shape the grid, as we use the bracing
Hennik and Houtman [7] showed that designing a pneumatic not only to stiffen the gridshell, but also to manipulate its shape after
formwork for a given target shape is far from trivial. A final disad erection.This enables us to keep the space underneath the shell free of
vantage is that pneumatic formwork covers the whole construction structural elements. Additionally, the number of braces and the radius of
area underneath the shell during casting and curing of the concrete, the cross-section of the grid rods are minimized, and feasibility of the
which makes this system unusable for use over existing structures, design is ensured by constraining the axial force in the braces. Finally,
rivers, or roads that have to remain operational during construction. the robustness of the optimized design with respect to load uncertainties
3. Membranes supported by cable nets do not obstruct the space un is improved by limiting the displacements under an additional load case.
derneath the shell during construction. Van Mele and Block [8], and The proposed design approach is tested by designing and building a
Veenendaal and Block [9] developed a design approach for this type structurally representative scale model.
of formwork. Motivated by the fact that the structural properties of a In a recently published paper, the authors describe a novel shape
shell are potentially very sensitive to their shape, they employed optimization approach for bending-active gridshells [19]. The objective
numerical optimization to fit a cable-net to a target shape, while of the current paper is different, but the setup of the structural model is
taking the weight of the wet concrete into account. Later, Liew et al. identical. Hence, also small parts of the sensitivity analysis are identical.
[10] extended the approach to adaptively account for design im Nevertheless, both the structural model and the sensitivity analysis are
perfections. With a large prototype for an actual roof, Block et al. described in full in the current paper (the latter in the appendix) for a
[11,12] showed the potential of this type of formwork. However, a better comprehensibility.
drawback is its limitation to anticlastic shapes, while shells acting in The remainder of this paper is organized as follows. First, Section 2
pure compression under gravitational loading generally have some gives an overview of all steps in the proposed design framework for a
form of synclastic curvature. gridshell acting as the falsework for a concrete shell. Next, Section 3
4. The use of bending-active elements in formwork for concrete shells is explains how the gridshell is modeled using finite elements. Section 4
relatively new. Popescu et al. [13] incorporated just a few bending- discusses the design and construction of a scale model, and finally,
active elements in a knitted flexible formwork. Tang and Pedreschi conclusions are given in Section 5.
[14], Coar et al. [15] and Cuviliers et al. [16] on the other hand,
proposed the use of bending-active gridshells as (lost) falsework. 2. Design framework
Because bending-active gridshells are generally erected from an
initially flat grid of continuous rods, with pinned connections be This section describes the framework for the design of bending-active
tween crossing rods, they do not require the production of specially gridshells as the falsework for concrete shells. The goal of this frame
customized beams or joints. This simplifies the construction process work is to design a gridshell that closely approximates a given target
and increases the reusability of the components. Moreover, the shape under the weight of the wet concrete. Additionally, the con
flexibility of the rods, in combination with the shear degree of struction effort and material use should be minimized by reducing the
freedom provided by the pinned connections, allows the structure to number of braces and the cross-section of the grid rods. Moreover, the
be deformed into complex curved shapes which can have both bending moments in the rods and normal forces in the braces should be
anticlastic and synclastic curvature. Consequently, these structures limited to ensure that the design is feasible. Finally, the design should be
have high potential to be used as falsework for concrete shells, and robust by limiting deviations from the target shape under increased
they also allow to keep the construction area underneath the shell loading. The design framework assumes a regular grid, round rod sec
free. However, despite their construction efficiency, their design is tions, and a bracing system with adjustable lengths. Moreover, the
2
wet concrete loading. Each step in the flowchart is explained in the

following paragraphs.
Input parameters
The target shape is determined in advance by the designer as the
desired shape for the actual concrete shell. This can be any smooth
shape, with both positive and/or negative Gaussian curvature, as long as
the maximum curvature does not exceed the maximum achievable
curvature of the used rods. However, since the rods’ cross-section is
determined later in the design framework, it is not possible to know the
maximum allowable curvature in the shell beforehand. Additionally, the
designer must choose the distance between grid rods, based on practical
considerations. Next, the mechanical properties of the cables should be
determined. Finally, the designer must choose a catalogue of available
rod sections.
(1) Generation of Chebyshev net
The first operation of the design framework is to draw a potential
grid or Chebyshev net on the target surface. Several techniques are
described in the literature, which can be categorized into two groups
[17]. Methods in the first group treat the problem as a purely geomet
rical problem. The most commonly used method in this group is the
compass method, originally proposed by Frei Otto [20], which requires
the designer to choose a mesh density, and two intersecting curves on
the surface. Other methods use numerical optimization to draw a grid on
[21,22], or near the target surface [23], allowing additional control over
some grid characteristics such as the curvature of the lines or angle
between line segments. The second group treats the problem as a ficti
tious physical problem, pulling or pushing a flat grid towards the target
surface using fictitious forces [24], springs [25], or geometric con
straints [26]. In this paper, we follow Kuijenhoven and Hoogenboom
[25] as we pull an over-sized grid towards the target surface with
fictitious springs, trimming excess parts, and using implicit dynamic
relaxation to solve the nonlinear equilibrium equations [27]. Compared
to a purely geometrical approach, this approach has the advantage that
the grid has a certain stiffness. Consequently, the grid will relax to a state
of low deformation energy, as the rods try to straighten out as much as
possible on the surface. Additionally, the stiffness of the connections can
be tuned to control the angle between crossing rods. Furthermore, the
forces in the cables can be used to determine to what extent the grid is
forced to the target surface. However, the fact that the material prop
erties are entirely fictitious at this stage makes it hard to directly choose
appropriate values for these properties. A trial and error approach is
often required to obtain the desired geometry. For this reason, other
techniques are potentially worth considering at this stage.
The resulting Chebyshev net is used to determine the length and
connectivity of each rod in the grid, the locations of the supports, and
the maximum allowable cross-section, derived from the maximum cur
vature in the grid. The maximum allowable cross-section can be derived
directly from the internal forces generated by the nonlinear finite
element solver. Assuming a circular cross-section, the maximum
allowableradius for the grid rods is given by:
If f
Fig. 1. Flowchart of the proposed design framework. rg,max = √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ (1)
M 2y + M 2z
bracing elements are assumed to only withstand tensile forces. Conse
quently, they are simply referred to as cables in the following. where I is the moment of inertia, ff is the longitudinal flexural strength,
A flowchart of the proposed design framework is given in Fig. 1. The and My and Mz are the bending moments around both principal axes.
design process consists of two parts. In the first part, the initial design is Additionally, the Chebyshev netwill serve as the target grid in the sec
generated, following a similar approach to the one proposed by Pigna ond part of the design framework.
telli et al. [18]: a so-called Chebyshev net is drawn on the target surface (2) Initial design
to extract the dimensions and layout of the corresponding flat grid, lo The second operation entails the generation of an initial design,
cations of the supports and maximum cross-section of the grid rods. This using the grid layout and support locations from the previous step, and
information is used to generate the initial design. Moreover, the Che assigning the maximum allowable cross-section to all grid rods. More
byshev net is used as a target grid in the second part, where the unde over, cables are modeled at all allowable positions. The cables are given
formed length of the cables and the cross-section of the grid rods are a certain level of prestress by setting their undeformed lengths l0c to a
optimized so that the resulting gridshell fits the target shape under the fraction of their current lengths lc (e.g. l0c = 0.99lc ), which is defined as
3
the distances between the nodes they connect in the target grid, can be added to the optimization problem, e.g. to account for non-
generated in the previous step. This layout of cables serves as a ground uniform loading during the pouring of concrete if it is not applied in
structure(i.e. a structure containing all possible cable elements) for the layers. After the optimization process converges, cables that are not
optimization in steps 3 and 4. In order for the optimization to converge, tensioned under the design load are removed from the numerical model.
the loaded gridshell in its current, initial design should be close enough The optimization problem is solved by means of a gradient-based
to the target grid: it should not show excessive deformations like global optimizer. The gradients of the terms in the objective function and the
snap-trough buckling under the wet concrete loading. If excessive de constraints are determined analytically. This sensitivity analysis is
formations do occur, a denser grid should be chosen in the first step. described in appendix A, along with a verification of the resulting ex
(3) Optimization of rod cross-section and cable lengths pressions by means of the finite difference method.
If the deformations of the initial design under loading are acceptable, (4) Optimization of cable lengths
the prestress of the cables and cross-section of the rods can be optimized. Although the radius of the cross-section of the rods is a discrete
This third operation in the flowchart forms the key aspect of the design variable in reality, it is optimized as a continuous variable. Conse
framework. The optimization problem is formulated as follows: quently, the optimized radius must be rounded to the nearest higher
radius available in the chosen section catalogue. This may affect the
∑
kc
min f = w1 dT d + w2 nc,i + w3 rg (2) performance and feasibility of the optimized gridshell. Therefore, as a
l0c ,rg i=1 fourth and final operation in the design framework, the initial lengths of
the cables are optimized again with the same objective and constraints,
s.t. l0c,min ⩽l0c ⩽l0c,max (3) while keeping the cross-section of the rods fixed. Rounding the sections
up makes the grid stiffer, which means more cables and/or larger cable
rg,min ⩽rg ⩽rg,max (4) forces are needed to deform the grid towards the target shape. On the
other hand, the grid will be less sensitive to additional load cases.
nc ⩽nc,max (5) Therefore, it is likely that a feasible solution can be found in this step if a
feasible solution was found in the previous step, while the objective
dT1 d1 ⩽d 2 (6) function increases slightly. As an extension to the proposed method, one
might consider the use of a penalization or projection strategy to steer
where the undeformed length of each cable, collected in a vector l0c and the continuous rod section towards a discrete solution – this might lead
the radius of the cross-section of all grid rods rg are the design variables, to a better optimum than rounding and re-optimizing. However, as
w1 ,w2 , and w3 are weighting factors, d is a vector collecting the distances opposed to penalization or projection strategies, rounding and re-
between each node and the corresponding node in the target grid, nc,i is optimizing ensures that a perfectly discrete solution is obtained.
the normal force in cable i and the i-th element of nc , kc is the number of
cables, l0c,min and l0c,max are vectors collecting the minimal and maximal 3. Finite element model
initial length for each cable, rg,min and rg,max are the minimal and maximal
cross-section radius, given by the chosen section catalogue, nc,max is a The behavior of the gridshell is modeled with nonlinear finite ele
vector collecting the maximum allowable tensile force in each cable, d1 ments, using implicit dynamic relaxation to iteratively find the struc
is a vector collecting the distances between each node and its corre ture’s equilibrium position. For more details on implicit dynamic
sponding node in the target grid for an additional load case, and d2 is a relaxation,which has been developed specifically for the simulation of
maximum allowable value for dT1 d1 . The distances collected in vectors d bending-active structures, the reader is referred to [27,28].
and d1 are measured as the Euclidian distances between the current grid The structure is in equilibrium when the external loads are balanced
nodes and the nodes of the target grid generated in the first step of the by the internal forces, i.e. all residual forces are zero:
design framework. For node i: r = p− f = 0 (8)
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
di = (xi − xti )2 + (yi − yti )2 + (zi − zti )2 (7) where r are the residual forces, f are the internal forces, and p are the
external loads. In practice, the residual forces will never be exactly zero,
where di is the distance of node i and the i-th element of vector d, xi , yi , and the structure is assumed to be in equilibrium when the solver re
and zi are the x-, y-, and z-coordinates of node i, and xti ,yti , and zti are the duces the residual forces below a certain threshold. The grid rods are
x-, y-, and z-coordinates of the corresponding target node. modeled using co-rotational beam elements [29], a short description of
Eq. (2) shows that the objective is threefold; first, the goal is to which is given next in Section 3.1.In order to account for the eccentricity
minimize the distance between the loaded gridshell and the target grid. of crossing rods, two separate nodes are defined for every connection.
Second, the number of cables is penalized using L1 regularization, which The connections are modeled as short co-rotational beam elements be
effectively reduces the tension in unnecessary cables to zero. Third, the tween these nodes, with low or no torsional stiffness to allow rotations in
rod section is minimized to reduce the material use and the construction the plane of the grid during erection. These elements do have non-zero
effort by reducing the magnitude of the erection forces required to bend axial and bending stiffness – the actual values must be determined
the grid into its final shape. The share of each part of the objective experimentally. Cables are modeled as beam elements with no torsional
function is managed by tuning the weighting factors w1 , w2 , and w3 , for or bending stiffness, and no stiffness in compression.
which the following units are chosen: w1 [m− 2 ], w2 [N− 1 ], w3 [m− 1 ]. This
results in a dimensionless objective function.
3.1. Co-rotational beam elements
Box constraints on the undeformed cable lengths l0c (Eq. (3)) can be
included to improve convergence by limiting the design domain. Box
Co-rotational beam elements are specifically developed to handle
constraints on the radius rg of the cross-section of the rods (Eq. (4))
large displacements [30]. They allow arbitrarily large translations and
ensure that the optimized rod section is within the bounds of available
rotations, as long as the strains remain small. This is achieved by
cross-sections of the chosen section catalogue. A constraint on the
decoupling rigid body motion and local beam deformations. The basis of
normal force in the cables nc (Eq. (5)) ensures that the design is feasible.
the co-rotational beam element as described by Crisfield [29] is briefly
Finally, a constraint on the deviation from the target grid under an
reviewed here. A more extensive explanation can be found in [29] or
additional load case (Eq. (6)) makes the design more robust with respect
[31].
to load variations.If needed, similar constraints for additional load cases
To distinguish between rigid body motion and local beam de
4
Fig. 3. Orientation of the local beam deformations (δ θx,a θy,a θz,a θx,b θy,b θz,b )
Fig. 2. Definition of the nodal basis vectors (ax , ay , az ) and (bx , by , bz ) as well as and corresponding internal forces (N Mx,a My,a Mz,a Mx,b My,b Mz,b ).
the element basis vectors (ex , ey , ez ) of a deformed beam element.
two principal axes at the beam ends.
formations, a local co-rotated reference configuration is fitted to the The local rotational deformations θ are expressed in terms of the
actual deformed beam. Local deformations are traced by comparing the nodal and element basis vectors:
actual deformed beam to the co-rotated reference beam. The co-rotated
reference beam is represented by a set of local element basis vectors 2sin(θx,a ) = − (az )T ey + (ez )T ay
Te = [ex ey ez ], where ex lies in the direction of the vector connecting the 2sin(θx,b ) = − (bz )T ey + (ez )T by
2sin(θy,a ) = − (az )T ex + (ez )T ax
beam ends. This set of element basis vectors is constructed from the (12)
2sin(θy,b ) = − (bz )T ex + (ez )T bx
known current beam end orientations, which are in turn represented by
2sin(θz,a ) = − (ay )T ex + (ey )T ax
the nodal basis vectors Ta = [ax ay az ] and Tb = [bx by bz ] (Fig. 2). Note
2sin(θz,b ) = − (by )T ex + (ey )T bx
that ax and bx point in the direction of the beam’s centerline, while ay
and az , and by and bz point in the direction of the cross-section’s prin where θx , θy and θz are the deformation angles of the beam ends around
cipal axes. For the actual procedure to construct Te , the reader is the local x-, y-, and z-axis respectively (Fig. 3).
referred to the original paper [29]. For an undeformed beam, both sets The axial deformation δ is simply the difference between the current
of nodal basis vectors are equal to the element basis vectors.
element length l and the undeformed element length l0 :
In each step of the nonlinear solver, the nodal basis vectors are
updated using Rodrigues’ rotation formula [31,32]. For beam end a we δ = l − l0 (13)
get:
with the length defined as the Euclidian distance between beam ends a
Tta = R(Δλa )Tt−a Δt
(9) and b:
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
where Tta collects the current nodal basis vectors, Tt−a Δt collects the nodal l = (xa − xb )2 + (ya − yb )2 + (za − zb )2 (14)
basis vectors of the previous iteration, and R(λ) is an orthogonal rotation
matrix describing the rotation around a given pseudo-vector λ = where xa is the x-coordinate of node a, xb is the x-coordinate of node b,
{α β γ}T over a given angle λ = ‖λ‖, where the double vertical lines and so on.
denote the Euclidean norm: The local internal forces are expressed in terms of the local de
formations:
sin(λ) 1 − cos(λ)
R(λ) = I + S(λ) + S(λ)S(λ) (10)
(15)
e e
λ λ2 f = K ue
where I is a three-by-three identity matrix, and S is a skew-symmetric e

where f is the vector collecting all local element forces (Fig. 3):
matrix defined as:
(16)
e
⎡ ⎤ f = [N Mx,a My,a Mz,a Mx,b My,b Mz,b ]T
0 − γ β
S(λ) = ⎣ γ 0 − α⎦ (11)
− β α 0 and ue is the vector collecting all local deformations (Fig. 3):
ue = [δ θx,a θy,a θz,a θx,b θy,b θz,b ]T (17)

In Eq. (9), the vector Δλa = [Δαa Δβa Δγa ]T is the rotation vector
describing the change in orientation of beam end a between this itera e
and K is the local element stiffness matrix, which is based on
tion and the previous one. In our case, these angle changes are calcu
Euler–Bernoulli beam theory. Furthermore, the material is assumed to
lated by the implicit dynamic relaxation solver. The updating procedure
be homogeneous, isotropic, and linearly elastic:
for the nodal basis vectors of beam end b is analogous. ⎡ ⎤
The deformation of the beam is described by seven independent local EA 0 0 0 0 0 0
beam deformations1, recorded at the beam ends. An axial deformation δ ⎢ 0
⎢ GJ 0 0 − GJ 0 0 ⎥ ⎥
⎢ 0 0 ⎥
gives rise to a normal force N, and six rotational deformations θ cause six 1⎢⎢ 0 4EI y 0 0 2EI y ⎥
(18)
e
corresponding bending moments M. These rotational deformations K = 0⎢ 0 0 0 4EI z 0 0 2EI z ⎥
⎥
l ⎢
⎢ 0 − GJ 0 0 GJ 0 0 ⎥
describe the beam’s twist as well as bending around the cross-section’s ⎣ 0
⎥
0 2EI y 0 0 4EI y 0 ⎦
0 0 0 2EI z 0 0 4EI z
1
Referred to as ‘strains’ in the original paper. However, because the con e
The global element forces f e are related to the local forces f as follows:
cerning quantities can not be regarded as strains in the strict mechanical sense,
the term ‘deformations’ is used here instead.
5
Fig. 5. Openings in the shell design were closed to define the target shape for
the supporting gridshell.
Fig. 4. A funicular shell is designed using RhinoVault, where the two-

dimensional form (left) and force (right) diagrams are manipulated to alter
the three-dimensional force network (top).
Fig. 9. Top view. Loads (black dots) are applied every two nodes in the area
(19)
e
f e = Te f
covered with concrete.
where f e = [fx,a fy,a fz,a mx,a my,a mz,a fx,b fy,b fz,b mx,b my,b mz,b ]T is the
how the methodology proposed in this paper can be used to overcome
vector containing forces and moments in the global x-, y-, and z-axis for
these issues. Therefore, design choices are made specifically to uncover
beam end a and beam end b, and Te is a transformation matrix defined as
possible problems. It is not our aim to make an exhaustive design and
( e )T
∂u analysis fulfilling all relevant norms.
Te = (20)
∂xe
where xe = [xa ya za αa βa γa xb yb zb αb βb γ b ]T is a vector collecting 4.1. Design

translational and rotational coordinates of beam end a and beam end b
in the global x-, y-, and z-direction. For the actual calculation of the As a preliminary step, the input parameters for the design framework
transformation matrix Te , the reader is referred to the original paper were determined. A concrete shell was designed using RhinoVault [33],
[29]. The full internal force vector f is assembled from the element which is a plug-in for Rhino to design funicular shells. The plugin is
forces f e , with e = 1… m, where m is the number of elements. based on the Thrust Network Analysis method, developed by Block and
The element tangential stiffness matrix is obtained from differenti Ochsendorf [34]. The 3D form diagram, and its corresponding 2D form
ation of the internal forces f e (Eq. (19)) with respect to the element and force diagrams are shown in Fig. 4. Although the envisioned shell
position vector xe : has a few openings, we decided that the supporting gridshell had to rest
e
completely on the ground. Therefore, the openings in the shell were
df e ∂f e ∂f due ∂f e dTe closed to define the actual target shape for the gridshell falsework
Ke = = e +
dxe ∂f ∂ue dxe ∂Te dxe (21) (Fig. 5).
=TKT e e eT
+ Keg = Kee + Keg Requirements regarding the workability of the grid rods and their
availability in the required scale limited the options for the rods’ cross-
The resulting stiffness matrix consists of an elastic part Kee and a geo sections. Therefore, the cross-section of the rods was chosen beforehand,
metric part Keg . The derivation of the geometric stiffness matrix Keg can implying that only the undeformed lengths of the bracing cables were
selected as design variables, and the fourth step of the design framework
be found in the original paper [29]. The element stiffness matrix Kei for
(Fig. 1) was omitted. The chosen rods have a diameter of d = 2 mm, and
each element i = 1… m is used to assemble the full tangential stiffness
are made of glass-fibre reinforced polymer, with a Young’s modulus of
matrix K.
40 GPa. Each rod segment spanning between two connections was
modeled with four co-rotational beam elements to capture potential
4. Prototype
local beam buckling with sufficient accuracy, while keeping the required
computation time acceptable. The spacing between grid rods was set to
In order to validate the proposed design framework, and to test the
55 mm. The supports were modeled as pinned. The bracing elements
proposed way of controlling the gridshell’s shape by changing the
were modeled as single cable elements with an axial stiffness of EA =
lengths of the cables, a scale model has been designed and built. The
187 N, based on preliminary tests. Finally, the connections between grid
gridshell model has a footprint of around 0.9 m by 0.6 m. Although the
rods were modeled as short beam elements perpendicular to the grid
model was designed as an independent structure on scale 1:1, it can be
shell’s surface, simulating the eccentricity between crossing rods.
regarded as a 1 to 20 scale representation of an actual gridshell. The
The structural properties of the connections were initially guessed.
following Section 4.1 describes the design of the model. Its construction
However, it was observed that they had a significant influence on the
is discussed in Section 4.2.
structural behavior of the model. Therefore, in a second stage, they were
The purpose of this prototype example it to show the possible chal
calibrated based on data from a test on the physical model, before
lenges that can be encountered during the design and construction of a
bracing was applied. The test data consisted of two 3D scans: one
gridshell serving as the formwork of a concrete shell, and demonstrate
without loading and one with point loads of 300 g on the locations
6
Table 1
The scale model is equivalent to a gridshell of 20 times its size, carrying a 5 cm
thick layer of concrete.
scale model equivalent
gridshell
length scale 1/1 20/1

weight scale 1/1 400/1
gridshell length 0.90 m 18.0 m

gridshell width 0.55 m 11.0 m
gridshell height 0.21 m 4.2 m Fig. 7. The initial, flat grid is derived from the target grid.
distance between rods 0.055 m 1.1 m
radius of cross-section of rods 0.001 m 0.02 m
shell thickness 0.0025 m 0.05 m
density of concrete 50000 kg/m3 2500 kg/m3
distributed load 1.2 × 103 N/ 1.2 × 103 N/m2
m2
equivalent point load per connection 3.7 N 1.5 × 103 N
node*
*
Approximation assuming a square tributary area.
Fig. 8. The locations of the supports is derived from the target grid.
shown in Fig. 9. The scans measured the location of 42 nodes distributed

over the gridshell using a Vicon motion capture setup [35]. To calibrate
the structural parameters of the connections, an inverse problem had to
be solved, which was achieved by slightly adapting the numerical
optimization approach described above: the structural parameters of the
connections were selected as design variables to minimize the discrep
ancy between predictions by the numerical model and data obtained
from the test on the physical model. The resulting parameters are: EA =
7.15 × 104 N, GJ = 7.38 × 10− 4 Nm2 , EIy = EIz = 7.85 × 10− 4 Nm2.
These parameters were validated by comparing predictions and mea
surements for an additional load case. For a large-scale gridshell, we
suggest to determine the connection properties by means of experiments
on a full-scale mockup structure consisting of at least two crossing bars.
A comparison between the grid properties of the scale model and an
equivalent large-scale gridshell is shown in Table 1 to more easily
interpret the dimensions and loads in the model. The approximated
equivalent point load per connection node is 3.7 N for the chosen grid
spacing, corresponding to 5 cm of concrete, assuming a square tributary
area for each load. However, for practical reasons, loads are only applied
every 2 connection nodes. Consequently, the weight of all loads was set
to 700 g after rounding and compensating for the fact that the tribu
taryareas are no longer perfectly square for the erected gridshell.
4.1.1. Generation of target grid

Next, a target grid was generated by pulling an oversized flat grid
towards the surface using implicit dynamic relaxation to simulate the
bending process. The grid was forced on the surface by assigning a force
density to a series of fictitious springs that were allowed to slide over the
surface. After a predefined number of iterations of the nonlinear solver,
the oversized grid fitted the target surface, and excess parts were trim
med (Fig. 6). The corresponding initial, flat grid (Fig. 7) and and the
positions of the supports (Fig. 8) were derived from the trimmed grid.
Moreover, the maximum allowable radius for the cross-section of the
grid rods was found to be 1.3 mm using Eq. (1). This is larger than the
cross-section we chose, meaning that the chosen cross-section is feasible.
4.1.2. Generation of initial design

Next, the initial design for the optimization was generated by
assigning the actual physical properties to all elements in the finite
element model used in the previous step. Although the flowchart (Fig. 1)
Fig. 6. A target grid is generated by pulling an oversized flat grid towards the suggests to assign the maximum allowable cross-section, the cross-
target surface and trimming away excess parts. section chosen previously is assigned instead, as it is not considered as
7
Fig. 12. Deviation from target grid (indicated in color) without loads. (For interpretation of the references to colour in this figure legend, the reader is referred to the
web version of this article.)
Fig. 13. Deviation from target grid (indicated in color) under the design load case. (For interpretation of the references to colour in this figure legend, the reader is
referred to the web version of this article.)
8
Fig. 14. Deviation from target grid (indicated in color) under the additional load case. (For interpretation of the references to colour in this figure legend, the reader
is referred to the web version of this article.)
Fig. 10. Convergence history of the objective function for a fixed rod diameter
of 2 mm. The optimizer reduces the objective function f from 1 to 0.54 in 49
∑c
iterations by reducing both terms f1 = w1 dT d and f2 = w2 ni=1 Ni . Fig. 11. Feasibility of the design with a fixed rod diameter of 2 mm during
optimization. c1 = (dT1 /d1 )/d2 − 1 is reduced from 23.4 to 0.0, while
a design variable in this case. Cables are placed on every feasible loca c2 = nc /nc,max − 1 increases from − 0.7 to − 0.4, which means that the optimized
tion, excluding cables that would have to be connected to the supports of design is feasible.
the gridshell and cables that would rise above the shell surface, as these
would hinder the application of the concrete. The undeformed lengths of manageable for the optimizer, and, consequently, we can proceed to the
the cables were initially set to 99 % of the distances between the nodes next step.
they connect in the target grid, determined in the previous step. The
initial design was subjected to point loads of 700 g (6.9 N), placed on 4.1.3. Optimization of undeformed cable lengths
every 2 nodes in the area that would be covered by the actual shell In this step, the undeformed lengths of the cables are optimized ac
(Fig. 9). Furthermore, the initial design was subjected to an additional cording to the optimization approach discussed in Section 2. The
load case of 1.2 times the design load in order to constrain the de weighting factors are chosen so that the value of the objective function
formations from this additional load case in the following optimization (2) is 1 for the initial design, and the first and second term respectively
step. If the initial design deforms too heavily under these load condi initially contribute 70 % and 30 % to make sure the optimized design fits
tions, it is too far from its optimum for the optimization algorithm to the target sufficiently well, while reducing the number of cables as much
converge. Figs. 12b, 13b, and 14b show the unloaded initial design, the as possible. Because only the undeformed cable lengths are chosen as the
initial design loaded with 6.9 N per two connection nodes, and the initial design variables for this case, the last term in Eq. (2), minimizing the
design loaded with 8.2 N per two connection nodes, respectively, where cross-section of the rods, is excluded by setting the corresponding
the colors indicate the distance to the target grid. Although deviations weighting factor w3 to zero. The resulting values for the other two
from the target shape are large for the additional load case, they are still
9
Fig. 15. Convergence history of the objective function for a variable rod
diameter. The optimizer reduces the objective function f from 1 to 0.52 in 49 Fig. 16. Feasibility of the design with a variable rod diameter during optimi
∑c
iterations by reducing the terms f1 = w1 dT d, f2 = w2 ni=1 Ni , and f3 = w3 rg . zation. c1 = (dT1 /d1 )/d2 − 1 increases from − 0.78 to − 0.05, and
c2 = nc /nc,max − 1 increases from − 0.37 to − 0.01, which means that the opti
weighting factors are: w1 = 41 m− 2 and w2 = 5.3 × 10− 4 N− 1.Con mized design is feasible.
straints (3), (5), and (6) were taken into account, where each value of
l0c,min was set to 1 cm, l0c,max was set to 1.1 times the initial undeformed
cable lengths, each value of nmax was set to 20 N, constraint (6)
considered an additional load case of 1.2 times the design load, and d2
was set to 0.037 m2, which corresponds to a RMS deviation from the
target of 5 mm per node.
The method of moving asymptotes [36] was chosen as the optimi
zation algorithm. The convergence history is shown in Fig. 10 for the
objective function (Eq. (2)), which shows a reduction from 1 to 0.54 in
only 49 iterations, as both terms in the objective function decreased. As
a result, the optimized gridshell fits the target better than the initial
design, while requiring only 130 cables instead of the 293 in the initial
design. The convergence history of the constraints is given in Fig. 11,
showing (dT1 /d1 )/d2 − 1 and the maximum value of nc /nc,max − 1 for each
iteration, such that negative values indicate a feasible design. The
constraint on the normal forces in the cables does not become active, as
the normal forces do not exceed 12.0 N during the optimization. On the
other hand, the constraint limiting the deviation from the target grid
under additional loading is heavily violated in the initial design. The
optimizer is capable of reducing this deviation until the design is
feasible. The deviation of the optimized gridshell from the target grid is Fig. 18. Rod connections are made using an elastic nylon band with duct-tape
shown in Fig. 13d for the design load, and Fig. 14d for the additional on the sides to restrict sliding of the knot.
load case.
To show the importance of taking the applied loads into account, an additional load case of 1.2 times the design load.
optimized design where the loads were neglected during optimization is In order to assess the impact of fixing the rod diameter in advance,
also considered. One could argue that this design is achievable with and to demonstrate the full optimization approach, we perform the
currently available design methods, by deriving the ideal (deformed) optimization a second time, now disregarding the constraint on the rod
cable lengths from the target grid, and using this information to itera diameter. The maximum allowable rod radius of 1.3 mm is used as initial
tively tension the cables in the physical gridshell until their lengths value. The weighting factors in the objective function are now chosen so
match the ideal lengths, before applying the final loads. Such a design is that, for the initial design, the first, second, and third term in the
compared to the design optimized according to the procedures described objective function equal 0.6, 0.2, and 0.2, respectively. This results in
previously, as well as the initial design and an unbraced design in the following values: w1 = 70 m− 2, w2 = 3.05 × 10− 4 N− 1, w3 = 154
Figs. 12–14 for the three following cases: without loads, under the m− 1. The weighting factors in this example are set to balance the effect
design load case, and under the additional load case. of all parts of the objective function to demonstrate the potential and
Comparison of the unloaded designs (Fig. 12) shows that the limitations of the approach. In a real application, the first part of the
unbraced gridshell does not fit the target shape very well, with de objective function would typically be most important.
viations exceeding 1 cm in a large area of the gridshell. This shows the The optimization problem is solved in 49 iterations. Figs. 15 and 16
need for a system to control the shape of the relaxed gridshell, such as show the convergence history of the objective function and the con
the bracing cables with adaptive lengths suggested in this paper. Com straints, respectively. The final value of the rod radius is 1.12 mm, which
parison of the loaded designs (Fig. 13) shows that some type of bracing is slightly more than the fixed value of 1 mm considered before. Drop
system is required anyway, as the unbraced gridshell completely snaps ping the constraint on the rod radius results in a larger design space,
through under the design load. Although the initial design and the which should give rise to a better optimum: the optimized design in
optimized design which does not consider the wet concrete loading fit cludes 86 cables, which is less than the 130 cables obtained before. The
the target grid very well without loads (Fig. 12), they perform much
misfit dT d between the final shape and the target shape of the shell is
worse than the optimized design once the design load is applied ∑c
44 % smaller than before, and the L1 norm ni=1 Ni of the cable forces is
(Fig. 13). Furthermore, they are much more sensitive to an increased
54 % smaller. These gains are achieved at the expense of the rod radius
loading, despite having more than twice as many cables: Fig. 14 shows
rg , which is 12 % larger than before. If this is not desirable, the
detrimental deformations for all designs but the optimized design for the
10
Fig. 17. The supports in the physical model consist of a small socket and a
tension cable running through the support plate.
optimization problem has to be solved again with different values for the
weighting terms in the objective function.
4.2. Construction
In this subsection, the construction of the scale model designed

above is discussed. The construction went through several stages. First,
the flat grid was assembled by tying the individual rods together with a
square lash using an elastic band to approximate cylindrical hinges.
Sliding of the knot was further restricted by applying a thin strip of duct-
tape next to the knot. The result (Fig. 18) is a connection that is rela
tively flexible for rotations in the plane of the grid, and stiff with respect
to other deformations. Subsequently, the grid was erected by putting
each rod end in a superficial socket, with a thin hole running through the
support plate to add a tension cable (Fig. 17). The resulting support is
able to generate reaction forces in almost any direction.
Fig. 20 shows the gridshell’s geometry at this stage, and a compar
ison between the physical and numerical model is made by visually
superimposing the numerical model on the physical model. Addition
ally, the geometry is captured by scanning several marked nodes on the
model. The scan is performed using a Vicon motion capture setup [35]
and 42 spherical markers distributed over the gridshell. An average
error of 0.15 mm and a variability lower than 0.025 mm have been re
ported for static experiments [37]. However, due to slightly inaccurate
placement of the markers, the maximum error of the results is expected Fig. 20. Geometry of the physical model without braces and without loads,
to be around 2 mm. The distance between the scanned nodes and the with the corresponding numerical model in blue. (For interpretation of the
modeled nodes is 3.1 mm on average, and at most 7.4 mm (Table 2). references to colour in this figure legend, the reader is referred to the web
Note that the maximum difference between the numerical model and the version of this article.)
physical model at this stage is equal to the maximal deviation from the
target of the optimized numerical model. In other words, for the current Once all cables were placed, their undeformed lengths were set to their
level of accuracy, it would be pointless to further minimize the predicted optimized values by tensioning the cable ties. Because we had no way of
deviation from the target. measuring the forces in the cables, the cables were tensioned based only
To further check the accuracy of the finite element model, the scale on their initial lengths. A comparison of the numerical model and the
model is subjected to two additional load cases before adding the physical model is shown in Fig. 24. Although the shape of the physical
bracing cables. The first load case consists of point loads of 300 g acting model matches the shape of the numerical model relatively well on first
on the same nodes as the design load case (Fig. 9). The scale model is sight, the accuracy is slightly worse than before the cables were added,
compared to the finite element model under these loading conditions in with deviations up to 14.0 mm, with an average of 4.1 mm.
Fig. 21. Another scan showed an average distance of 4.0 mm, and Before loading the structure with the final design load, an additional
maximum distance of 8.8 mm between the physical model and the nu test was performed first, applying the same load case as before, con
merical model (Table 2). The second load case consists of 10 point loads sisting of point loads of 300 g. The response of the scale model shows a
of 500 g acting in the middle of the gridshell (Fig. 19). The distance very similar mismatch as before the loads were applied, indicating that
between the numerical model and the physical model ranges up to 8.0 the relative displacements are quite accurately predicted by the nu
mm, with an average of 3.2 mm (Fig. 22, Table 2). merical model (Fig. 25). Deviations range up to 16.3 mm, with an
In the next stage, bracing cables were added to the scale model in the average of 4.3 mm (Table 2). As expected, adding bracing cables
optimized locations. The bracing consists of two double nylon wire resulted in a much stiffer gridshell. Finally, the design load (point loads
strings, and a tensioning mechanism composed of 2 cable ties (Fig. 23).
11
Table 2
Distance between numerical model and physical model for five scenarios
measured at the points shown in Fig. 27.
distance [mm]
pos. \case 1 (Fig. 20) 2 (Fig. 21) 3 (Fig. 22) 4 (Fig. 24) 5 (Fig. 25)
1 2.1 2.3 4.0 2.1 2.2

2 4.9 5.0 2.1 1.5 1.4
3 2.0 1.4 2.8 4.1 4.4
4 2.4 3.3 0.9 4.2 4.0
5 5.1 5.3 4.0 4.4 4.3
6 4.0 4.8 1.8 3.1 2.8
7 0.9 3.3 2.3 6.1 7.1
8 2.1 3.3 2.7 2.1 2.7
9 2.4 4.6 3.5 5.8 6.5
10 3.0 4.8 3.4 3.6 3.6
11 1.4 2.9 2.2 3.6 4.3
12 2.5 3.6 3.7 5.6 5.1
13 3.6 5.7 2.4 5.8 7.3
14 3.0 3.3 3.4 2.0 2.3
15 2.3 7.1 4.0 6.2 7.5
16 7.3 8.8 4.0 3.4 4.9
17 2.3 2.4 3.7 4.0 5.2
18 3.8 1.8 4.4 3.3 3.0
19 5.4 6.3 2.2 4.4 4.4
20 5.6 6.2 4.7 3.9 3.0
21 3.5 5.3 1.1 14.0 16.3
22 2.3 6.1 8.0 12.5 9.7
23 7.4 5.8 6.3 5.2 5.0
24 1.7 2.8 2.4 2.5 1.5
25 3.8 3.4 2.4 5.2 4.7
26 3.3 5.8 5.1 4.3 3.4
27 0.8 2.8 3.9 1.7 2.6
28 2.2 3.8 4.6 1.8 1.9
29 2.8 2.6 2.9 2.1 2.6
30 2.4 1.8 1.5 1.5 1.9
31 1.2 2.3 3.0 2.1 1.1
32 3.8 4.3 1.8 2.0 2.8
33 2.8 3.3 2.4 3.0 3.2
34 3.1 3.8 2.5 3.7 4.1
35 2.2 2.6 2.9 2.4 2.3
36 2.8 4.3 2.7 3.0 3.3
37 4.4 6.1 1.1 5.5 7.4
of 700 g) was applied to the scale model. The difference between the
numerical model and the physical model is a lot higher for this load case,
which is clearly visible in Fig. 26. At this stage, the model was near
collapse, and no additional scans were performed. It appears that the
proposed optimization approach pushes the design to its limits, and, as a
result, imperfections have a significant impact on the performance of the
structure.
These imperfections were very hard to control for a model of such Fig. 21. Geometry of the physical model without braces with 300 g on the
small scale. Several uncertainties influence the design, and it is hard to nodes indicated in Fig. 9, with the corresponding numerical model in blue.
pinpoint the exact cause of the observed mismatch. Imperfections with
respect to the consistency of the cable stiffness, irregularities in the grid
layout, uncertainties in the pre-tension in the cables, small imperfections
of the loads, dynamic effects, imperfections in support location and
stiffness, variations in stiffness of the rods, creep effects, and the small
inherent error in accuracy of the used beam model are all possible
causes. It is expected that most of these imperfections can be drastically
reduced if a gridshell is constructed on a larger scale, making use of a
more informed technique to tension the bracing cables. Therefore, the
next step should be to build a prototype on a larger scale, and find the
most important sources of uncertainty. The impact of the remaining
uncertainties can be reduced by means of robust design optimization
[38]; one could e.g. define a joint probability distribution for all un
certain parameters and minimize the variance of the misfit between the Fig. 19. Top view. Ten point loads (black dots) are applied in the middle of
shape of the loaded gridshell and the target shape. the gridshell.
Despite the observed deviations between the measured geometry and
the expected geometry, some general conclusions and recommendations
are made. Comparisons with smaller loads show a much better match,
indicating that the method is more effective for stiffer designs, for
12
Fig. 23. Braces are made from two double nylon wire strings and a tensioning
system made from cable ties.
Fig. 22. Geometry of the physical model without braces with 500 g on the
nodes indicated in Fig. 19, with the corresponding numerical model in blue.
example using stiffer cables, larger cross-sections or a denser grid.

Additionally, it is worth investigating the effectiveness of taking more
than one additional load case into account, simulating more asymmetric
loading for example, to increase the robustness of the design. This would
come at the cost of increased computation time, as a nonlinear structural
analysis is required for every iteration of the optimization process for
every considered load case. However, in practice, a sufficiently high
accuracy and robustness are crucial, and are often worth the additional
computational cost. Alternatively, the effectiveness of more thorough
approaches to obtain a robust design, developed for other optimization
problems could be explored [39]. However, some level of inaccuracy
will always remain. Therefore, in addition to improving the accuracy
and robustness of the design, the proposed optimization approach could
be employed to post-tension the cables in the gridshell after construc
tion, similar to the approach proposed by Liew et al. [10] for cable net
falsework. This strategy requires an accurate real-time measuring sys
tem to control the grid geometry. Further research should clarify which Fig. 24. Geometry of the physical model with braces without loads, with the
corresponding numerical model in blue.
13
Fig. 25. Geometry of the physical model with braces and with 300 g on the Fig. 26. Geometry of the physical model with braces and with 700 g on the
nodes indicated in Fig. 9, with the corresponding numerical model in blue. nodes indicated in Fig. 9, with the corresponding numerical model in blue. (For
interpretation of the references to colour in this figure legend, the reader is
adaptations are needed to the existing optimization setup to apply it in referred to the web version of this article.)
such an active control approach.
5. Conclusions
This paper presents a framework for the design of gridshells as

falsework for concrete shells. As a key part of the design framework, a
novel optimization approach was proposed to fit a gridshell to a given
target shape, while taking external loading into account to make sure the
gridshell fits the target shape when the concrete is applied. Practical
constraints and penalization terms make sure the final design is feasible,
and material use and construction time are minimized. A bracing system
with adjustable lengths was suggested to allow control of the gridshell’s
shape after its relaxation. A scale model was constructed to test the
design framework in practice and formulate recommendations for
further research. Fig. 27. Top view. 37 points are measured on the physical model to compare to
Numerical results showed promising improvements over existing the corresponding points of the numerical model.
design techniques on four areas; (1) deviations from the target shape
14
were reduced, (2) the required number of braces was reduced, (3) the adaptively control the cable lengths could be explored.
forces in the cables were effectively constrained, and (4) the robustness
was improved by constraining deformations under additional loading.
Furthermore, the cross-section of the grid rods can also be minimized. Declaration of Competing Interest
However, for practical reasons, this was not applied in the design of the
scale model. In order for the optimization to be effective in practice, it is The authors declare that they have no known competing financial
crucial that the numerical model accurately simulates the behavior of interests or personal relationships that could have appeared to influence
the physical model. Although the construction of a scale model showed a the work reported in this paper.
relatively good match for smaller loads, larger deviations were observed
for the design load. Therefore, future research should focus on finding Acknowledgements
and eliminating the origin of the experienced inaccuracy, for example,
by constructing a gridshell on a larger scale and accurately monitoring We would like to thank K. Langlois and T. Bacek for their assistance
the gridshell’s geometry and forces in the cables through all construction with scanning the scale model. This work was supported by the Research
stages. Additionally, the potential of an active control system to Foundation - Flanders (FWO; Grant No. G0C2315N).
Appendix A. Sensitivities
In order to allow for the use of a gradient-based optimization algorithm, a sensitivity analysis of the optimization problem defined in Section 2 is
performed. Below, the objective function (Eq. (2)) and the left-hand side of constraints (5) and (6) are differentiated consecutively with respect to the
undeformed cable lengths l0c and the radius of the cross-section of the grid rods rb .
A.1. Differentiation with respect to the undeformed cable lengths
In this subsection, the derivatives of the objective function (Eq. (2)) and the left-hand side of constraints (5) and (6) with respect to the undeformed
cable lengths l0c are first expressed in terms of the following three total derivatives: (1) the derivative dd/dl0c of the distance vector under the design
load with respect to the undeformed cable lengths, (2) the derivative dd1 /dl0c of the distance vector under the additional load case with respect to the
undeformed cable lengths and (3) the derivative dnc /dl0c of the normal forces in the cables with respect to the undeformed cable lengths.
The following expression shows how the objective function f is a function of the undeformed cable lengths l0c :
( 0) ( ( ( )) ( ( )))
̂f l = f d x l0 , nc l0 , x l0
c c c c (A.1)
where d is the distance vector, whose elements are defined in Eq. (7), x is a vector collecting the positional information of the nodes (i.e. coordinates
and rotation vector components corresponding respectively to each translational degree of freedom and each rotational degree of freedom of the finite
element model), and nc is the vector collecting the normal forces in the cables. Consequently, differentiation of the objective function with respect to
the undeformed cable lengths gives:
df ∂f dd ∂f dnc
= + (A.2)
dl0c ∂d dl0c ∂nc dl0c
where:
∂f
= 2w1 dT (A.3)
∂d
and ∂f/nc is elaborated for each cable i individually:

∂f
= w2 (A.4)
∂nc,i
Differentiation of the left-hand side of constraint (5) also involves the derivation of dnc /dl0c . Differentiation of the left-hand side of constraint (6) gives:
d(dT1 d1 ) dd1
= 2dT1 0 (A.5)
dl0c dlc
A.1.1. Differentiation of the distance vector

The distance vector d depends on the undeformed cable lengths as follows:
( 0) ( ( ))
̂ l = d x l0
d c c (A.6)
Consequently, the derivative of the distance vector d is given by:
15
dd ∂d dx
= (A.7)
dl0c ∂x dl0c
where ∂d/∂x is elaborated for each node individually. For node i, differentiation of Eq. (7) gives:
∂di ∂di dxi ∂di dyi ∂di dzi
= + + (A.8)
∂x ∂xi dx ∂yi dx ∂zi dx
The derivative of the distance di with respect to the x-coordinate of node i is given by:
∂di xi − xti
= (A.9)
∂xi di
where xti is the x-coordinate of node i in the target grid. In Eq. (A.8), dxi /dx can be written as:
dxi
= Lxi (A.10)
dx
where Lxi is a selection matrix selecting the x-coordinate xi of node i from the position vector x:
x i = Lx i x (A.11)
Similar expressions can be derived for the derivatives in terms of y-, and z-coordinates in Eq. (A.8).
The derivative of the position vector x with respect to the undeformed cable lengths l0c in Eq. (A.7) is found by differentiation of equilibrium Eq. (8):
dr dp df
= − =0 (A.12)
dl0c dl0c dl0c
where dp/dl0 is zero, as the undeformed cable lengths l0c do not influence the magnitude of the loads, nor the degrees of freedom on which the loads act.
On the other hand, the dependency of the internal forces f on the undeformed cable lengths l0c is given by:
( )
̂f (l0 ) = f l0 , x(l0 )
c c c (A.13)
Differentiation of the internal forces f results in:

df ∂f ∂f dx
= + (A.14)
dl0c ∂l0c ∂x dl0c
Next, combining Eqs. (A.12) and (A.14) gives:

∂f dx ∂f
= − 0 (A.15)
∂x dl0c ∂lc
where ∂f/∂x is the tangential stiffness matrix. Eq. (A.15) is rewritten as:
dx ∂f
K = − 0 (A.16)
dl0c ∂lc
which can be solved for dx/dl0c when ∂f/∂l0c is known. Differentiation of the internal forces f with respect to the undeformed cable lengths l0c gives:
∂f ∑ kc
∂f dl0i
0
= 0 0
(A.17)
∂lc i=1 ∂li dlc
where kc is the number of cables and dl0i /dl0c is a selection matrix Sci selecting element i from all cable elements:
l0i = Sci l0c (A.18)
The derivative ∂f/∂l0i in Eq. (A.17) is further elaborated:

∂f ∂f ∂f ei
= (A.19)
∂l0i ∂f ei ∂l0i
where f ei collects the internal forces of element i and:
16
∂f ∂f ∂f F ( )T
= = LF LFi (A.20)
∂f ei ∂f F ∂f ei
where f F is the full internal force vector, collecting the internal forces corresponding to every degree of freedom, including fixed degrees of freedom,
and LF and LFi are selection matrices: LF selects the unconstrained degrees of freedom from all degrees of freedom and LFi selects the degrees of
freedom of element i from all degrees of freedom:
f = LF f F (A.21)
f ei = LFi f F (A.22)
The derivative ∂f ei /∂l0i in Eq. (A.19) is found by differentiation of Eq. (19):

e
∂f ei ∂Tei e ∂f
= 0 f i + Tei 0i (A.23)
∂l0i ∂li ∂li
where ∂Tei /∂l0i is zero. Differentiation of Eq. (15) gives:

e e e
∂f i ∂Ki e e ∂u
0
= 0 ui + Ki 0i (A.24)
∂li ∂li ∂li
e
where the derivatives of Ki and uei are found by differentiation of Eqs. (18), and (13) and (17):
e e
∂Ki K
= − 0i (A.25)
∂l0i li
∂uei
= [− 1 0 0 0 0 0 0]T (A.26)
∂l0i
e
Inserting Eqs. (A.25) and (A.26) into Eq. (A.24) gives ∂f i /∂l0i ; which can be used in Eq. (A.23) to find ∂f ei /∂l0i . Inserting this derivative into Eq. (A.19),
together with ∂f/∂f ei
from Eq. (A.20) gives ∂f/∂l0i ,
which is required to solve Eq. (A.16), taking Eq. (A.17) and (A.18) into account. Finally, dd/dl0c is
0
calculated by inserting dx/dlc from Eq. (A.16) and ∂d/∂x from Eqs. (A.8)–(A.10) into Eq. (A.7). Differentiation of distance vector d1 follows along the
same lines.
A.1.2. Differentiation of the normal forces in the cables

Next, dnc /dl0c is derived, which is required to determine the second term in Eq. (A.2), and for the differentiation of the left-hand side of constraint
(5). The differentiation is performed for individual cables i and j. The normal force nc,i in cable i depends on the undeformed cable length l0j of cable j as
follows:
( ) ( e( ( ( )) e ))
n c,i l0j = nc,i f i uei l0j , xei x(l0j ) , Ki (l0j )
̂ (A.27)
Differentiation of the normal force nc,i in cable i with respect to the undeformed cable length of cable j gives:
e
dnc,i ∂nc,i df i
= e 0 (A.28)
dl0j ∂f i dlj
e
where, because nc,i = Ni is the first element of the internal force vector f i :
∂nc,i
e = [1 0 0 0 0 0 0] (A.29)
∂f i
e
Next df i /dl0j is further elaborated according to Eq. (15):
e e e
df i dKi e e du
0
= 0 ui + Ki 0i (A.30)
dlj dlj dlj
e e e
where, if j = i, the total derivative dKi /dl0j is equal to the partial derivative ∂Ki /∂l0i given in Eq. (A.25), and if j ∕
= i, dKi /dl0j is zero. Next, the de
rivative duei /dl0j is further elaborated:
duei ∂uei ∂uei dxei
= 0+ e 0 (A.31)
dl0j ∂lj ∂xi dlj
17
where, if j = i, ∂uei /∂l0j is given by Eq. (A.26), and if j ∕

= i, ∂uei /∂l0j is zero. Furthermore, ∂uei /∂xei is the transformation matrix Tei defined in Eq. (20), and
dxei /dl0j is further elaborated:
dxei ∂xe dxF
= i 0 (A.32)
dl0j ∂xF dlj
where xF is the full position vector collecting all positional information for all degrees of freedom, including fixed degrees of freedom, and:
∂xei
= LFi (A.33)
∂xF
where LFi is the selection matrix introduced in Eq. (A.22). The next derivative in Eq. (A.32) is elaborated further:
dxF ∂xF dx
= (A.34)
dl0j ∂x dl0j
where dx/dl0j is obtained from Eq. (A.16), and:

∂xF ( F )T
= L (A.35)
∂x
where LF the selection matrix introduced in Eq. (A.21).

With ∂xF /∂x and dx/dl0c from Eqs. (A.35) and (A.16) respectively, dxF /dl0j can be calculated using Eq. (A.34). This derivative is then used in Eq.
(A.32), together with ∂xei /∂xF from Eq. (A.33) to obtain dxei /dl0j . Next, duei /dl0i is calculated from Eq. (A.31) with ∂ue /∂l0j and ∂uei /∂xei from Eqs. (A.26)
e e
and (20), respectively. The derivative duei /dl0j is then used in Eq. (A.30) to find df i /dl0j with dKi /dl0j from Eq. (A.25). Finally, dnc,i /dl0j is calculated using
e e
Eq. (A.28) with the previously calculated df i /dl0j and ∂nc,i /∂f i from Eq. (A.29).
A.2. Differentiation with respect to the radius of the cross-section of the grid rods
Similar to the previous subsection, the derivatives of the objective function (Eq. (2)) and the left-hand side of constraints (5) and (6) with respect to
the radius of the cross-section of the grid rods rg are first expressed in terms of the following three total derivatives: (1) dd/drg , (2) dd1 /drg and (3)
dnc /drg .The calculation of these total derivatives themselves is tackled next.
The objective function f is a function of the radius rg of the cross-section of the grid rods as follows:
( ) ( ( ( )) ( ( )) )
̂f rg = f d x rg , nc x rg , rg (A.36)
Differentiation of the objective function f with respect to the radius rg of the cross-section of the grid rods gives:
df ∂f dd ∂f dnc ∂f
= + + (A.37)
drg ∂d drg ∂nc drg ∂rg
where ∂f/∂d and ∂f/∂nc are defined in Eqs. (A.3) and (A.4) respectively, and:
∂f
= w3 (A.38)
∂rg
Differentiation of the left-hand side of constraint (5) also involves the derivation of dnc /drg . Differentiation of the left hand side of constraint (6) gives:
d(dT1 d1 ) dd1
= dT1 (A.39)
drg drg
A.2.1. Differentiation of the distance vector

The distance vector d depends on the cross-section of the radius of the grid rods as follows:
( ) ( ( ))
̂
d rg = d x rg (A.40)
Consequently, elaborating dd/drg gives:

dd ∂d dx
= (A.41)
drg ∂x drg
18
where ∂d/∂x is defined for any node i in Eq. (A.8), and dx/drg is derived by differentiation of the equilibrium Eq. (8):
dr dp df
= − =0 (A.42)
drg drg drg
where dp/drg is zero, as the radius rg of the cross-section of the grid rods does not influence the magnitude of the loads, nor the degrees of freedom on
which the loads act. On the other hand, the dependency of the internal forces f on the radius rg of the cross-section of the grid rods is given by:
( ) ( ( ))
̂f rg = f rg , x rg (A.43)
Differentiation of the internal forces f results in:

df ∂f ∂f dx
= + (A.44)
drg ∂rg ∂x drg
Next, combining Eqs. (A.42) and (A.44) gives:

∂f dx ∂f
= − (A.45)
∂x drg ∂rg
where ∂f/∂x is the tangential stiffness matrix. Eq. (A.45) is rewritten as:
dx ∂f
K = − (A.46)
drg ∂rg
The radius rg of the cross-section of the grid rods indirectly influences the internal forces f through the section properties A,J,Iy , and Iz . Consequently,
elaborating ∂f/∂rg gives:
k ( )
∂f ∑ g
∂f dAi ∂f dJi ∂f dIy,i ∂f dIz,i
= + + + (A.47)
∂rg i=1
∂Ai drg ∂Ji drg ∂Iy,i drg ∂Iz,i drg
where kg is the number of elements used to model the grid rods and Ai , Ji , Iy,i , and Iz,i are the section properties of grid rod element i. The derivatives
dAi /drg , dJi /drg , dIy,i /drg and dIz,i /drg are derived assuming circular cross-sections:
dAi
= 2πrg (A.48)
drg
dJi
= 2π r3g (A.49)
drg
dIy,i dIz,i
= = πr3g (A.50)
drg drg
Next, ∂f/∂Ai in Eq. (A.47) is derived. Differentiation of the internal forces f with respect to the area Ai of the cross-section of element i gives:
df ∂f df e
= e i (A.51)
dAi ∂f i dAi
where ∂f/∂f ei is defined in Eq. (A.20). Differentiation of Eq. (19) gives:

e
df ei dTe e df
= i f + Tei i (A.52)
dAi dAi i dAi
where dTei /dAi is zero. Differentiation of Eq. (15) gives:

e e e
df i dKi e e du
= u + Ki i (A.53)
dAi dAi dAi
e
where duei /dAi is zero, and the derivative of K is found by differentiation of Eq. (18):
⎡ ⎤
e
E 0 … 0
dK 1⎢ 0 0 ⎥
= 0⎢ ⎥ (A.54)
dA l ⎣⋮ ⋱ ⎦
0 0
19
Differentiation of the internal forces f with respect to other cross-section properties J, Iy , Iz can be found following the same approach, with:
⎡ ⎤
0 … 0
⎢0 G 0 0 − G 0 0⎥
⎢ ⎥
e ⎢0 … 0⎥
dK 1⎢ ⎥
= 0⎢ 0 … 0⎥ (A.55)
dJ l ⎢
⎢0 − G 0
⎥
⎢ 0 G 0 0⎥ ⎥
⎣0 … 0⎦
0 … 0
⎡ ⎤
0 … 0
⎢0 … 0⎥
⎢ ⎥
e ⎢0 0 4E 0 0 2E 0⎥
dK 1⎢ ⎥
= 0⎢ 0 … 0⎥ (A.56)
dIy l ⎢
⎢0
⎥
⎢ … 0⎥⎥
⎣0 0 2E 0 0 4E 0⎦
0 … 0
⎡ ⎤
0 … 0
⎢0 … 0 ⎥
⎢ ⎥
e ⎢0 … 0 ⎥
dK 1⎢ ⎥
= 0⎢ 0 0 0 4E 0 0 2E ⎥ (A.57)
dIz l ⎢
⎢0
⎥
⎢ … 0 ⎥
⎥
⎣0 … 0 ⎦
0 0 0 2E 0 0 4E
e
Eq. (A.54) can be combined with Eq. (A.53) to find df i /dAi . Next, df/dAi can be calculated using Eqs. (A.51) and (A.53). Similarly, using Eqs. (A.55),
(A.56), and (A.57) instead of (A.54), df/dJi , df/dIy,i , and df/dIz,i can also be calculated. These derivatives can be used together with the derivatives
defined in Eqs. (A.48), (A.49), and (A.50) in Eq. (A.44) to find df/drg , which is required to solve Eq. (A.46) for dx/drg . Finally, inserting dx/drg into Eq.
(A.41), together with dd/dx from Eq. (A.8) gives dd/drg . Differentiation of distance vector d1 follows along the same lines.
A.2.2. Differentiation of the normal forces in the cables

The normal force in a cable i depends on the radius of the cross-section of the grid rods as follows:
( ) ( e( ( ( ))))
n c,i rg = nc,i f i uei xei x(rg )
̂ (A.58)
Applying the chain rule to elaborate dnc,i /drg gives:

e
dnc,i ∂nc,i df i
= e (A.59)
drg ∂f i drg
e
where ∂nc,i /∂f i is given in Eq. (A.29), and:
e e
df i e du
= Ki i (A.60)
drg drg
where:
duei ∂uei dxei
= (A.61)
drg ∂xei drg
where ∂uei /∂xei is the transformation matrix Tei defined in Eq. (20). Further elaboration of Eq. (A.61) follows the same steps as previously developed for
the undeformed cable lengths l0c . The corresponding expressions are obtained by replacing l0c by rg in Eqs. (A.32), (A.33), (A.34), (A.35), and using Eq.
(A.46) to determine dx/drg .
A.3. Finite difference verification
In this subsection, the analytical expressions for the design sensitivities determined above are verified numerically. A simple example is chosen to
validate thesensitivities. A bending-active arch is erected from a 4 m long, initially straight flexible rod, with its ends pinned 8/π m apart (Fig. A.28).
The rod has a Young’s modulus E = 33 GPa, and is discretized into 8 elements. The diameter of the cross-section of the rod is initially set to 10 mm.
After the rod is erected, it is braced with cables that have a diameter of 1 mm and Young’s modulus E = 210 GPa. The location of the braces is shown in
Fig. A.28. The undeformed lengths of the braces are initially set to 0.995 times the distances between the nodes they connect, measured in the rod’s
naturally bent condition, i.e. the shape which the rod adopts for the given boundary conditions, without external loads. Finally, the arch is loaded with
a point load P = 100 N in the middle.
The target shape for the rod is a semi-circle. A target node is defined on the semi-circle for each node of the modeled rod. The undeformed lengths of
the bracing cables and the radius of the cross-section of the rod are regarded as the design variables. The weighting factors in the objective function
20
Fig. A.28. A braced (thin lines) slender rod (thick line) forms the initial design for the validation of the sensitivities. Its target shape is discretized using equidistant
nodes (blue). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)
∑c
(Eq. (2)) are chosen so that, for the initial design, the first term (w1 dT d) has a value of 0.7, the second term (w2 ni=1 Ni ) has a value of 0.1, and the third
term (w3 rg ) has a value of 0.2, resulting in an objective function of 1. The corresponding values for the weighting factors are w1 = 33.3 m− 2, w2 =
1.92 × 10− 5 N− 1, w3 = 40.0 m− 1, respectively.
For the initial design, the sensitivities of the objective function f with respect to the undeformed cable lengths l0c and the radius rg of the cross-
section of the rod are calculated analytically using the expressions derived in Sections A.1 and A.2. The results are compared to the same sensitiv
ities calculated numerically by means of the forward finite difference method2. In the forward finite difference method, an approximation of the
derivative df/dx of a function f(x) with respect to x is determined as follows:
df f (x + ∊) − f (x)
≈ (A.62)
dx ∊
where the increment ∊ is a small positive number. For the test case considered, frepresents the objective function, and x stands for the undeformed
cable lengths and the radius of the cross-section of the rod. The finite difference calculation is performed for increments ∊ ranging from 10− 1 m to
10− 10 . Table A.3 shows that the analytically calculatedsensitivities deviate slightly from the converged values of the numerically calcu
latedsensitivities, with an average deviation of 0.04 % between the analytically calculatedsensitivities and thesensitivities calculated numerically with
∊ = 10− 10 . Closer investigation has shown that this small inaccuracy is primarily caused by the fact that the analytical expression for the tangential
stiffness matrix given by Crisfield is not perfectly consistent, as explained in Section 2.6 of the original paper [29]. In an attempt to eliminate this error,
we recalculated the sensitivities using the analytical expressions from the previous subsections, but now using a numerically calculated stiffness
matrix: the stiffness matrix, which represents the derivative df/dx of the internal forces f with respect to the nodal coordinates x, is now calculated by
means of the forward finite difference method, using ∊ = 10− 10 m. The results are given in Table A.3. They are much closer to the numerically obtained
sensitivities: the average deviation from the sensitivities calculated numerically with ∊ = 10− 10 is only 6 × 10− 5 %. However, using the numerically
calculated stiffness matrix results in a small asymmetry in thesensitivities. Moreover, we observed that the inaccuracy caused by analytically
calculating the tangential stiffness matrix has limited influence on the convergence behavior of the optimization problem. As the numerical calculation
is computationally more demanding, we prefer to use the analytical expression for the stiffness matrix.
The inconsistency of the stiffness matrix also influences the convergence speed of the nonlinear analyses performed during each optimization
Table A.3
Comparison of the derivatives of the objective function f (Eq. (2)) of the braced arch shown in Fig. A.28. Analytical derivatives are compared to the forward finite
difference derivatives for a decreasing increment ∊.
K ∂f/∂l0c [m− 1] ∂f/∂rg [m− 1]
analytical analyt. 41.6801 0.5652 − 28.4215 − 39.1266 − 28.4215 0.5652 41.6801 56.7718
numer. 41.6806 0.5659 − 28.4222 − 39.1284 − 28.4220 0.5663 41.6811 56.7721
∊ [m] ∂f/∂l0c [m− 1] ∂f/∂rg [m− 1]
finite differences 10− 1 15.2969 13.8808 0.5378 − 6.9063 0.5378 13.8808 15.2969 40.5957
10 − 2 152.9685 42.7250 − 15.3648 − 33.2679 − 15.3648 42.7250 152.9685 48.5799
10− 3 59.4205 4.0898 − 27.1031 − 38.4377 − 27.1031 4.0898 59.4205 55.5385
10− 4 43.0310 0.9140 − 28.2895 − 39.0576 − 28.2895 0.9140 43.0310 56.6165
10− 5 41.8132 0.6008 − 28.4089 − 39.1214 − 28.4089 0.6008 41.8132 56.7560
10− 6 41.6941 0.5696 − 28.4208 − 39.1278 − 28.4208 0.5696 41.6941 56.7704
10− 7 41.6822 0.5664 − 28.4220 − 39.1284 − 28.4220 0.5664 41.6822 56.7719
10− 8 41.6810 0.5661 − 28.4221 − 39.1285 − 28.4221 0.5661 41.6810 56.7720
10− 9 41.6809 0.5661 − 28.4221 − 39.1285 − 28.4221 0.5661 41.6809 56.7720
10− 10 41.6809 0.5661 − 28.4222 − 39.1284 − 28.4221 0.5661 41.6809 56.7720
2
Although the central finite difference method is more accurate, the authors have chosen the forward finite difference method because it can be determined with
only a single additional function evaluation per design variable (as f (x) is already computed for the determination of the analytical derivative), instead of two for the
central finite difference method. The authors claim that a higher accuracy of the numerical derivatives would not change the conclusion of this analysis.
21
iteration. However, the inconsistency is small – Crisfield reports a convergence which is ‘close to quadratic’ using the Newton–Raphson method [29] –
and quadratic convergence is lost anyway in favor of numerical stability by using implicit dynamic relaxation [27]. Therefore, the authors consider
this influence to be negligible.
References 1971 vis Mai 1973 am Institut fur Leichte Flachentragwerke. Stuttgart: Institute for
Lightweight Structures; 1974.
[21] Bouhaya L, Baverel O, Caron J-F. Optimization of gridshell bar orientation using a
[1] Ramm E, Bletzinger K-U, Reitinger R. Shape optimization of shell structures. Bull
simplified genetic approach. Struct Multidiscip Optim 2014;50(5):839–48. https://
Int Assoc Shell Spatial Struct 1993;34(2):103–22. https://doi.org/10.1080/
doi.org/10.1007/s00158-014-1088-9.
12506559.1993.10511083.
[22] Pone S, Mirra G, Pignatelli E, Lancia D, Colabella S. Specialised algorithms for
[2] Tomás A, Martí P. Shape and size optimisation of concrete shells. Eng Struct 2010;
different project stages in a postformed timber gridshell design, in. In: Proceedings
32(6):1650–8. https://doi.org/10.1016/j.engstruct.2010.02.013.
of the international conference on structures and architecture; 2016. https://doi.
[3] Billig K. Concrete shell roofs with flexible moulds. J ICE 1946;25(3):228–31.
org/10.1201/b20891-33.
[4] Kromoser B, Huber P. Pneumatic formwork systems in structural engineering. Adv
[23] Hernández EL, Baverel O, Gengnagel C. On the design and construction of elastic
Mater Sci Eng 2016;2016. https://doi.org/10.1155/2016/4724036.
gridshells with irregular meshes. Int J Space Struct 2013;28(3–4):161–74. https://
[5] Sobek W. Concrete shells constructed on pneumatic formwork. In: Heki K, editor.
doi.org/10.1260/0266-3511.28.3-4.161.
Proceedings lASS Symposium, Vol. 1; 1986. p. 337–44.
[24] Bouhaya L, Baverel O, Caron J-F. Mapping two-way continuous elastic grid on an
[6] Sobek W. Auf Pneumatisch Gestützten Schalungen Hergestellte Betonschalen,
imposed surface: Application to grid shells. In: Domingo A, Lázaro C, editors.
Ursula Sobek; 1987.
Proceedings of the International Association for Shell and Spatial Structures (IASS)
[7] van Hennik PC, Houtman R. Textile Composites and Inflatable Structures II., Vol. 8
Symposium 2009, Valencia; 2009. p. 989–97.
of Computational Methods in Applied Sciences, Springer, Ch. Pneumatic Formwork
[25] Kuijvenhoven M, Hoogenboom PCJ. Particle-spring method for form finding grid
for Irregular Curved Thin Shells; 2008. p. 99–116. doi:10.1007/978-1-4020-6856-
shell structures consisting of flexible members. J Int Assoc Shell Spatial Struct
0_7.
2012;53(1):31–8.
[8] Van Mele T, Block P. A novel form finding method for fabric formwork for concrete
[26] Li J-M, Knippers J. Designing regular and irregular elastic gridshells by six DOF
shells. J Int Assoc Shell Spatial Struct 2011;52(4):217–24.
dynamic relaxation, in. In: Proceedings of design modeling symposium, Berlin;
[9] Veenendaal D, Block P. Design process for a prototype concrete shells using a
2013.
hybrid cable-net and fabric formwork. Eng Struct 2014;75:39–50. https://doi.org/
[27] Rombouts J, Lombaert G, De Laet L, Schevenels M. A fast and accurate dynamic
10.1016/j.engstruct.2014.05.036.
relaxation approach for form-finding and analysis of bending-active structures. Int
[10] Liew A, Stürz YR, Guillaume S, Van Mele T, Smith RS, Block P. Active control of a
J Space Struct 2019;34(1–2):40–53. https://doi.org/10.1177/
rod-net formwork system prototype. Autom Constr 2018;96:128–40. https://doi.
0956059919864279.
org/10.1016/j.autcon.2018.09.002.
[28] Rombouts J, Lombaert G, De Laet L, Schevenels M. On the equivalence of dynamic
[11] Block P, Schlueter A, Veenendaal D, Bakker J, Begle M, Hischier I, Hofer J,
relaxation and the Newton-Raphson method. Int J Numer Meth Eng 2018;113(9):
Jayathissa P, Maxwell I, Mendez Echenagucia T, Nagy Z, Pigram D, Svetozarevic B,
1531–9. https://doi.org/10.1002/nme.5707.
Torsing R, Verbeek J, Willmann A, Lydon G. Nest hilo: Investigating lightweight
[29] Crisfield MA. A consistent co-rotational formulation for non-linear, three-
construction and adaptive energy systems. J Build Eng 2017;12:332–41. https://
dimensional, beam-elements. Comput Methods Appl Mech Eng 1990;81:131–50.
doi.org/10.1016/j.jobe.2017.06.013.
https://doi.org/10.1016/0045-7825(90)90106-V.
[12] Veenendaal D, Bakker J, Block P. Structural design of the flexibly formed, mesh-
[30] Belytschko T, Schwer L, Klein MJ. Large displacement, transient analysis of space
reinforced concrete sandwich shell roof of NEST Hilo. J Int Assoc Shell Spatial
frames. Int J Numer Meth Eng 1977;11(1):65–84. https://doi.org/10.1002/
Struct 2017;58(1):23–38. https://doi.org/10.20898/j.iass.2017.191.847.
nme.1620110108.
[13] Popescu M, Reiter L, Liew A, Van Mele T, Flatt R, Block P. Building in concrete with
[31] Crisfield MA. Non-linear finite element analysis of solids and structures, Vol. 2.
a knitted stay-in-place formwork: Prototype of a concrete shell bridge. Structures
Wiley; 1997.
2018;14:322–32. https://doi.org/10.1016/j.istruc.2018.03.001.
[32] Felippa CA, Haugen B. A unified formulation of small-strain corotational finite
[14] Tang G, Pedreschi R. Deployable gridshells as formwork for concrete shells. In:
elements: I. theory. Comput Methods Appl Mech Eng 2005;194:2285–335. https://
Proceedings of the international society of flexible formwork (ISOFF) Symposium
doi.org/10.1016/j.cma.2004.07.035.
2015; 2015.
[33] Rippmann M, Lachauer L, Block P. RhinoVAULT - interactive vault design. Int J
[15] Coar L, Cox M, Adriaenssens S, De Laet L. The design and construction of fabric
Space Struct 2012;27(4):219–30. https://doi.org/10.1260/0266-3511.27.4.219.
formed ice shells with bending active frames utilizing principle stress patterns. In:
[34] Block P, Ochsendorf J. Thrust network analysis: A new methodology for three-
Bögle A, Grohmann M, editors. Proceedings of the IASS Annual Symposium 2017,
dimensional equilibrium. J Int Assoc Shell Spatial Struct 2007;48(3):167–73.
Hamburg; 2017.
[35] Vicon, Intelligence in motion, www.vicon.com (Accessed March 2019).
[16] Cuvilliers P, Douthe C, Du Peloux L, Le Roy R. Hybrid structural skin: prototype of
[36] Svanberg K. The method of moving asymptotes—a new method for structural
elastic gridshell braced with a concrete envelop. In: Bögle A, Grohmann M, editors.
optimization. Int J Numer Meth Eng 1987;24(2):359–73. https://doi.org/10.1002/
Proceedings of the IASS Annual Symposium 2017, Hamburg; 2017.
nme.1620240207.
[17] Hernández EL. Design and optimisation of elastic gridshells. Ph.D. thesis.
[37] Merriaux P, Dupuis Y, Boutteau R, Vasseur P, Savatier X. A study of vicon system
Universität der Künste Berlin; 2015.
positioning performance, Sensors 2017;17(7). doi:10.3390/s17071591.
[18] Pignatelli E, Mirra G, Pone S. InFormer: designing forming actions in post-formed
[38] Beyer H, Sendoff B. Robust optimization - A comprehensive survey 2007;196:
gridshells by means of MOGAs. In: Bögle A, Grohmann M, editors. Proceedings of
3190–218.
the IASS Annual Symposium, Hamburg; 2017.
[39] Schuëller GI, Jensen HA. Computational methods in optimization considering
[19] Rombouts J, Lombaert G, De Laet L, Schevenels M. A novel shape optimization
uncertainties – an overview. Comput Methods Appl Mech Eng 2008;198(1):2–13.
approach for strained gridshells: design and construction of a simply supported
https://doi.org/10.1016/j.cma.2008.05.004.
gridshell. Eng Struct 2019;192:166–80.
[20] Otto F, Schauer E, Hennicke J, Hasegawa T. IL 10 Gitterschalen/ Grid shells:
Bericht über das japanisch-deutsche Forschungsprojekt STI, durchgeführt von Mai
22

1 s2.0 S0141029621005022 Main

Uploaded by

Copyright:

Available Formats

1 s2.0 S0141029621005022 Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S0141029621005022 Main

Uploaded by

Copyright:

Available Formats

Engineering Structures 240 (2021) 112352

Contents lists available at ScienceDirect

Designing bending-active gridshells as falsework for concrete shells through

1. Introduction This makes on-site construction more challenging. If concrete is used as

wet concrete loading. Each step in the flowchart is explained in the

where I is a three-by-three identity matrix, and S is a skew-symmetric e

ue = [δ θx,a θy,a θz,a θx,b θy,b θz,b ]T (17)

Fig. 4. A funicular shell is designed using RhinoVault, where the two-

where xe = [xa ya za αa βa γa xb yb zb αb βb γ b ]T is a vector collecting 4.1. Design

length scale 1/1 20/1

gridshell length 0.90 m 18.0 m

shown in Fig. 9. The scans measured the location of 42 nodes distributed

4.1.1. Generation of target grid

4.1.2. Generation of initial design

In this subsection, the construction of the scale model designed

1 2.1 2.3 4.0 2.1 2.2

example using stiffer cables, larger cross-sections or a denser grid.

This paper presents a framework for the design of gridshells as

A.1. Differentiation with respect to the undeformed cable lengths

and ∂f/nc is elaborated for each cable i individually:

A.1.1. Differentiation of the distance vector

Consequently, the derivative of the distance vector d is given by:

Differentiation of the internal forces f results in:

Next, combining Eqs. (A.12) and (A.14) gives:

The derivative ∂f/∂l0i in Eq. (A.17) is further elaborated:

where f ei collects the internal forces of element i and:

The derivative ∂f ei /∂l0i in Eq. (A.19) is found by differentiation of Eq. (19):

where ∂Tei /∂l0i is zero. Differentiation of Eq. (15) gives:

A.1.2. Differentiation of the normal forces in the cables

where, if j = i, ∂uei /∂l0j is given by Eq. (A.26), and if j ∕

where dx/dl0j is obtained from Eq. (A.16), and:

where LF the selection matrix introduced in Eq. (A.21).

A.2.1. Differentiation of the distance vector

Consequently, elaborating dd/drg gives:

Differentiation of the internal forces f results in:

Next, combining Eqs. (A.42) and (A.44) gives:

where ∂f/∂f ei is defined in Eq. (A.20). Differentiation of Eq. (19) gives:

where dTei /dAi is zero. Differentiation of Eq. (15) gives:

A.2.2. Differentiation of the normal forces in the cables

Applying the chain rule to elaborate dnc,i /drg gives:

A.3. Finite difference verification

You might also like