Conference PaperPDF Available

OptiGAN: Topological Optimization in Design Form-Finding With Conditional GANs


Abstract and Figures

With the rapid development of computers and technology in the 20th century, the topological optimisation (TO) method has spread worldwide in various fields. This novel structural optimisation approach has been applied in many disciplines, including architectural form-finding. Especially Bi-directional Evolutionary Structural Optimisation (BESO), which was proposed in the 1990s, is widely used by thousands of engineers and architects worldwide to design innovative and iconic buildings. To integrate topological optimisation with artificial intelligence (AI) algorithms and to leverage its power to improve the diversity and efficiency of the BESO topological optimisation method, this research explores a non-iterative approach to accelerate the topology optimisation process of structures in architectural form-finding via conditional generative adversarial networks (GANs), which is named as OptiGAN. Trained with topological optimisation results generated through Ameba software, OptiGAN is able to predict a wide range of optimised architectural and structural designs under defined conditions.
Content may be subject to copyright.
2,*School of Architecture and Urban Design & Centre for Innovative
Structures and Materials, RMIT University
3Tsinghua University
4ByteDance, 0000-0002-6294-378X
2, *, 0000-0003-1395-8747, 0000-0002-5033-3597, 0000-0002-9291-2554
Abstract. With the rapid development of computers and technology
in the 20th century, the topological optimisation (TO) method has
spread worldwide in various fields. This novel structural optimisation
approach has been applied in many disciplines, including architectural
form-finding. Especially Bi-directional Evolutionary Structural
Optimisation (BESO), which was proposed in the 1990s, is widely used
by thousands of engineers and architects worldwide to design
innovative and iconic buildings. To integrate topological optimisation
with artificial intelligence (AI) algorithms and to leverage its power to
improve the diversity and efficiency of the BESO topological
optimisation method, this research explores a non-iterative approach to
accelerate the topology optimisation process of structures in
architectural form-finding via conditional generative adversarial
networks (GANs), which is named as OptiGAN. Trained with
topological optimisation results generated through Ameba software,
OptiGAN is able to predict a wide range of optimised architectural and
structural designs under defined conditions.
Keywords. BESO (Bi-directional Evolutionary Structural
Optimisation); Artificial Intelligence; Deep Learning; Topological
Optimisation; Form-Finding; GAN (Generative Adversarial
Networks); SDG 12; SDG 9.
1. Introduction
Structural optimisation, including topology optimisation, plays a significant role in
architectural design. It can increase the performance of structures and the efficiency of
material use and thus reduce material waste and carbon impact in the fabrication and
construction process. By integrating topology optimisation with artificial intelligence
POST-CARBON, Proceedings of the 27th International Conference of the Association for Computer-
Aided Architectural Design Research in Asia (CAADRIA) 2022, Volume 1, 121-130. © 2022 and
published by the Association for Computer-Aided Architectural Design Research in Asia (CAADRIA),
Hong Kong.
(AI) for more efficient use of materials in the industry, it seeks to help achieve the
United Nations Sustainable Development Goal 12: Ensure sustainable consumption
and production patterns (United Nations, 2015).
Structural optimisations aim to achieve the best structural performance while meeting
the requirement of various constraints. Over the past three decades, high-speed
computers and rapid improvements in algorithms have been used to develop better
structural optimisation solutions by a number of engineering researchers.
1.1.1. Topology Optimisation
Topology optimisation is one of the most popular optimal structural design methods
for discrete structures, such as trusses and frames. It is developed to search for the
optimal spatial order and connectivity of the bars. Topology optimisation of continuum
structures is to find optimal designs by determining cavities' best locations and
geometries in the design domains.
Topology optimisation can be readily used to perform shape optimisation by simply
restricting the structural modification to the existing boundaries (Huang & Xie, 2010).
In the field of topology optimisation, there are several notable methods based on finite
element analysis (FEA) developed, such as the homogenisation method (Bendsøe &
Kikuchi, 1988), the solid isotropic material with penalisation (SIMP) method (Bendsøe
& Sigmund, 1999), the evolutionary structural optimisation (ESO) (Xie & Steven,
1993), the bi-directional evolutionary structural optimisation (BESO) (Huang & Xie,
2010; Huang et al., 2007) and the level-set method (LSM) (Wang et al., 2003). In this
paper, bi-directional evolutionary structural optimisation (BESO) proposed by Huang
and Xie (2010) is adopted to develop a new integrated topology optimisation algorithm
(Figure 1).
Figure 1 Bi-directional evolutionary structural optimisation (BESO) result
1.1.2. Bi-directional Evolutionary Structural Optimisation (BESO)
Bi-directional evolutionary structural optimisation (BESO) is the emerging technology
that is an extension of evolutionary structural optimisation (ESO) developed by Xie
and Steven in 1992 (Xie & Steven, 1993). Both ESO and BESO algorithms are based
on finite element analysis (FEA) for topology optimisation of continuum structures.
BESO algorithm aims to find the solution with the highest structural performance
under certain material limitations by removing or adding material elements step by step
(Bao et al., 2020). The ESO method also inspires the Extended ESO method, widely
used in architecture design projects, such as the Akutagawa River Side Project in Japan
by Ohmori and Qatar National Convention Centre by Arata Isozaki, to generate an
optimised model with not only high structural performance but also some different
characteristics to meet more functional requirements or aesthetic preferences. In the
past few years, Mike Xie and his team have modified many detailed control strategies
for topology optimisation in architectural design and development during the process
(Yan et al., 2021).
1.1.3. Ameba Software
Because of the benefit of form-finding through topology optimisation and Bi-
directional evolutionary structural optimisation (BESO), more and more designers and
architects seek to use topology optimisation methods to design buildings and furniture.
However, due to the complexity and slow speed to directly use the algorithm for
architectural design, a new Rhinoceros plug-in named Ameba, a topology optimisation
tool based on the BESO method and FEniCS open-source computing platform (Zhou
et al., 2018), has been developed. More and more architects and designers have gained
opportunities to use this intelligent method to work with computers interactively to
create innovative, efficient, and organic architectural forms using Ameba. In this work,
the authors use it as the topology optimisation tool to form the dataset for training
generative adversarial networks to assist and investigate the research.
Allowed by the development of deep learning algorithms and fast-growing
computational power, artificial neural networks, including generative adversarial
networks (GANs), have been increasingly used in architectural and structural
explorations such as topology optimisation in the design process.
1.2.1. Generative Adversarial Networks
A generative adversarial network (GAN) is a particular artificial neural network that
learns from a collection of examples and their probability distribution. It is then able to
generate more examples from the estimated probability distribution (Goodfellow et al.,
2020). A typical GAN often consists of a generator that defines a prior probability
distribution P(z) based on a vector z as input and a discriminator which examines
whether data x is real (sampled from the training examples) or fake (sampled from the
output of the generator).
GANs can further be extended to conditional models (cGANs) where both the
generator and discriminator are conditioned on extra information y as input (Mirza &
Osindero, 2014). Besides examining whether x is real or fake, the discriminator of a
cGAN also evaluates whether it matches the condition y. For example, when using
cGANs to solve topology optimisation problems, x can be the expected optimisation
results, given the corresponding boundary and load conditions of y.
1.2.2. Topology Optimisation via Deep Learning
In recent years, there has already been some research into solving topology
optimisation problems through artificial neural networks, especially GANs. For
example, the TopologyGAN (Nie et al., 2021) is developed on a cGAN, whose
generator combines a U-Net (Ronneberger et al., 2015) and ResNet (He et al., 2016).
It takes displacement, load boundary conditions and target volume fraction augmented
with dense initial fields computed over the unoptimised domain as the input to the
model. By doing so, this method vastly improves the accuracy of predicting topology
optimisation results compared to some baseline models. Another research by Yu et al.
(2019) transforms the boundary conditions into multi-channel images as input to a
convolutional-neural-network-based encoder to generate low-resolution topology
optimisation results. It inputs the low-resolution results into a GAN to produce final
results in high resolution. Differently, the proposed method utilises cGANs directly. It
requires very brief input to keep the models easy to operate and thus broaden the
potential range of users with or without professional structural knowledge.
This research suggests an approach to accelerate the topology optimisation process of
structures in architectural form-finding by replacing iterative calculation procedures
with an end-to-end algorithm via conditional generative adversarial networks. This
method is named OptiGAN by the authors as the ultimate goal is to generate topology
optimisation results efficiently and accurately. Trained with a small number of
topological optimisation results generated with Ameba software, the proposed method
is able to predict a wide range of optimised two-dimensional structural forms under
defined conditions.
2. Methods
To achieve the research target, a coarse-to-fine network of cGANs is developed and
trained with a dataset collected by the authors.
To train OptiGAN, an original dataset of topology optimisation results is collected with
Ameba software. In detail, the material is kept as steel during the data generating
process, and the volume fraction is set to 0.5 consistently. The parameters that can vary
from case to case are design domain, fixing edge and load conditions, which are the
very parameters to use as input parameters of OptiGAN. In practice, in the Ameba
script, the fixing edge of a square design domain is always set to the left edge as fixing
at other edges are seen as the same as rotating left-edge-fixing conditions in the later
data augmentation process when training the cGAN models. In the first stage of the
research, a total number of 1385 optimisation results are included in the dataset.
During the pre-processing, the input parameters are translated into a three-channel
input image at 256 * 256 pixels in size. By doing so, the model is provided with two-
dimensional spatial clues. Thus, the difficulty for training the model to predict two-
dimensional results can be reduced compared to using merely numerical inputs directly
without spatial suggestions. Specifically, as Figure 2 demonstrates, the design domain
and fixing edge are expressed in all three channels by assigning the value of 0 to the
corresponding pixels representing the geometries. Load locations and load unit vectors
projected onto the X and Y axis are documented in the first and second channel
respectively in values remapped into a range between 0 and 255. Values of the other
pixels are assigned 255 in all three channels by default.
Figure 2 Translated input image of 1000*1000 mm design domain, fixing at the left edge, load one
at (900, 1000) location in 45° angle, load two at (1000, 700) location in 180° angle. The values of
the rest pixels are all 255.
Unlike the previous researches mentioned in section 1.2.2, there are no dense initial
fields, low-resolution results, or any other inputs than design domain, fixing edge and
load conditions for the OptiGAN generator. Keeping the inputs simple can make it
potentially as easy as adjusting a few number sliders for the OptiGAN users to operate.
At the same time, it dramatically increases the difficulty for the generator to speculate
the results by providing less input information. To respond to the conflict, OptiGAN
adopts a coarse-to-fine network architecture.
Specifically, as Figure 3 demonstrates, there are two generators and discriminators
in the network. The initial input of conditions in the form of translated input images is
first fed to the coarse generator, which then predicts a rough output examined by the
coarse discriminator. Then the rough output together with the initial conditions are
input to the fine generator, which outputs the final results. Although compared to
conventional cGANs, the input of OptiGAN consists only the conditions y alone
without vector z, the networks can also learn from only the conditions. This choice of
input is also being suggested in Pix2Pix (Isola et al., 2017), one of the most successful
image-to-image translation models.
Figure 3 The architecture of OptiGAN
2.2.1. Generators and Discriminators
Both the coarse and fine generators used in OptiGAN are U-Net (Ronneberger et al.,
2015), which has a mirrored encoder-decoder network architecture with skip
connections between symmetrical layers. It can work effectively with very few training
data. Meanwhile, both discriminators are PatchGAN (Isola et al., 2017), which focuses
on 156 * 156 patches as through testing, such patch size provides the best outcome in
the tasks.
2.2.2. Objective
The objective that OptiGAN tries to optimise can be expressed as equation (1). It
consists of two parts, the cGAN loss (equation (2)) and L1 loss (equation (3)). Besides
generating images that look real, as the other goal is to eventually achieve results as
close to the BESO optimisation outcomes as possible, L1 loss is added to the total loss
with a considerable weight of λ to force the output to be close to the ground truth. L1
loss is chosen out of L2 loss because it encourages less blurring effect of images
compared to L2 loss. In the experiments, the weight is set to 125 (λ = 125) to emphasise
the importance of L1 loss in this particular task.
=arg min
 (,)+ ℒ1 () (1)
cGAN(,) = ,[log (,)] +[log(1− (,())] (2)
1()=,[ − ()1] (3)
OptiGAN is trained in two steps: the coarse generator and discriminator are trained
first, after which the fine generator and discriminator are trained. Both parts of the
network are trained for 200 epochs to achieve the demonstrated results. Following the
conventions, instead of minimising log(1− (,()), the trainings maximise
log((,()) to avoid saturating log(1− (,()) in the early training stage.
The initial learning rate is 0.0005 decaying to 0 in the last 100 epochs and the Adam
optimiser is used. Figure 4 shows the history of the cGAN loss and L1 loss during the
two training procedures.
Figure 4 The cGAN loss and L1 loss during the coarse training and fine training
The coarse generator and fine generator take different types of input and are paired
with different discriminators as introduced in section 2.1.1, so it is not considered
practical to compare the absolute value of the corresponding cGAN losses of the coarse
and fine models. Even though, the cGAN loss improves from the beginning to the last
epoch in both cases, as shown in Figure 4. In contrast, for both generators, the L1 loss
is calculated according to the same ground truth, and it can be discovered that the fine
generator further decreases the L1 loss based on the coarse generator, which indicates
that the fine generator is able to further improve the accuracy of predicted topology
optimisation results by OptiGAN.
3. Results
During the training process, the L1 loss of OptiGAN reduced from over 80 to less than
20. More importantly and precisely, the pixel-wise accuracy is used to evaluate the
performance of the models. It is equal to the percentage of accurate pixels in a
prediction out of the total pixels of that image. Tested with 150 randomly selected
pieces of data different from the training set, the average pixel-wise accuracy of
OptiGAN is able to achieve 83.15%. Figure 5 demonstrates some of the testing results
in different load conditions with pixel-wise differences between the predictions and
ground truth visualised for each case.
Figure 5 Some prediction results of OptiGAN with ground truth and pixel-wise difference
4. Discussion
OptiGAN demonstrates the ability of a novel approach and its application in
architectural and structural form-finding. It is the extension of the SwarmBESO (multi-
agent-based topology optimisation) method proposed by Bao & Yan in 2020 (Bao et
al., 2021) to improve the diversity of the topological optimisation generative method.
It has the potential to significantly help architects and engineers save material and
produce more efficient structural layouts and building envelopes. It is valuable to
integrate two intelligent computational design methods, deep learning and topology
optimisation, for designers in the conceptual design phases.
However, the research is in a rudimentary phase and is temporarily constrained in
a range of two-dimensional solutions. Although set as an input variable, the design
domain in the current dataset includes only square geometries, despite that it can
perform well in this geometric range, as demonstrated in the testing results. To truly
diversify the spectrum of results and further increase the accuracy, it is very crucial that
OptiGAN must be trained with much more data of various design domains and load
conditions. Future line of research also includes further equipping the model with the
ability to solve three-dimensional topology optimisation problems.
5. Conclusion
This research develops OptiGAN, a non-iterative method to accelerate the topology
optimisation process of structures in architectural form-finding via conditional
generative adversarial networks with high accuracy. It demonstrates the process of
integrating topology optimisation and generative adversarial networks to establish an
artificial intelligence (AI) based structural optimisation technique. This new
methodology holds great potential for practical application in architecture and
engineering fields. It increases the diversity of outcome of the topology optimisation
generative design such as the application of shell (Figure 6).
Figure 6 Diverse BESO results of shell optimisation
We thank Nanjing Ameba Engineering Structure Optimization Research Institute for
providing educational version of Ameba software to support our research. (For Ameba,
Bao, D. W., Yan, X., Snooks, R., & Xie, Y. M. (2020). Bioinspired Generative Architectural
Design Form-Finding and Advanced Robotic Fabrication Based on Structural
Performance. In Architectural Intelligence (pp. 147-170), Springer.
Bao, D. W., Yan, X., Snooks, R., & Xie, Y. M. (2021). SwarmBESO: Multi-agent and
evolutionary computational design based on the principles of structural performance. In
26th International Conference of the Association for Computer-Aided Architectural
Design Research in Asia: PROJECTIONS CAADRIA 2021, Hong Kong (pp. 241-250).
Bendsøe, M. P., & Kikuchi, N. (1988). We are generating optimal topologies in structural
design using a homogenisation method. Computer methods in applied mechanics
engineering, 71(2), 197-224. doi:
Bendsøe, M. P., & Sigmund, O. (1999). Material interpolation schemes in topology
optimisation. Archive of applied mechanics, 69(9), 635-654.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Bengio, Y.
(2020). Generative adversarial networks. Communications of the ACM, 63(11), 139-144.
He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. In
Proceedings of the IEEE conference on computer vision and pattern recognition, Las
Vegas, (pp. 770-778).
Huang, X., & Xie, M. (2010). Evolutionary topology optimisation of continuum structures:
methods and applications: John Wiley & Sons.
Huang, X., Xie, Y. M., & Burry, M. C. (2007). Advantages of bi-directional evolutionary
structural optimisation (BESO) over evolutionary structural optimisation (ESO). Advances
in Structural Engineering, 10(6), 727-737.
Isola, P., Zhu, J.-Y., Zhou, T., & Efros, A. A. (2017). Image-to-image translation with
conditional adversarial networks. In Proceedings of the IEEE conference on computer
vision and pattern recognition (pp. 1125-1134).
Mirza, M., & Osindero, S. (2014). Conditional generative adversarial nets. arXiv preprint
arXiv: 1411.1784.
Nie, Z., Lin, T., Jiang, H., & Kara, L. B. (2021). Topologygan: Topology optimisation using
generative adversarial networks based on physical fields over the initial domain. Journal
of Mechanical Design, 143(3), 031715.
Ronneberger, O., Fischer, P., & Brox, T. (2015). U-net: Convolutional networks for
biomedical image segmentation. In International Conference on Medical image
computing and computer-assisted intervention, (pp. 234-241).
United Nations. (2015, September 25). Transforming our world: The 2030 agenda for
sustainable development. Retrieved January 20, 2022, from: https://
Wang, M. Y., Wang, X., & Guo, D. (2003). A level set method for structural topology
optimisation. Computer methods in applied mechanics engineering, 192(1-2), 227-246.
Xie, Y. M., & Steven, G. P. (1993). A simple evolutionary procedure for structural
optimisation. Computers & structures, 49(5), 885-896. doi:
Yan, X., Bao, D., Zhou, Y., Xie, Y., & Cui, T. (2021). Detail control strategies for topology
optimisation in architectural design and development. Frontiers of Architectural
Research, 10, 1-17. doi:
Yu, Y., Hur, T., Jung, J., & Jang, I. G. (2019). Deep learning for determining a near-optimal
topological design without any iteration. Structural Multidisciplinary Optimization, 59(3),
787-799. doi:
Zhou, Q., Shen, W., Wang, J., Zhou, Y. Y., & Xie, Y. M. (2018). Ameba: A new topology
optimisation tool for architectural design. In International Association for Shell and
Spatial Structures (IASS), (pp. 1-8).
Full-text available
Topology optimization is a computationally expensive process, especially when complicated designs are studied, and this is mainly due to its finite element analysis and iterative solvers incorporated into the algorithm. In the current work, we investigated the application of deep learning methods to computationally accelerate topology optimization. We tested and comparatively analyzed three types of improved neural network models using three different structured datasets and achieved satisfactory results that allowed for the generation of topology optimized structures in 2D and 3D domains. The results of the studies show that the improved Res-U-Net and U-Net are reliable and effective methods among deep learning approaches for the computational acceleration of topology optimization problems. Moreover, based on the results, it is evaluated that Res-U-Net gives better results than U-Net for higher iterations. We also showed that the proposed CNN method is highly accurate and required much less training time compared to existing methods.
Full-text available
With the ability to generate forms with high efficiency and elegant geometry, to-pology optimization has been increasingly used in architectural and structural designs. However , the conventional topology optimization techniques aim at achieving the structurally most efficient solution without any potential for architects or designers to control the design details. This paper introduces three strategies based on Bi-directional Evolutionary Structural Optimization (BESO) method to artificially pre-design the topological optimized structures. These strategies have been successfully applied in the computational morphogenesis of various structures for solving practical design problems. The results demonstrate that the developed methodology can provide the designer with structurally efficient and topologically different solutions according to their proposed designs with multi-filter radii, multi-volume fractions, and multi-weighting coefficients. This work establishes a general approach to integrating objective topology optimization methods with subjective human design preferences, which has great potential for practical applications in architecture and engineering industry.
Conference Paper
Full-text available
This paper posits a design approach that integrates multi-agent generative algorithms and structural topology optimisation to design intricate, structurally efficient forms. The research proposes a connection between two dichotomous principles: architectural complexity and structural efficiency. Both multi-agent algorithms and Bi-directional evolutionary structural optimisation (BESO) (Huang and Xie 2010), are emerging techniques that have significant potential in the design of form and structure.This research proposes a structural behaviour feedback loop through encoding BESO structural rules within the logic of multi-agent algorithms. This hybridisation of topology optimisation and swarm intelligence, described here as SwarmBESO, is demonstrated through two simple structural models. The paper concludes by speculating on the potential of this approach for the design of intricate, complex structures and their potential realisation through additive manufacturing.
Full-text available
In topology optimization using deep learning, load and boundary conditions represented as vectors or sparse matrices often miss the opportunity to encode a rich view of the design problem, leading to less than ideal generalization results. We propose a new data-driven topology optimization model called TopologyGAN that takes advantage of various physical fields computed on the original, unoptimized material domain, as inputs to the generator of a conditional generative adversarial network (cGAN). Compared to a baseline cGAN, TopologyGAN achieves a nearly 3x reduction in the mean squared error and a 2.5x reduction in the mean absolute error on test problems involving previously unseen boundary conditions. Built on several existing network models, we also introduce a hybrid network called U-SE(Squeeze-and-Excitation)-ResNet for the generator that further increases the overall accuracy. We publicly share our full implementation and trained network.
Full-text available
Due to the potential to generate forms with high efficiency and elegant geometry, topology optimization is widely used in architectural and structural designs. This paper presents a working flow of form-finding and robotic fabrication based BESO (Bi-directional Evolutionary Structure Optimization) optimization method. In case there are some other functional requirements or condition limitations, some useful modifications are also implemented in the process. With this kind of working flow, it is convenient to foreknow or control the structural optimization direction before the optimization process. Furthermore, some fabrication details of the optimized model will be discussed because there are also many notable technical points between computational optimization and robotic fabrication.
Full-text available
In this study, we propose a novel deep learning-based method to predict an optimized structure for a given boundary condition and optimization setting without using any iterative scheme. For this purpose, first, using open-source topology optimization code, datasets of the optimized structures paired with the corresponding information on boundary conditions and optimization settings are generated at low (32 x 32) and high (128 x 128) resolutions. To construct the artificial neural network for the proposed method, a convolutional neural network (CNN)-based encoder and decoder network is trained using the training dataset generated at low resolution. Then, as a two-stage refinement, the conditional generative adversarial network (cGAN) is trained with the optimized structures paired at both low and high resolutions, and is connected to the trained CNN-based encoder and decoder network. The performance evaluation results of the integrated network demonstrate that the proposed method can determine a near-optimal structure in terms of pixel values and compliance with negligible computational time.
Full-text available
We propose a new framework for estimating generative models via an adversarial process, in which we simultaneously train two models: a generative model G that captures the data distribution, and a discriminative model D that estimates the probability that a sample came from the training data rather than G. The training procedure for G is to maximize the probability of D making a mistake. This framework corresponds to a minimax two-player game. In the space of arbitrary functions G and D, a unique solution exists, with G recovering the training data distribution and D equal to 1/2 everywhere. In the case where G and D are defined by multilayer perceptrons, the entire system can be trained with backpropagation. There is no need for any Markov chains or unrolled approximate inference networks during either training or generation of samples. Experiments demonstrate the potential of the framework through qualitative and quantitative evaluation of the generated samples.
Conference Paper
There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at .
The evolutionary structural optimization (ESO) method evolves a structure from the full design domain towards an optimum by gradually removing inefficient material. The bi-directional ESO (BESO) may start from any initial design and evolve a structure to an optimum by adding and removing material simultaneously. In this paper, a detailed comparison between ESO and BESO has been carried out for stiffness optimization problems. Both 2D and 3D examples shows that the BESO method possesses many advantages over the ESO method such as computational efficiency, robustness of the method and manufacturability of the final topology.