144

Generative Design: An Explorative Study

Francesco Buonamici1 , Monica Carfagni1 , Rocco Furferi1, Yary Volpe1 and Lapo Governi1

1Department of Industrial Engineering of Florence, [email protected]

2Department of Industrial Engineering of Florence, [email protected]
3Department of Industrial Engineering of Florence, [email protected]
4Department of Industrial Engineering of Florence, [email protected]
5Department of Industrial Engineering of Florence, [email protected]

Corresponding author: Francesco Buonamici, francesco.buonamici@unifi.

Abstract. Generative design tools have recently become an interesting solution to

tackle design problems in several technical fields. This article takes into
consideration the specific field of mechanical design and aims at describing
available generative design solutions capable of dealing with structural optimization
problems. The study provides a practical description on the workflow and
performances of a specific software system implementing a generative approach for
the generation of a set of alternative solutions for a static structural design
problem. The software analyzed is Autodesk’s Generative Design, hosted in Fusion
360. The article discusses the functioning of the software and its performances; an
enhanced focus on the features oriented to the generation of manufacturable
shapes is provided in the text. In order to provide a practical and effective
procedure, a literature case study was selected to test the software.

Keywords: Generative Design; Topology Optimization; CAD; Structural

Optimization; Biomimetic Design
DOI: https://doi.org/10.14733/cadaps.2021.144-155

1 INTRODUCTION
Generative approaches are recently becoming more and more applied in a variety of technical
fields. By implementing artificial intelligence tools, they can elaborate and propose to a human
user a series of plausible solutions for a design problem. That is, a number of alternative
configurations that i) satisfy a set of imposed design constraints and ii) try to maximize a goal
function passed to the algorithm. The proposed alternatives are the result of an iterative
exploration of the related solution space that is guided by an artificial intelligence.
Thanks to the significant increase of computing power available, these tools have recently
observed a growing interest in the design community. Generative Design (GD) has taken its first
steps in the architectural field [7,8][13] and has generally been first applied in open-problem
scenarios characterized by large design spaces. In this context, the term “Generative Design”

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 145

refers to a series of tools, implementing artificial intelligence methods and algorithms, applied to
solve design problems. From a practical point of view, GD tools essentially seek for a solution of a
problem expressed with a mathematical formulation; this often results in an iterative optimization
process that tries to minimize an objective function. Accordingly, GD has proven itself useful to
identify uncommon solutions that do not fall within the typical set of shapes or configurations used
(consider, for instance, Figure 1). As a result, GD tools have been first exploited mainly to
encourage divergent thinking and creativity. This remains a distinctive aspect of the technology
that has brought to its application in areas where aesthetics and innovativeness are important in
the product development process, such as product design [3][14] or the automotive sector (e.g.
[1][9]).

Figure 1: Philippe Starck's A.I. chair for Kartell [14].

More recently, GD has been applied in the mechanical field to pursue performance-driven design
[4][11][16]. In this case, research focused on the development of AI tools to deal with structural
optimizations; in this case, the part’s plausible topologies are coupled with their FE-simulated
structural behaviors to identify the most performing shapes. As in Topology Optimization (TO),
which shares many features to GD, parts can be optimized according to compliance or minimum
weight while maintaining a safety factor or a deflection. While the application of this kind of GD
resources as everyday tools is still a distant target, their use is being encouraged by the
development of more reliable and easy-to-use tools. Simultaneously, the functionalities offered by
such software systems have been continuously improved, especially considering the ranges of
possible design constraints and load conditions offered.
One of the companies who has arguably invested more in the development of GD tools and in
their integration within traditional CAD environments is Autodesk. The software company has
launched its Dreamcatcher project [12], dedicated to the development of GD tools back in 2014;
five years of development have brought to the release of the first version of GD commercial
software. Autodesk’s tool is called “Generative Design” [15] and is hosted within Fusion 360 [5], a
parametric CAD modeler. Autodesk’s GD (AGD) is essentially a suite of the CAD application that
optimizes the shape of a component to comply with a static structural load condition. Cloud
computing functionalities are exploited to run multiple FE analyses and obtain a result in an
acceptable time window.
While TO analyses have become a standard tool, GD potentialities have not been fully
explored, as AGD is the first tool that can be applied without a heavy tuning and setup phase.
These features have brought GD tools within everyone’s reach, but the skills required for a
conscious application are not yet fully formed and widespread. Moreover, there is not much
information on the performance of the technology and the benefit that its application could

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 146

guarantee, especially in the industrial context. Considering AGD, its implementation details, its
framework and the features offered by the software have not been fully discussed yet.
Accordingly, this article aims at a practical and effective description of the design workflow
offered by AGD. The study has been carried out applying AGD to a static structural optimization
problem and carrying out the entire design phase using the tools offered by the AGD suite. As
explained in the remainder of the paper, the case study was selected from the literature in order to
be able to compare the results with those obtained by means of traditional tools.

2 CASE STUDY
The case study selected to test the functioning of the AGD is the design challenge proposed at
[2]. The component to be designed is a gripper arm that is part of a robot deputed to the handle
meteors. The overall shape of the part is assigned and depicted in Figure 2; the areas to be
maintained intact are marked in red in the figure. The overall goal consists of the reduction of
weight of the component within the limits imposed by the design requirements summarized in
Table 1. The load case is characterized by a static load of 20000 N applied orthogonal to the
gripping surface (see Figure 2), while the part is constrained on two cylindrical surfaces that
interact with the rest of the robot. A maximum deflection of 8mm is allowed while guaranteeing a
safety factor of 3.

Parameter Value
Force applied to gripping surface 20000 N
Max deflections 8 mm
Safety factor >3
Material ASTM A36

Table 1: Design requirements of the challenge.

Figure 2: Design constraints of the part to be optimized.

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 147

3 AGD FRAMEWORK
The framework proposed by AGD heavily relies on cloud-computing: while most phases are carried
out by the user on its personal workstation, all the optimization and the analyses are performed in
external servers. The phases composing the AGD’s framework are similar to those that can be
found within a TO analysis. The main difference is the generation, at a first level, of a series of
shapes that need to be “explored” to identify the most effective solution. The overall framework
can be schematized in the following phases, depicted in Figure 3:
• Objectives – two options (“Minimize mass” and “Maximize stiffness”) can be selected as
goal for the analysis. In both cases, a SF is required. If the second option is selected, the
user should provide also a target mass that the optimization should achieve.
• Geometry – the user defines the areas that should be left intact by the optimization
(Preserve regions) and the volumes that must remain empty (Obstacle regions). One of the
great differences w.r.t. the classic TO approach is that AGD does not require the definition
of a starting volume (Design space) to be progressively hollowed. The optimization can be
initiated with a plausible Starting Shape (SS) that can guide the optimization, but this
choice is optional.
• Load cases – AGD supports forces, pressures and bearing loads. Gravity can be included in
the analysis as well. Available constraints are identified as fixed, pinned and frictionless.
Multiple load cases can be considered by the solver; dynamic conditions, on the other
hand, cannot be introduced. All the loads and constraints need to be applied to preserve
regions.
• Manufacturing constraints – they can be provided by the user to guide the analysis towards
shapes that can be manufactured using a specific process (additive manufacturing, 5-axis
milling, 3-axis milling), hence reducing the production costs for the part fabrication. This
advanced feature is being introduced also in some TO software.
• Material –only linear-elastic models can be used up to this moment. AGD allows the
concurrent selection of up to ten materials in a single analysis.
• Input Check & computation – AGD checks if the required information is provided. In case of
a positive answer, Cloud-computing is performed after the payment of a fixed fee (cloud
credits).
• Results – once that the results are downloaded on the local machine, these are available to
the user; depending on the setting of the optimization, the results may be in the order of
dozens.
• Exploration – AGD has a dedicated environment that offers visualization tools to map the
results in an ordered manner to help the user to identify the best possible solution. Results
can be plotted according to their mechanical and physical properties.
• Selection – the user identifies the design that best fits the desired behavior and exports it
from the visualization environment.
• Export – the design is isolated and made available for further modifications to the user; the
CAD geometry of the part is imported within the modelling environment of Fusion 360. This
phase is associated with a cost, as the software requires an extra payment for every design
exported from the visualization environment.
• Modification – once that the design is exported, it should be edited with traditional CAD
tools to amend defects that are typically present in complex shapes, characterized by
multiple B-REP surfaces. This is required also in TO applications, where the tessellated
shape produced by the optimization need to be edited and beautified to allow fabrication.
• Validation – The performances of the exported shape need to be validated by performing
an additional set of FE analyses. Independently from the modifications introduced in the
previous phases, some discrepancies can be found by comparing FE results obtained in the
exploration phase and the final design. Accordingly, a validation is usually recommended to
assess the final mechanical properties of the part.

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 148

The user is facilitated in the execution of all the steps by a simplified GUI that proposes the steps
in a chronological order. Moreover, a series of automatic intermediate controls indicate to the user
the possible lack of data essential to perform a subsequent task.

Figure 3: Autodesk Generative Design Framework.

4 RESULTS
The AGD framework has been applied to the selected test case in order to test the functionalities
of the software. It’s important to note that the scope of this paper is to evaluate the features
offered by the AGD tool rather than carry out a mere design activity of a component. As a
consequence, while the part optimization will be carried out following the indication provided at
[2], a more general perspective will be maintained to try to understand how the tool can be used
beyond the individual case study.
In order to consider as many aspects as possible, four different materials and production
technologies were introduced in the analysis, regardless of the requirements imposed by the case
study. The materials were selected from Fusion 360 library, starting from materials similar to the
A36 steel required by the challenge and including also high-performance ones (e.g. Ti6Al4V,
AISI304). Four possible alternatives were selected as possible production technologies: 5-axes
milling, 3-axes milling, additive manufacturing and unrestricted. Each technology introduces some
design constraint to allow manufacturing of the produced part. Milling, for example, requires the
definition of a tool geometry to compute volumes to leave empty to allow tool access. Additive
manufacturing considers the generation of overhang surfaces. The unrestricted modality, on the
other hand, corresponds to an unconstrained optimization. Two separate analyses have been
carried out: a first one, without any starting shape used to guide the optimization, and a second
one that uses the shape visible in Figure2 as starting configuration.
Each analysis brought to the generation of 28 shapes that the user can consider as plausible
solutions. Tables 2 and 3 show that candidate solutions can be divided into categories. AGD
differentiates between converged and completed solutions. The outcomes marked as “completed”
(5 in the considered application) have either not met the design criteria set during the setup, or
the optimization failed to produce a fully converged result. Accordingly, while these results are
made available anyway to the user, they should be carefully taken into consideration because they
might not be suitable for the application.

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 149

Manufacturing constraint # Solutions Converged #Solutions Completed

Unrestricted 4 0
5-axes milling 3 1
3-axes milling 7 1
Additive 9 3
Total 23 5

Table 2: Number of solutions produced by the AGD depending on the manufacturing constraint.

Material # Solutions Converged #Solutions Completed

ASTM A36 6 1
ASTM A572 6 1
Titanium 6Al4V 5 2
AISI 304 6 1
Total 23 5

Table 3: Number of solutions produced by the AGD depending on the material.

The solutions can be explored by means of graphs automatically generated by the software, which
allow an efficient mapping of the solutions: the user can select which parameters use on the axes
and how to group the solution to identify general trends. Figure 4 shows an example where the
solutions generated in the analysis are mapped in terms of mass vs. max displacement. These
graphs are functional, as they allow to dynamically compare solutions that differ significantly under
multiple aspects (e.g. material, manufacturing technology, macroscopic shape, etc.) this is a
crucial step considering the high number of solutions generated by the AGD.

Figure 4: Results dispersion graph produced by the AGD exploration environment, mass vs max
displacement, solutions grouped according to the material.

Figure 5 shows a subset of shapes produced in the two AGD studies (with or without starting
shape). The introduction of manufacturing constraints evidently influences the type of results
produced. Solutions obtained under the same hypothesis (grouped by color or by row in Figure 5)

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 150

evidently share similarities in the geometry produced by the algorithm, with main features that are
maintained constant within each group of solutions. Even considering all the solutions produced
with a starting shape (row 2 and row 4 of Figure 5) is possible to identify significant similarities.
Specifically, the envelope of all the shapes roughly resembles the starting shape.
However, some limits in the compliance of the imposed manufacturing constraints can be
identified. While the global features required by the selected process are verified in the produced
results, the software occasionally fails on a local level in the generation of details that are not
manufacturable or introducing errors that affect the surface quality. A valid example, with this
respect, is represented by the generation of corrugated overhang surfaces in AM parts, as showed
in Figure6. As a consequence, a subsequent editing phase is required in order to refine and polish
the shapes identified by the AGD analysis.

Figure 5: A selection of shapes produced by the two analyses.

Figure 6: Local defects introduced in AM parts.

Table 4 reports a selection of results produced by the AGD. The structural behavior of the
generated shape is valid for the selected application, as they satisfy the constraints of Table 1. The
mechanical performances of the best solutions produced by the AGD are comparable with the best

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 151

solutions obtained by means of TO or traditional design methods (e.g. possible solutions proposed
in the challenge) (Table 5).
The element of novelty that is introduced by the AGD is the possibility of choosing among a
series of solutions, equally valid from a structural point of view. Accordingly, the user can draw
from their own experience in order to identify the solution that fits all the design criteria, even the
ones not expressed in the AGD study. Depending on the application, different considerations might
be relevant and could improve the value of the product: ergonomics, aestethics, manufacturability
and industrial know-how.

No SS #7 No SS #8 With SS #4 With SS #8

Material A36 A572 A36 A572

Manufacturing M5ax U AM U
Mass [kg] 33.6 25.2 24.3 19.4
Max Displacement
2.19 3.2 1.78 2.37
[mm]

Table 4: Mechanical performance of a subset of AGD-generated results.

TO#1 TO#2 Challenge#1 Challenge#2

Material A36 A36 A36 A36

Manufacturing - - - -
Mass [kg] 44.6 63.3 38.4 33.1
Max
Displacement 1.08 0.67 1.09 1.67
[mm]

Table 5: Mechanical performance of a subset of AGD-generated results, compared with results

obtained by means of traditional techniques (TO and shapes generated by a human user through a
trial-and-error process). Challenge #1 and #2 are results proposed in the context of the design
challenge.

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 152

In order to test the tools to export a result from the visualization environment, the solution
marked as “no SS#7” of Table 4 (visible in Figure7) ohas been selected as best candidate and
therefore exported. During this phase, a T-spline model is generated from the mesh obtained in
the analysis. This introduces some local defects in the model that need to be manually amended.
Finite Element analyses (see Figure 7) were carried out on the final shape of the part to confirm
previously-obtained results. The same analysis has been performed also on an external FE
software system to confirm the results (see Figure 8).

Figure 7: FE analyses performed within Fusion 360 on the final part: a) safety factor b) Von-Mises
Stress values.

Figure 8: FE analyses performed within Solidworks Simulation on the final part: Von-Mises Stress
values.

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 153

Finally, this study investigated the manufacturability of the produced shape. As previously
discussed, AGD allows the introduction of manufacturing constraints that are taken into
consideration in the analyses. However, while their general effect is visible comparing the class of
shapes produced in each category of Figure 4, no additional data on the effects introduced by
activating such constraints is provided by the software. Accordingly, the manufacturability of the
exported shape has been verified using the CAM software included in Fusion 360 (see Figure 9) to
compute possible machining paths (the component was optimized hypothesizing a milling
operation). After that the manufacturability of the part was verified, the fabrication time obtained
in the simulation1 has been compared with the one required to manufacture a topologically
optimized part (TO#2 in Table 5). This was done to identify if the introduction of manufacturing
constraints allows the identification of shapes that are optimized even considering the
manufacturability.
A total fabrication time of 2:24:58, corresponding to a total machining path length of 175.199
m, was computed for the AGD part, while a time of 2:41:34 and a 204.96m path were obtained for
the TO part. It’s important to highlight that this represents only a partial confirmation of the
advantages achievable with AGD manufacturing constraints as a much larger pool of possible
solutions should be considered to integrate the present study on this aspect. Moreover, different
test case should be considered. In fact, the one that was selected in this study is essentially prone
to generate planar shapes with a variable presence of reticular structures. More complex load
conditions or different solution spaces might be able to better highlight the differences between
AGD results and the ones obtained with traditional methods, especially considering the
manufacturability of the parts.

Figure 9: CAM simulations carried out on the a) AGD part and b) TO part to evaluate their
fabricability.

5 CONCLUSIONS
In spite of the continuous development of structural optimization software tools, designers’
experience will always remain an essential element of the design process. Indeed, the ability to
analyze design problems and to identify driving factors that play a major role towards the

1 This was computed on a scaled-down part in order to ease the computation of machining paths and work with a part
with global dimensions suitable for a milling operation.

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 154

achievement of a high-quality result remains a human strength that cannot be easily mimicked by
AI tools. AGD proposes a promising approach to the design problem, exploiting advanced
computation tools where their application is most favorable, i.e. to tackle problems that can be
easily expressed mathematically, and leaving to the user the task of identifying the best solution
among a set of equally valid candidates.
The results of this study indicate that the performances offered by AGD allow the application of
the considered tool in real case scenarios. The optimization performances seem to be comparable
with the traditional tools which are nowadays commonly applied in design (e.g. TO [6][10]). As
showed by the comparison of results in Table 1, the application of AGD has not imposed severe
limitations in the identification of the global optimum. Moreover, a direct comparison of the shapes
produced by the tool highlights the effects caused by the introduction of manufacturing constraints
in the analysis: this innovative feature could raise the interest towards AGD, as this is one of the
most important factors considered at the moment in AM applications [10]. Although it’s possible to
identify some similarities between the behaviour of traditional TO methods and the application of
AGD if a SS is provided to start the optimization, the two approaches remain fundamentally
separate. TO always proceed by removing material from the design volume, while AGD maintains
the possibility of adding material and, generally, to deviate from the initial SS provided.
Accordingly, the indication of a SS serves as guideline for the optimization, rather that as a proper
design space.
However, strong limitations are perceived considering, for instance, the available tools for
modelling of structural conditions. Currently available loads and constraints limit the application of
the AGD to a subset of the typical mechanical situations (i.e. static pure-structural optimizations).
With this respect, the advisable introduction of dynamic and thermic loads would be the first step
to promote a wider application of the tool. It’s also important to highlight that, while Fusion 360 is
advertised as a low-cost solution (and completely free for non-professional use), the AGD tools
requires the payment of a fixed fee to launch each analysis or to export a single result in order to
obtain a CAD model for downstream applications.
Sadly, a more in-depth technical discussion on the functioning of the AGD tool is not
achievable, at the moment, due to the limitness of available information provided by the
developers. This will be a key area to focus on during future work: indeed, a higher comprension
of the mechanisms responsible for the design choices in engineering task is advisable in order to
provide the user with all the desiredable tools to make informed choice. The present paper is only
a first step towards this goal. Considering its overall performances, AGD could be considered as a
solid starting point for a design activity whenever: i) the skillset of the designer is not sufficient to
identify right from the start a valid shape for a critical component; ii) performance design needs to
be integrated with other constraints.

REFERENCES
[1] Autodesk generative design used on electric VW bus concept,
https://www.pesmedia.com/autodesk-generative-design-fusion-360-volkswagen-vw-bus/,
Autodesk.
[2] Autodesk Robot Gripper Arm Design Challenge | Engineering & Design Challenges,
https://grabcad.com/challenges/autodesk-robot-gripper-arm-design-challenge, GrabCAD.
[3] Alcaide-Marzal, J.; Diego-Mas, J.A.; Acosta-Zazueta, G.: A 3D shape generative method for
aesthetic product design, Design Studies, 66, 2020, 144–176.
https://doi.org/10.1016/j.destud.2019.11.003
[4] Arias-Rosales, A.; Osorio-Gómez, G.: Albatros Create: an interactive and generative tool for
the design and 3D modeling of wind turbines with wavy leading edge, International Journal
on Interactive Design and Manufacturing, 2020. https://doi.org/10.1007/s12008-020-
00655-y
[5] Cloud Powered 3D CAD/CAM Software for Product Design | Fusion 360,

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net
 155

https://www.autodesk.com/products/fusion-360/overview, Autodesk.
[6] Bianchi, M.; Buonamici, F.; Furferi, R.; Vanni, N.: Design and optimization of a
flexion/extension mechanism for a hand exoskeleton system, Proceedings of the ASME
Design Engineering Technical Conference, 1A-2016, 2016.
https://doi.org/10.1115/DETC2016-59466
[7] Caetano, I.; Santos, L.; Leitão, A.: Computational design in architecture: Defining
parametric, generative, and algorithmic design, Frontiers of Architectural Research, 2020.
https://doi.org/10.1016/j.foar.2019.12.008
[8] Caldas, L.G.; Norford, L.K.: A design optimization tool based on a genetic algorithm,
Automation in Construction, 11, 2002, 173–184. https://doi.org/10.1016/S0926-
5805(00)00096-0
[9] In Full Swing - Autodesk’s generative tools shift Lightning Motorcycles into gear - TCT
Magazine, https://www.tctmagazine.com/3d-software-news/in-full-swing-autodesk-
generative-design-lightning-motorcycles/
[10] Meli, E.; Rindi, A.; Ridolfi, A.; Furferi, R.; Buonamici, F.; Iurisci, G.; et al.: Design and
production of innovative turbomachinery components via topology optimization and additive
manufacturing, International Journal of Rotating Machinery, 2019.
https://doi.org/10.1155/2019/9546831
[11] Plocher, J.; Panesar, A.: Review on design and structural optimisation in additive
manufacturing: Towards next-generation lightweight structures, Materials and Design, 183,
2019. https://doi.org/10.1016/j.matdes.2019.108164
[12] Project Dreamcatcher | Autodesk Research,
https://www.autodeskresearch.com/projects/dreamcatcher, Autodesk.
[13] Shea, K.; Aish, R.; Gourtovaia, M.: Towards integrated performance-driven generative
design tools, Automation in Construction, 14, 2005, 253–264.
https://doi.org/10.1016/j.autcon.2004.07.002
[14] Starck: | design | furniture | chairs, https://www.starck.com/a-i-for-kartell-by-starck-
powered-by-autodesk-kartell-p3534, Starck.
[15] What is Generative Design | Tools & Software,
https://www.autodesk.com/solutions/generative-design, Autodesk.
[16] 6 Examples of Generative Design in Manufacturing,
https://www.autodesk.com/redshift/generative-manufacturing/, Autodesk.

Computer-Aided Design & Applications, 18(1), 2021, 144-155

© 2021 CAD Solutions, LLC, http://www.cad-journal.net