Artificial Intelligence for Engineering Design, Analysis and Manufacturing (2014), 28, 277–283.

# Cambridge University Press 2014 0890-0604/14 $25.00

doi:10.1017/S0890060414000225

Generative spatial performance design system

BENJAMIN P. COOREY1 AND JULIE R. JUPP2

1
School of Architecture, University of Technology, Sydney, Sydney, Australia
2
School of the Built Environment, University of Technology, Sydney, Sydney, Australia
(RECEIVED March 15, 2013; ACCEPTED February 25, 2014)

Abstract
Architectural spatial design is a wicked problem that can have a multitude of solutions for any given brief. The information
needed to resolve architectural design problems is often not readily available during the early conceptual stages, requiring
proposals to be evaluated only after an initial solution is reached. This “solution-driven” design approach focuses on the
generation of designs as a means to explore the solution space. Generative design can be achieved computationally through
parametric and algorithmic processes. However, utilizing a large repertoire of organiational patterns and design precedent
knowledge together with the precise criteria of spatial evaluation can present design challenges even to an experienced ar-
chitect. In the implementation of a parametric design process lies an opportunity to supplement the designer’s knowledge
with computational decision support that provides real-time spatial feedback during conceptual design. This paper presents
an approach based on a generative multiperformance framework, configured for generating and optimizing architectural
designs based on a precedent design. The system is constructed using a parametric modeling environment enabling the cap-
ture of precedent designs, extraction of spatial analytics, and demonstration of how populations can be used to drive the
generation and optimization of alternate spatial solutions. A pilot study implementing the complete workflow of the system
is used to illustrate the benefits of coupling parametric modeling with structured precedent analysis and design generation.
Keywords: Generative Design; Parametric Design; Performance Driven Design; Space Syntax; Spatial Analysis

1. INTRODUCTION ferentiates between “limiting preconceptions,” which “restrict

the freedom of the designer,” and “enabling preconceptions,”
A building is more than a list of activities or rooms; it is a pat-
which “give the designer greater freedom.” Preconceptions
tern of space that abides by conventions that determine the
can be in the form of guiding principles, primary generators,
type and number of spaces required, how they are connected,
or design schemas. The repertoire of organizational patterns
sequenced, and which activities should be grouped together
and design precedent knowledge as well as the precise criteria
or segregated. Architectural spatial design is a wicked prob-
and computation of spatial evaluation required for generative
lem that has a multitude of solutions for any given brief.
exploration is more than what can be expected from the accu-
The information needed to resolve architectural design prob-
mulated knowledge of an experienced architect (Schumacher,
lems is often not readily available during the early conceptual
2012). There is a need to supplement a designer’s knowledge
stages, requiring proposals to be evaluated only after an initial
with a computational decision support system that can provide
solution is reached. This “solution-driven” design approach
real-time spatial feedback during conceptual design.
focuses on the generation of solutions as a means to explore
This paper presents an exploration of parametric modeling
large solution spaces. Generative design can be achieved
techniques as a mechanism for constructing design logic
computationally through parametric and algorithmic pro-
models, in particular to support spatial feedback during de-
cesses. This requires the codification of design logic and
sign. The explorations are based around the four mathemati-
hence an understanding of the design process.
cal models described in Alfaris and Merello’s (2008) genera-
Kruger and Cross (2006) show that designers rely heavily on
tive multiperformance design system (GMPDS) as a
preconceptions in solution-driven design. Janssen (2005) dif-
conceptual and theoretical starting point for developing a spa-
tial performance design system; the system is implemented
Reprint requests to: Benjamin P. Coorey, School of Architecture, Univer-
sity of Technology Sydney, Sydney, Australia. E-mail: Benjamin.Coorey@ through a pilot study to demonstrate how individual paramet-
uts.edu.au ric logic models can be integrated.
277

https://doi.org/10.1017/S0890060414000225 Published online by Cambridge University Press

 278 B.P. Coorey and J.R. Jupp

2. PARAMETRIC DESIGN SYSTEMS (1999) provide insight into techniques for comparing and
identifying spatial patterns within a collection of buildings
Conceptual design calls for the design of multiple options
using gamma analysis. In addition to gamma analysis indica-
with a multitude of interdisciplinary criteria at a point where
tors, which are topological by nature, the parametric model
there is limited knowledge on the design problem. To achieve
allows the extraction of geometric spatial properties such as
novel solutions, design as exploration (Gero, 1994) is re-
area, perimeter, length, width, and spatial proportion. In
quired, and generative systems have the potential to offer
this way, the range of analytical metrics available from an
computational support for the population and evaluation of
analytical model forms a multicriteria problem that needs
solutions. The GMPDS developed by Alfaris and Merello
an evaluative model to determine performance. In related
(2008) describes a domain-independent computational
work, Heitor et al. (2004) combine shape grammar and space
framework to generate intelligent variations of an initial de-
syntax to formulate, generate, and evaluate designs, and Eloy
sign concept using multicriteria evaluation. The GMPDS de-
and Duarte (2011) use space syntax to provide a means of de-
composes a design problem into four mathematical models:
scribing and evaluating spatial properties to “increase the
synthesis, analysis, evaluation, and optimization. This modu-
likelihood of generating solutions that closely correspond to
larization allows a strategic approach, provides clarity to solv-
the user’s requirements” (Heitor et al., 2004, p. 494).
ing complex design problems, and is particularly suited to
parametric modeling. It allows a design problem to be decom-
posed into parametric submodels. The function of each model 2.3. Evaluative models
is summarized below.
Evaluation models determine how the analytical model is
interpreted to allow decision making. This is typically ap-
2.1. Synthesis models proached by deriving an objective function, commonly re-
Synthesis models abstract design intentions into a collection ferred to as a fitness function. The definition of this function
of design parameters, rules, or algorithms. Synthesis models is critical in guiding the search for a solution and with ill-de-
can be constructed via parametric or algorithmic descriptions. fined multicriteria design problems, often result in conflicting
Algorithmic models describe design through rules and algo- criteria. Evaluation can occur before or after the search for so-
rithms, while parametric models describe a design as a series lutions. Preevaluation requires the decision maker to aggre-
of relationships driven by parameters. gate multiple objectives into a single fitness function. These
Parametric models have been used to generate design var- can be weighted independently to preference certain objec-
iation (Park et al., 2005; Hernandez, 2006; Almusharaf & tives.
Mahjoub, 2010; Coorey, 2010). Because the design is com- While evaluation models are data driven and not reliant on
posed as a system of relationships, large populations of design geometric design software, the advantage for parametric inte-
solutions can be generated by varying the parameters. When gration lies in the feedback loop that can generate and opti-
generative parametric systems comprise analysis, evaluative, mize alternatives in the synthesis model.
and/or optimization models, a significant opportunity arises
with the capacity to improve design solutions (Littlefield, 2.4. Optimization models
2008; Sakamoto & Ferré, 2008; Gun, 2010; Hensel et al.,
2010; Peters, 2010). Optimization models enable the search for a solution. This in-
volves cycling through the design space established by the
synthesis model and comparing design instances to evalu-
2.2. Analysis models
ation criteria in order to identify more feasible solutions. Typ-
Analysis models determine characteristics from a design so- ically, optimization models can be classified as heuristic or
lution that are relevant to a specific discipline. Design is often numerical. Heuristic algorithms are nongradient methods
multidisciplinary and requires more than one analysis model. such as evolutionary algorithms and simulated annealing,
Analysis models are deterministic and should always produce while numerical techniques are gradient based such as New-
the same result. Depending on the amount of information re- ton’s method and steepest descent.
quired, there are low-fidelity or high-fidelity models. Low-fi- Heuristic algorithms are suited to complex problems asso-
delity models typically implement observation data, approx- ciated with the wicked problems typical in design. Develop-
imations, and heuristics, while high-fidelity models are ments in parametric modeling software are allowing heuristic
theoretical models that are either physics based or mathemat- optimization methods to be incorporated into the parametric
ically derived. workflow. Heuristic methods are appealing because they
Spatial analytics are typically low-fidelity indicators that can often find a good solution; however, optimality is not
can be readily integrated into a parametric model. Space syn- guaranteed and a different design solution may be found
tax analysis has been used to understand and predict the be- each time the optimization is run. In saying this, no optimiza-
havior of urban and interior spaces. Gamma analysis (Hillier tion technique is guaranteed to find the global optimum of a
& Hanson, 1984) interprets urban space syntax measures for nonlinear, nonconvex problem. Heuristic optimization is akin
permeability of interior spaces. Ostwald (2011) and Bafna to a designer’s method of solution-driven design.

https://doi.org/10.1017/S0890060414000225 Published online by Cambridge University Press

 Generative spatial performance design system 279

3. GENERATIVE SPATIAL PERFORMANCE niques in the development of a performative design system.

DESIGN SYSTEM (GSPDS) The system is set up to limit designer input to producing a
sketch design only. The exploration was conducted using
This section describes a series of spatial parametric logic
the parametric modeling software Grasshopper, and the var-
models that, when combined, enable a performance-based
ious models can be seen through the modularized parametric
design system, which we have named GSPDS. The following
definitions, visualized in Figure 2. Details of the four para-
sections will describe the overall conceptual framework and
metric models are described in the following sections.
components in more detail. The system consists of the follow-
ing four models and functions:
4.1. Synthesis model
1. Synthesis model
The role of the synthesis model is to abstract and capture the
† Construct an abstract parametric rig that represents a core spatial configuration of a building as a parametric model
spatial configuration. (Fig. 2) that has the capacity to generate alternatives. This is
† With a parametric rig, reconfigure and generate an al- implemented as two design procedures that are illustrated in
ternate solution. Figure 3.
The first part of the synthesis model interprets the design-
2. Analysis model
er’s initial sketch into a “spatial configuration rig” by taking a
† Calculate spatial analysis metrics through the para- series of room outlines (closed polylines), a set of relation-
metric rig to determine the design spatial properties. ships between them (lines), and an entry (point) as a set of in-
† Store parametric rig and spatial analytics into a spa- puts to produce a network graph consisting of nodes and re-
tial database. lationships. Each node represents a space and stores the
† Visualize analytical information in real time on de- spatial properties parametrically as an output.
sign geometry. The second part of the synthesis model remaps the spatial
configuration rig onto a randomized point cloud (which can
3. Evaluation model
be repopulated to produce different starting positions for
† Evaluate the configuration error of the generated so- the rig as a generative mechanism). This reconfigured net-
lution. work can then be optimized to produce alternate solutions.
This part randomizes the locations of the spaces while main-
4. Optimization model
taining their topological connection. This allows essentially
† Optimize the solution according to the evaluation. an infinite amount of variations for space allocation. At this
point, they maintain their topological relationships but their
geometric spacing is distorted, which is later resolved in the
We present the GSPDS (Fig. 1) as a solution to required ob-
optimization model described in Section 4.4.
jectives. The above system is demonstrated in a pilot study
that shows an implementation of the system through a fic-
tional case study. 4.2. Analysis model
This pilot study is focused on the development of a spatial
4. PILOT STUDY
performance design system and requires spatial analysis of
This pilot study demonstrates the capacity of the GSPDS to residential buildings. As discussed in Section 2, there are a
capture spatial information from a precedent, analyze spatial number of spatial indicators that can be derived from the syn-
properties, generate design variations, evaluate a solution ac- thesis model. These include space syntax gamma measures to
cording to certain criteria, and optimize for that criteria. The drive the spatial analytics model, including integration, con-
study is demonstrated through the design exploration of a sin- trol, and topological depth as well as geometric measures in-
gle residential building design. The goal of this study is to cluding area, length, width, and proportion (see Hillier &
demonstrate the appropriateness of parametric modeling tech- Hanson, 1984).

Fig. 1. A model for generative parametric spatial design exploration.

https://doi.org/10.1017/S0890060414000225 Published online by Cambridge University Press

 280 B.P. Coorey and J.R. Jupp

Fig. 2. A parametric definition for the generative spatial performance design system.

The analysis model calculates values directly from the syn- space. The sum of each receiving space is calculated; spaces
thesis model in the parametric modeling environment. This that have values . 1 indicate a “strong control,” and spaces
provides the benefit of visualizing real-time analysis of the , 1 indicate a “weak control” of their immediate neighbors.
original design geometry. The analytical data is then used as the input to the visualiza-
Topological depth provides a measure to determine the tion procedure, which maps the chosen analytic back onto the
depth distribution of spaces within the building. The measure individual rooms and colors them according to a black to white
of integration can be understood as how deep or shallow gradient field, where a value of 0 is represented as white and a
each space is, compared to the limit of how deep or shallow value of 1 is represented as black. An illustration of the differ-
they theoretically could be. This is also known as relative ent measures can be seen in Figure 4 showing the topological
asymmetry and is calculated by averaging the mean depth of depth, integration, control, and area overlaid onto one of the
the original space to every other space in the system. This pro- precedent residential building designs. Once calculated, the
vides a number between 0 and 1 for each space, with low val- analysis can be stored in a spatial database for later retrieval
ues indicating shallow spaces that tend to integrate the network, (see Coorey & Jupp, 2013).
while high values indicate deep spaces that are segregated. The
measure of control is a local indicator determining how impor-
4.3. Evaluation model
tant a space is in controlling the flow of movement around it.
Each space has a number “n” of immediate neighbors. Control The evaluation model is required to set a goal for design op-
is calculated by assigning a proportion 1/n to each neighboring timization. While in reality this should be a multicriteria

Fig. 3. An interpreted designer’s sketch (left) remapped onto a generative point grid (right). Spatial analytical measures. From top left
clockwise: topological depth, integration, control, and area.

https://doi.org/10.1017/S0890060414000225 Published online by Cambridge University Press

 Generative spatial performance design system 281

Fig. 4. Spatial analytical measures. From top left clockwise: topological depth, integration, control, and area.

evaluation problem, this case study demonstrates the logic by three bedrooms. With the different programs mapped, the de-
considering one evaluation criteria: spatial proximity. This signer can choose the configuration from existing buildings
model determines the validity of a generated solution by com- “A” or “B” to be mapped onto the hybrid network. This
paring the generated solution’s spatial proximity distances to will maintain the base topology of the chosen building; how-
the original specification in the spatial configuration rig. The ever, it will add in additional links to any bifurcated spatial
evaluation model can then determine the configuration error programs using the same method of hybridization described
for each spatial relationship and, through a simple aggrega- above. An equally weighted hybrid network between two
tion procedure, determine the average configuration error houses can be seen in Figure 5, which can then be used as
for the generated design. Optimization of this criterion will an evaluation source to provide novel typologies.
produce alternate design configurations that are reconfigured
with the exact same three-dimensional spatial proximity as
4.4. Optimization model
per the original design, which we will call “optimized” de-
signs. The pilot study concludes with an optimization module that at-
To explore the ability to generate novel topologies, a hy- tempts to take the generated configuration, with a configura-
bridization procedure was developed. Two existing designs tion error determined by the evaluation model, and optimize
were hybridized by connecting similar spaces from one the solution to reduce that error, creating an optimized design.
design to another to form a novel configuration to evaluate The optimization model achieves this through the parametric
against. The hybrid network is then used to map a new sketch plugin Kangaroo (Piker, 2013), which is a physics-based
design onto a target design through alignment of programma- “springs system.” Each spatial relationship is represented by
tic spaces. Each design can differ in the number and type of a “spring.” Each spring has a rest length, which is the desired
spaces. The hybrid network takes this into account by bifur- spatial proximity to its neighbor. The evaluation model deter-
cating or collapsing programs into each other. For example, mines the configuration error, and the spring system iteratively
if one building has one bedroom and the other has three, three reduces this error by expanding or contracting until the whole
connections will be drawn from the single bedroom to the system resolves, that is, the configuration error is minimized.

Fig. 5. The hybrid network.

https://doi.org/10.1017/S0890060414000225 Published online by Cambridge University Press

 282 B.P. Coorey and J.R. Jupp

Fig. 6. Generated spatial design variations.

Physics-based systems have been used to generate spatial enable the development of design solutions that have a
configurations (Arvin & House, 2002) and were explored more integrated spatial rationale. While the system was con-
for suitability in the parametric environment. The desired dis- figured to limit the designer’s input, it would be beneficial
tances are inputs from the original design or hybrid design, to allow designer input for influencing the optimization pro-
and the optimization model attempts to resolve those relation- cess. Finally, the spatial analytical data was able to be cap-
ships. Figure 6 shows an original network, with two alternate tured into a spatial database; however, the evaluation and op-
generated network solutions and the mapping of the original timization models were not developed to utilize this precedent
spaces onto the new network. knowledge. Future work will explore the determination of
evaluation rules as compared to the parameters captured
and controlled by a precedent database.
5. CONCLUSION
This paper presented the GSPDS, which integrates a series of
parametric logic models into a system for design exploration. REFERENCES
The main contribution of this paper is a demonstration of Alfaris, A., & Merello, R. (2008). The generative multi-performance design
workflow that allows the modularization of design logic system. Proc. Siliconþ Skin: Biological Processes and Computation
that can be modeled parametrically for integration into a de- (28th ACADIA), Minneapolis.
Almusharaf, M., & Mahjoub, E. (2010). A performance-based design ap-
sign decision support system. The system presented here is proach for early tall building form development. In CAADjCitiesjSustain-
configured for the specific and significant architectural prob- ability: 5th Int. Conf. Proc. Arab Society for Computer Aided Architec-
lem of residential spatial design. In the pilot study, the tural Design, pp. 39–50. Fez, Morocco: Arab Society for Computer
Aided Architectural Design.
demonstrated workflow highlights the computational power Arvin, S.A., & House, D.H. (2002). Modeling architectural design objectives
and flexibility of coupling generative parametric design in physically based space planning. Automation in Construction 11(2),
with precedent analysis for optimizing design solutions. 213–225.
Bafna, S. (1999). The morphology of early modernist residential plans: ge-
While this study validates the conceptual framework, the re- ometry and genotypical trends in Mies van der Rohe’s designs. Proc.
sultant optimization is a single criterion proximity optimiza- Space Syntax-II Int. Symposium, Brasilia.
tion and not necessarily an optimization of a spatial network Coorey, B. (2010). Scalability: parametric strategies from exoskeletons to the
city. In New Frontiers: Proc. 15th Conf. Computer Aided Architectural
according to multiple spatial analysis principles as per the Design Research in Asia (Dave, B., Li, A.I., Gu, N., & Park, H.-J.,
initial goal. Eds.), Hong Kong, April 7–10.
However, the system demonstrates the advantages of using Coorey, B., & Jupp, J. (2013). A schema for capturing and comparing para-
metric spatial data. Open Systems: Proc. 18th Conf. Computer Aided
parametric modeling for the development of a design system, Architectural Design Research in Asia (CAADRIA) (Stouffs, R., Janssen,
namely, the ability to modularize a design problem, achieve P.H.T., Roudavski, S., & Tunçer, B., Eds.), Singapore, May 15–17.
real-time analytical feedback, and work in a visual program- Eloy, S., & Duarte, J.P. (2011). Transformation grammar for housing reha-
bilitation. Nexus Network Journal 13(1), 49–71.
ming environment that provides instant feedback for design Gero, J.S. (1994). Towards a model of exploration in computer-aided design.
logic development. The limitations of such a system are in In Formal Design Methods for CAD (Gero, J.S., & Tyugu, E., Eds.), pp.
the capacity to compose a fitness value that integrates a series 315–336. Amsterdam: North-Holland.
Gun, O. (2010). Geometric gestures. In Elements of Parametric Design
of spatial criteria in a manner that allows the assessment of the (Woodbury, R., Ed.). London: Routledge.
solution’s spatial performance. Ongoing research is address- Heitor, T.V., Duarte, J.P., & Rafaela, M. (2004). Combining grammars and
ing this limitation. space syntax: formulating, generating and evaluating designs. Interna-
tional Journal of Architectural Computing 2(4), 492–515.
Future research requires further analysis of the generated Hensel, M., Menges, A., & Weinstock, M. (2010). Emergent Technologies
outcomes, especially in terms of the validity of the generated and Design. New York: Routledge.
spatial program. This may be enhanced by introducing addi- Hernandez, C. (2006). Thinking parametric design: introducing parametric
Gaudi. Design Studies 27(3), 309–324.
tional goals for the evaluation model that provide better con- Hillier, B., & Hanson, J. (1984). The Social Logic of Space. Cambridge:
trol over the variables of the optimization process. This will Cambridge University Press.

https://doi.org/10.1017/S0890060414000225 Published online by Cambridge University Press

 Generative spatial performance design system 283
Janssen, P.H.T. (2005). A design method and computational architecture for in parametric and computational design. His expertise in-
generating and evolving building designs. PhD Thesis, Hong Kong Poly-
technic University.
cludes parametric modeling, building information modeling,
Kruger, C., & Cross, N. (2006). Solution driven versus problem driven de- and architectural systems design. His research focuses on the
sign: strategies and outcomes, Design Studies 27(5), 527–548. integration of generative, analytical, and evaluative tech-
Littlefield, D. (2008). Space Craft: Developments in Architectural Comput-
ing. Marylebone: Royal Institute of British Architects.
niques for performative-based design.
Ostwald, M.J. (2011). A justified plan graph analysis of the early houses
(1975–1982) of Glenn Murcutt. Nexus Network Journal 13(3), 737–762.
Julie R. Jupp is an Associate Professor in the School of Built
Park, S.M., Elnimeiri, M., Sharpe, D.C., & Krawczyk, R.J. (2005). Tall Environment at the University of Technology, Sydney. She
Building Form Generation by Parametric Design Process. Chicago: Illi- specializes in design and construction technologies in the
nois Institute of Technology.
Peters, B. (2010). The new Elephant House. In Elements of Parametric De-
built environment. Her expertise covers areas related to build-
sign (Woodbury, R., Ed.). London: Routledge. ing information modeling, parametric modeling, and knowl-
Piker, Daniel. (2013). Kangaroo Plugin for Grasshopper [Computer Soft- edge and information management through life. Her research
ware]. Accessed at http://www.grasshopper3d.com/group/kangaroo
October 1, 2013.
thus encompasses key areas such as technological require-
Sakamoto, T., & Ferré, A. (2008). From Control to Design: Parametric/Al- ments of digital modeling, analysis, and simulation, informa-
gorithmic Architecture. Barcelona: Actar-D. tion technology and implementation issues, and governance
Schumacher, P. (Ed.). (2012). The Autopoiesis of Architecture: A New
Framework for Architecture, Vol. 2. Chichester: Wiley.
mechanisms relative to contractual, relational, data.

Benjamin P. Coorey is a Lecturer in the School of Architec-

ture at the University of Technology, Sydney. He specializes

https://doi.org/10.1017/S0890060414000225 Published online by Cambridge University Press