Design-Built

Rationalization Strategies
and Applications
Stylianos Dritsas

international journal of architectural computing issue 04, volume 10 575

 Design-Built Rationalization Strategies and
Applications
Stylianos Dritsas

Abstract
Rationalisation of architectural design is paramount to manufacturing
and its construction.This paper presents a methodology of
rationalisation of building envelope geometry.We discuss methods for
understanding and addressing design complexity; review two theoretical
models of rationalisation: the pre-rational and post-rational design
principles; illustrate their benefits and limitations and demonstrate their
meeting point proposing an integrated performance-oriented model for
analysis and design of building envelopes, using digital design techniques.

576
1. INTRODUCTION
Rationalisation is a widely used term across disciplines such the social
sciences, mathematics, engineering and architecture [1].We may categorize
the broad range of available definitions into either (a) retrospective
explanation of unconscious or ad-hoc behaviours via rational thinking or (b)
prospective application of rational thought in establishing causal
relationships. Rationalisation as an effort to systematize thinking and making
seems to encompass both a notion of analysis as well as synthesis.Within
the domain of architecture one may rationalise, or more precisely interpret,
historical precedents to attribute sociocultural, economic and technological
instigators for its inception and realisation. Rationalisation in the design-built
process amounts to the effort of creating a structured methodology for
translating a design from its conceptual phase to its end production while
addressing the arising technical implications.
In this paper we study the rationalisation of a design process in two
domains (a) strategies: we discuss the methods for structuring a design
process and offer a sketch for a unified pre + post rational framework and
(b) applications: we present an approach to address a particular family of
rationalisation problems aiming to control dimensional component variance
in contemporary digital design.The methodology used in both domains is
based on architectural design-computation and the primary interface to
design information is geometry. Both strategies and applications aim at
simplifying or compressing a design’s information density optimising its
geometry towards manufacturing and construction ends.

2. CONTEXT
Design rationalisation is not a new phenomenon but perhaps innate to the
architectural design-built condition: Classical design media such as sketches
and drawings express the desire of formalizing design and construction by
figurative documentation in addition to information conveyed by natural
language [1].What is new today is the medium, namely design-computation.
The emergence of a digital design paradigm occurred during the past few
decades expanded the scope of design opportunity while it increased its
complexity both formally and configurationally [2, 3]. Parametric models
allow us to organize the design process using dynamic geometric and
numerical relationships thus enhancing intuition by offering an opportunity
to observe design considerations interacting with one another. Our digital
models are not figurative or experiential representations of architecture but
they compute design.The thought process used to compose design models
is central in contemporary practice of design as they are the prime medium
of explicitly expressing design intent.Within this context we focus our study
on a computational version of design rationalisation as part of the process
seeking to address complexity in contemporary design and provide the
means of gaining insight, overview and control.

Design-Built Rationalization Strategies and Applications 577

 3. BACKGROUND
The principles of pre-rational and post-rational analysis and design are two
prominent methodologies of architectural rationalisation in contemporary
digital design.Their scope spans across the translation of concept design to
detail development and eventually construction documentation.They have
been extensively discussed by Shelden [4], Glymph [5],Whitehead [6],
Hesselgren [7], Ceccato [8].
Pre-rational principles are based on layered application of logical rules,
formalizing design as a process of interpolating composition of simple
figurative and functional primitives: lines and arcs, quadratic and developable
surfaces assembled using rule-based and associative modelling techniques.
The International Terminal Waterloo Station (1993) and Eden Project (2001)
by Nicholas Ghrimsaw and Partners; London City Hall (2002), 30 St.Mary
Axe (2004) by Foster and Partners; offer some indicative examples of pre-
rational processes where the design is strongly coupled with its generating
constructive geometry principles which provisionally embed manufacturing
and construction logic.
The post-rational perspective is rather an engineering systems approach
to architectural rationalisation in that it is heavily based on analysis and
optimisation processes using design-computation. An important assumption
made is that design is an input which is retrospectively approximated to a
rational tectonic model up to a notional design intent error tolerance.The
Guggenheim Museum of Bilbao (1997) and Walt Disney Concert Hall (2003)
by Ghery Partners;The Hungerburg Funicular (2006) and Heydar Aliyev
Center (2012) by Zaha Hadid Architects, offer examples of design post-
rationalisation where construction logic is fitted onto a rather more
intuitively arrived at design expression.
While the distinction between pre and post rationalisation is a helpful
formalization for the design-built process it is also artificial.We cannot
assume that every post-rational process is preceded by an utterly
unstructured design thinking process but merely one that is less explicit. In
similar fashion we may picture a circumstance where pre-rational approach
is employed as a form of reverse-engineering of an in-mind design concept
which is also merely less explicit. Composite notions of rationalisation have
been studied [9,10] as both means of resolving this dichotomy and mapping
more realistically design rationalisation in practice which seldom employs
one method or another. Insofar as extending the discussion we approach
design-built rationalisation from a design rather than analytical perspective.
Our inquiry is primarily focused on questions pertaining when a particular
method seems appropriate and what are the implications in terms of
process and product complexity. Characterisation may be seen as an
attribution process identifying the load-balance between pre and post
rational principles employed.

578 Stylianos Dritsas

 4. FRAMEWORK
Our framework identifies three design rationalisation components: (a)
Description, (b) Analysis and (c) Optimisation.Those map onto the phases
of generation, evaluation and selection within a design process.They are
embodied within computer aided design modelling using for instance
parametric or generative techniques; performance metric extraction using
methods such as quantity, structural, and environmental computations; and
finally automated design candidate filtering using artificial intelligence
heuristics, numerical solvers and search techniques (Figure 1).

Figure 1: Design rationalisation from

general concepts to methods of
implementation.

The framework is affine to classical models by Asimow [11] and Archer [12]
however the reorder of events with design description, or synthesis,
preceding analysis is used to allow for less explicit and exploratory design
processes which may be formalized subsequently. Most importantly for
contemporary rationalisation within digital media description, analysis and
optimisation which are traditionally human operations, are assumed being
embodied by computation or a composition between thereof. Finally, we
allow these components to nest in one another in order to achieve
complex design graphs.
Indicatively, we may employ the framework in a wide range of
architectural and engineering applications such as tall building design [7, 13]
where cladding unit configuration is computed using procedurally generated
geometry while performance metrics such as wall cavity and floor areas are
optimized via divide and conquer strategies. For the same building [14]
present a structural engineering study where description is based on finite
element modelling, structural analysis is used to evaluate the building
performance and stochastic optimisation is employed to improve its static
behaviour and reduce material use.

4.1. Design description

The concept of design description encompasses both visual and non-visual
information which yields a design state. It is foundational for rationalisation
as no analysis and optimisation can occur without it. Description aims at

Design-Built Rationalization Strategies and Applications 579

 establishing explicit relationships of generation and by implication capture
the design intent. In architectural design, description is embodied in
representations; a rather more computable version thereof is geometry.
However, for engineering purposes the forces conveyed in material while
may be visualized they are inherently non-visual design information.
While design description may provision for construction logic by
embedding rules and constraints this is neither a prerequisite nor something
to avoid. It is exactly this allowance that enables the same framework to
express both pre and post-rational workflows.The more rules employed
during the descriptive phase the more a particular implementation becomes
pre-rationally conditioned and in effect proactively reduces its input and
output domains. However we might as well establish a generative process
that contains no prior bias and defers evaluation to metric analysis. In effect
we may allow ourselves the freedom of a larger domain for design
opportunity and in a consecutive step attempt to enforce reason from
within or from outside the system’s description.

4.2. Design analysis

Metric analysis expresses design evaluation both quantitatively and
qualitatively. Analysis looks at one or more performance criteria such as
human, building and environmental factors.The notion of metric analysis
acquires a form of design performance introspection within digital design as
we inquire the descriptive model to report those metrics procedurally.
Extraction of direct quantitative metrics such as lengths, areas and volumes
is trivial given the analytical nature of computational geometry. However,
complex or qualitative indicators require design and implementation of
specialized analysis processes hence, design analysis becomes a domain of its
own.
Pre-rational design contains integrated evaluation logic and it employs
exactly those embedded criteria for generative purposes. However, it is not
always possible to evaluate a design directly from its driving parameters; we
still need additional external evaluation. Most importantly it is also beneficial
to employ unbiased assessment processes disaggregated from the
descriptive process as we may for instance catch logical mistakes and false
assumptions. Post-rational design by definition places most of the work in
this segment as it does not assume control over the generative description
process.
While this segment is attributed as analysis it is essential to note that
the selection, compilation and creation of design performance indicators
from first principles or empirical studies is an act of art and design in as
much a task of scientific analysis. Our case study demonstrates that even
simple visual assessment criteria are very difficult to objectively classify and
depend of personal or project specific goals.

580 Stylianos Dritsas

4.3. Design optimisation
The notion of design optimisation is arguably the logical end to the
availability of digital generative and analytical assessment systems. Its goal is
to establish feedback loops and embody the notion of computer mediated
decision making within design-computation.The scope of design
optimisation may be localized and focus on narrow bands of the system or
span across driving parameters and composite output metric indices.
Optimisation may be as trivial as in generating a limited number of design
options, characterizing them according to certain design criteria and
selecting the most prominent, or it may be as complex as incorporating
quasi or fully automated iterative search algorithms sampling vast domains
before converging to a few design candidates.
Pre-rational design integrates optimisation logic to the extent that
generation rules strategically target specific neighbourhoods of tested and
proved solutions. Associative modelling using ruled-surfaces for instance
proactively prunes a range of design possibilities exhibiting curvature
characteristics that may not be as easily manufactured. However we may
still use disjoint analysis and optimisation to break out of the tight boxes of
pre-rational systems and explore the effect of complex design interactions
that may not be easily modelled parametrically. Optimisation is absolutely
necessary for post-rational systems as their descriptive domains are
typically boarder and retro fitting directly implies an embedded notion of
search.

5. APPLICATION
We will proceed with examining an application of the model via a case
study of a computer aided design rationalisation of a complex building
envelope.The particular design process is derived from an on-going project
that cannot be disclosed at the point of publication upon client’s request.
However the methods used are applicable to a wide range of free-form
building envelope designs.

5.1. Description
The design intent is to create a curved wall in plan, and potentially in
section, that may be simply generated and manufactured.We employ a pre-
rational strategy of piece-wise tangent surfaces of revolution (Figure 2), also
known as translational surfaces [5, 15], to define the cladding setting-out
geometry.The result envelope belongs to a family of surfaces exhibiting
certain desirable characteristics: (a) components are circular arcs which are
simple to manufacture and geometrically offset without deterioration, (b)
surface patches along the primary surface directions are planar quads thus
suitable for curtain wall applications, and (c) the visual constraint of tangent
continuity or apparent smoothness is ensured by definition.

Design-Built Rationalization Strategies and Applications 581

  Figure 2:
Rational
envelope
definition as a
function of
continuous and
discontinuous
radius graphs.

In detail, we establish relationships between consecutive arcs such that at

the transition points they share radial direction but not magnitude. A planar
sectional profile may be then swept along the base path to generate the
design surface. In addition, the section may be arbitrarily complex without
violating the desired geometric properties.
The particular strategy has been presented in the past [5, 6, 15, 16]. Our
implementation is a generalization of this technique using a smooth curve
graph, or law curve, to produce continuously varying mesh geometries.
Discontinuous graphs produce traditional piece-wise revolute surfaces while
smooth graphs generate surfaces of pseudo-continuous curvature. Using the
same principle we may also parameterize arc lengths, by additional graphs,
to increase resolution at regions of higher curvature. Alternative pre-
rational techniques using displacement graphs [17] can be used to mitigate
numerical and formal shortcoming of radius-driven geometries while [18-
20] offer in depth studies of planar quadrilateral meshes and developable
surfaces from a post-rational perspective.

5.2. Analysis
Setting-out geometry typically spans large regions of a building’s envelope
and it requires subdivision into smaller components which can be
manufactured using standard industrial methods. Each arc is thus divided by
a constant chord length into equal segments to pre-constraint the number
of dimensionally unique parts (Figure 3). Control of the number of types is a
logistical consideration as volumes of varying parts increase manufacturing
and assembly time/cost and encumber the process of building information
management on and off site.

582 Stylianos Dritsas


Figure 3:
Geometric
principle and
expressions of
approximation
errors.

Parts of large arc radius may be approximated by linear segments without

visual distortion such that manufacturing costs of bending may be culled.
There exists no objective threshold determining when the simplification is
generally acceptable as the decision is bound to visual preferences,
fabrication constraints and cost considerations. However, in principle
shallow arcs of sagittal lengths measuring a few millimetres may be
linearized with nominal errors (Figure 4).

Figure 4: Geometric approximation

errors due to linearisation.

We further observe that a chain of linear elements along an arc suggests

that joints between segments require mitring at the bisecting angles, thus
for each segment there is potentially a unique cutting pattern. Reducing
manufacturing complexity implies averaging or squaring those angles such

Design-Built Rationalization Strategies and Applications 583

  Figure 5:
Interactive
visualisation of
envelope
performance
metrics.

that processing cost is reduced.The approximation induces a visual artefact

known as the splay error where small gaps occur between consecutive
segments (Figure 5). Again, there is no general consensus for when splay
errors may be ignored or indeed introduces significant visual nuance. It is
however common to ignore splay errors when they remain below typical
architectural detailing and manufacturing tolerances. For example a few
millimetres are within an acceptable range since even planar walls exhibit
this dimensional range of detailing gaps at the joints.

5.3. Optimisation
The pre-rational system presented is geometrically predictable in indicating
coarsely the number of unique elements required.We deduce that the
number of unique segments (n) in plan times the number of unique
segments in section (m) yields an upper bound of (n • m) total types. A
composite cylindrical envelope independent of the number of patches
requires one unit type, simple conic surfaces result to a number of types
equal to the number of sectional segments and composites thereof require
at most the product of patches times the number of sectional segments
(Figure 6).
The aim of design optimisation in this case study is to answer the
following design inquiry: Can we further reduce the total number of unique
types and simplify the design and construction? This turns out to be a
common problem, or a family of problems, where formal complexity results
to dimensionally varying parts. Relevant previous work includes [9]
attempting to rationalize the number of unique linear elements of a spatial
frame; [21] study a similar problem from an engineering perspective such
that by modifying the overall form and force distribution achieve using a

584 Stylianos Dritsas


Figure 6:
Dimensional
variation for
different surface
characteristics.

single cross sectional element in a bridge design; and [22, 23] identify
typologies of curvature in fitting free-form envelopes with simple
developable surfaces.
Our study was performed at a later phase of design development where
there was no opportunity to modify the overall building form. Instead the
rationale is based on the insight that the design contains variably sized parts
which are variably spaced due to segmentation. Units and gaps between
thereof lie at different building scales which hint of an opportunity to
leverage part variance concentrating it into detailing voids instead of physical
members. In particular, we focus at the transom carriers and attempt to
reduce their length variance while allowing it to fluctuate in between.

Description
Variation control implies a notion of similarity classification where parts of
certain closeness are averaged and refitted in the original design as long as
they do not alter it dramatically. Our case study is a one-dimensional
optimisation of linear member sizes thus we do not need to examine
geometric affinity which is far more challenging. Identifying a fixed number
of unique types which best approximate variation among a set of samples
bound by an error metric is a data clustering problem. In particular it is an
NP-hard problem addressed typically via numerical computation [24, 25].

Design-Built Rationalization Strategies and Applications 585

  Figure 7:
Understanding
component
variations using
frequency
histograms at
5mm intervals.

Analysis
In order to understand the idea of variance concretely and measure the
performance of our optimisation we need a benchmark strategy.We thus
turn to statistics and use frequency histograms to record part size
distributions and later compare clustering strategies. Cladding element sizes
span the domain of real numbers which for complex envelopes every
instance may be potentially unique.This is however a purely theoretical limit
as manufacturing accuracies and construction tolerances allow us to define
a discrete lower bound. For simplicity we define a pseudo-theoretical
distribution increment of one millimetre, assuming this is the absolutely very
least one may round unit sizes, and effectively count the number of non-
empty histogram bins (Figure 7).
Generally we wish to perform better so we define a parametric
increment bin-size of a few millimetres, round elements downwards to their
nearest cluster, measure the average and maximum errors and count the
number of non-empty or unique unit types.The maximum rounding error is
bound by the cluster’s size, while the average error is approximately half of
that. For 5mm bins we are certain that the worst rounding error will be
strictly bellow 5mm while on average the error will be approximately
2.5mm.The process is computationally trivial as it may be computed using a
standard spread sheet.Yet expectedly the results are suboptimal as by
definition the histogram approach disregards all prior density characteristics
of the distribution. By implication this means that we have no control over
the end-effect of rationalisation as for instance gaps among units may

586 Stylianos Dritsas


Figure 8:
Constrained k-
means algorithm
process
diagram.

abruptly vary causing visual nuance.What we thus need is a strategy that

may allow us to control how rounding errors present themselves in the final
design expression.

Optimisation
For selecting better cluster pivots we developed a constrained k-means
algorithm. K-means clustering is a statistical method for minimising
intercluster variance [26, 27].The algorithm selects a number of cluster
partitions at random, associates each element with its closest cluster and
repositions the partition centres at the mean of each cluster’s elements
(Figure 8).The algorithm converges after a few iterations but it is sensitive
towards the initial seed-cluster locations and requires repeated attempts.
Our application (Figure 9) uses a modified k-means method as we
cannot allow units to increase dimensions upwards to their nearest cluster;
elongated units may result to geometric clashes. Instead, we pin a cluster to
the lowest bound of the range and skew the final cluster centres to their
respective lowest bound.This subtle modification challenges the distance
metric of the original algorithm. Intuitively it doubles its error as it coerces
nodes from a clusters’ mid-point half way to its lower bound.We describe
our method as two-times k-means strategy. Even so, the algorithm still
performs on average 11% and 35% better than the benchmark in maximum
and average errors. Moreover, k-means’ intercluster error minimization
ensures that gaps will appear smooth within typology groups and only
suddenly jump between clusters.
However, while k-means improves intercluster error it cannot address
the problem of minimising the maximum error which is often visually more
pronounced. A qualitative interpretation suggests that minimising the
maximum error expresses the desire to control the largest visible gap
change across an envelope while minimising the average error implies that
small transitions need to be smooth.To minimise the maximum intercluster

Design-Built Rationalization Strategies and Applications 587

  Figure 9:
Interactive
visualization. Left:
Theoretical
unclustered
version, Right:
Aggressive
clustering
produces visual
artefacts which
are marked as
black units.

error we employ the k-tMM strategy [28, 29].The algorithm splits an initial
cluster containing every envelope unit into sub-clusters while maintaining
the min/max criterion (Figure 10). Eventually, each unit is rounded to the
lowest bound of its cluster and the algorithm deterministically converges
within two times the optimal [28]. Our implementation achieves on average
16% and 50% better than the benchmark strategy in terms of maximum and
average errors, respectively.

 Figure 10:
Min/max
clustering
algorithm process
diagram.

Overall, we achieve improved results with both min/average and min/max

compared to our initial benchmark (Figure 11). In terms of complexity k-tMM
is simpler to implement, it runs faster, it is much more predictable and yields
better results for large cluster sizes in both maximum and average errors
perhaps due to k-means modifications which skewed what should have been a
symmetric metric. Our parametric model produces the original design
geometry as well as its end optimisation results at interactive speeds.Thus it
is eases the process of design exploration and relays overall form refinement
to the design actor by merging description, analysis and optimisation.

588 Stylianos Dritsas


Figure 11:
Maximum and
average errors
of alternative
clustering
strategies.

6. CONCLUSION
We presented a design-built rationalisation framework alongside a case
study demonstrating how we may integrate pre and post rational principles
to address the problem of high levels of part variation in formally complex
designs.We conclude by offering observations on the thinking process of
rationalisation systems within digital media.
The question we propose as per when or how one may select a pre or
post rational approach to rationalisation cannot be seen as an either/or
proposition. Instead we may use a set of indicators to assist and guide the
decision making process. Pre-rational principles offer a visual paradigm for
expressing design information via primarily associative geometric modelling
technique. Component/configuration organizational principles allow pre-
rational modelling to achieve complex design compositions.We attribute
this notion as systemic compactness whereby simple operations yield
expressively complex results. Interestingly this behaviour is achieved while
ensuring certain upper bounds in terms of complexity expectation. In our
case study we can predict coarsely the expected part variance already from
input parameters generating the design surface. Sensitivity-wise pre-rational
models behave also fairly linear towards small parametric variations.We
attribute this notion as systemic predictability, which is the degree we can
foresee how generative actions effect results.
The rules captured by pre-rational systems are usually simple and may
be replicated with limited or no computation using traditional media.
Revolute building envelopes may be inefficiently drawn by hand because
intrinsically they involve merely affine geometric transformations. Pre-
rational methods in this respect are somewhat calculation averse. Simplicity
at the descriptive level though maps naturally to conventional design

Design-Built Rationalization Strategies and Applications 589

 processes and allows design teams of various degrees of comfort with
computation to engage and contribute. For instance cladding setting-out
geometry is used for space planning, structural steel layout, quantity surveys
etc.The ability for multiple parties to operate, evaluate and verify the design
at the same level of information complexity defines a notion systemic
reliability: it is possible to manually reproduce, gracefully decompose, easily
verify and communicate the products of a system.
While descriptive elegance and simplicity, intuitive capability for
provision and insight, as well as procedural and output dependability are
certainly appealing properties of pre-rational systems there are also
expressive and technical shortcomings.The computation averse and visual
oriented attributes of pre-rational systems limit expression and problem
solving from exploiting powerful capabilities within digital media. Simulated
form-finding and discrete computational geometry for instance do not fit
well in these models. Meanwhile, the deterministic directed acyclic graphs
constructed using pre-rational methods tend to fall short answering fairly
simply stated design inquiries which typically involve implicit or inverse
relationship interrogation [30].
Our post-rational sub-system within the optimisation study also exhibits
similar systemic properties. Descriptively we formalized our problem by
transposing the inquiry into the domain of statistics and data clustering.The
process of transforming geometric information into numerical, operating on
it using methods from domains of knowledge which are not traditionally
appear relevant to architectural design and eventually translate information
back to a formal domain in order to assess the implication is fascinating.
Moreover, in order to understand complexity we may refer to concepts
from information theory such as entropy, asymptotic analysis, NP-hardness
etc.These are very precise and powerful tools compared to observational
assessments of worst case scenarios. In addition, implementing optimization
algorithms, understanding and communicating their results and operating on
numerically optimized models are processes which are fairly foreign to
architectural design and we will need long term exposure and experience
before we can naturally integrate in practice.
As contemporary design praxis in architecture, engineering and the
construction industry is rapidly transforming by the introduction of design
information technologies we find ourselves creating highly sophisticated
building information models which we modestly use to leverage tasks such
as document control, trades coordination and the assessment of basic
building metrics such as quantity surveys and clash detection. Our study
suggests for an expanded paradigm of this process where we may no longer
rely only on prefabricated systems of smart parametric objects but embed
functional components of description, analysis and optimisation to take
advantage of the power within design information and address the
challenging issues arising during the translation of concept design to
construction.

590 Stylianos Dritsas

 Design computation has flourished over the past decades with main
focus of research placed on exploratory design studies and the development
of primarily descriptive systems. Metric analysis and design optimisation are
still relatively nascent at least within the realm of architecture. Analytical
methods for engineering such as the finite elements, the fluid dynamics and
a range of environmental evaluation techniques are far more developed. In
architecture, there is a great opportunity to invent design metrics capturing
qualitative aspects of design such as spatial and visual performance for
instance.This requires a certain mental leap from the notion of
representations to design information we can analyse and compute with.We
have been using photorealistic visualization technologies for the longest
time for example but we rarely appreciate those tools for their analytical
and predictive potential for they employ equally highly sophisticated
probabilistic light simulation techniques. Design optimisation holds also great
research potential not merely for resolving intractable design relationships
which we cannot possibly model by associative principles but also to assist
us identify unexpected and surprising design possibilities lying beyond
human intuition. In summary, the goal of the description-analysis-
optimisation model is not to remove and replace the design actor by
computation but to expand the information processing capability and
increase design information density.
With respect to the adoption of digital fabrication technologies for part
manufacturing and potentially the construction, we note that rationalization
is not contrary but complementary to the process.The downstream
demand for highly precise manufacturing information to drive numerically
controlled machinery will raise the expectations for procedural generation
and processing capability of upstream design information. Information
analysis, filtering, classification as well as concepts such as simplification and
compression will have to readily change orientation from catering
capabilities and limitation of conventional industrial mass production to
large scale bespoke manufacture.

ACKNOWLEDGEMENTS
The author would like to thank the International Design Centre at the
Singapore University of Technology and Design for supporting this research.

REFERENCES
1. Fischer,T., Geometry Rationalisation for Non-Standard Architecture, Architecture
Science, 2012, 5, 25-47.
2. Kolarevic, B., Architecture in the Digital Age: Design and Manufacturing,Taylor and
Francis Group, New York and London, 2003.
3. Kolarevic, B. and Klinger, K., Manufacturing Material Effects, Rethinking Design and
Making in Architecture, Routledge, New York, 2008.
4. Shelden, R. D., Digital surface representation and the constructability of Gehry’s
architecture, PhD Thesis, Massachusetts Institute of Technology, 2002.

Design-Built Rationalization Strategies and Applications 591

 5. Glymph J., Shelden, D., Ceccato, C., Mussel, J. and Schober, H., A parametric
strategy for free-form glass structures using quadrilateral planar facets,
Automation in Construction, 2004, 13, 187-202.
6. Whitehead, H., Laws of form, in: Kolarevic, B., ed., Architecture in the Digital Age:
Design and Manufacturing,Taylor and Francis Group, New York and London, 2003.
7. Hesselgren, L, Charitou, R. and Dritsas, S.,The Bishopsgate Tower case study,
International Journal of Architectural Computing, 2007, 1(5), 61-81.
8. Ceccato, C., Hesselgren, L., Pauly, M., Pottman, H., and Wallner, J., Advances in
Architectural Geometry, Springer-Verlag,Vienna, 2010.
9. Fisher,T., Rationalising Bubble Trusses for Batch Production, Automation in
Construction, 2007, 16, 45-53.
10. Loukissas,Y., Rulebuilding. Exploring Design Worlds through End-User
Programming, Master Thesis, Department of Architecture, Massachusetts Institute
of Technology, Cambridge, 2003.
11. Asimow, M., Introduction to Design, Prentice-Hall, Inc, New Jersey, 1962.
12. Archer, L. B.,The Structure of the Design Process. Design Methods in Architecture.
London, Lund Humphries, 1969.
13. Dritsas, S., Charitou, R. and Hesselgren, L., Computational Methods on Tall
Buildings - The Bishopsgate Tower, eCAADe Conference Proceedings,Volos, Greece,
778-785, 2006.
14. Luebkeman, C. and Shea, K., CDO: Computational Design + Optimisation in
Building Practice, The ARUP Journal, 2005, 17-21.
15. Schlaich, J. and Schober, H., Design Principles of Glass Roofs, Lightweight Structures
in Civil Engineering, Proceedings of the international symposium,Warsaw, Poland,
2002.
16. Gun,Y. O. and Wallin, J. N., Composing the Bits of Surfaces in Architectural
Practice, Methodologies and codes for generation, rationalisation, and analysis of
non-standard geometries. ACADIA Conference Proceedings. Halifax, Canada,
2007, 66-79.
17. Dritsas, S. and Becker, M., Research & Design in Shifting from Analogue to Digital.
ACADIA Conference Proceedings. Halifax, Canada, 2007, 56-65.
18. Leopoldseder, S. and Pottmann, H. Approximation of Developable Surfaces with Cone
Spline Surfaces, Institut fur Geometrie Technische Universitat Wien, 2003.
19. Pottmann, H. and Wallner, J., The focal geometry of circular and conical meshes,
Technische Universitat Wien,Vienna, 2006.
20. Eigensatz, M., Kilian, M., Schiftner, A., Mitra, J. N., Pottmann, H. and Pauly, M.,
Paneling Architectural Freeform Surfaces, ACM Transactions on Graphics, SIGGRAPH,
Los Angeles, USA, 2010, 29 (4), 45.
21. Kaijima, S. and Michalatos, P., Discretization of Continuous Surfaces as a Design
Concern, eCAADe Conference Proceedings, Frankfurt, Germany, 2007, 901-908.
22. Kaijima, S. and Michalatos, P., Computational Design Consultancy, eCAADe
Conference Proceedings, Antwerpen, Belgium, 2008, 311-318.
23. Flory, S. and Pottmann, H., Ruled Surfaces for Rationalisation and Design in
Architecture, ACADIA Conference Proceedings, New York, USA, 2010, 103-109.
24. Jain, A. K., Murty, M. N. and Flynn, P. J., Data clustering: A review. ACM Computing
Surveys, 1999, 31(3), 264–323.
25. Bern, M. and Eppstein, D., Approximation Algorithms for Geometric Problems, PWS
Publishing, Boston, USA, 1997.
26. MacQueen, J. B., Some Methods for Classification and Analysis of Multivariate

592 Stylianos Dritsas

 Observations. Fifth Symposium on Math, Statistics and Probability, Berkeley, USA,
1967, 281-297.
27. Wagstaff, K., Cardie, C., Rogers, S. and Schroedl, S., Constrained K-means
Clustering with Background Knowledge, Proceedings of the 18th International
Conference on Machine Learning, 2001, 577-584.
28. Gonzalez,T. F., Clustering to Minimize the Maximum Intercluster Distance,
Theoretical Computer Science, 1985, 38, 2-3, 293–306.
29. Xiang, Z., Color Image Quantization by Minimizing the Maximum Intercluster
Distance, ACM Transactions in Graphics, 1997, 16(3), 260-276.
30. Kilian, A. Design innovation through constraint modeling, International Journal of
Architectural Computing, 2006, 4(1), 87-105.

Stylianos Dritsas
Singapore University of Technology and Design
Architecture and Sustainable Design
L2R10, 20 Dover Drive, Singapore 138682
S Dritsas, [email protected]

Design-Built Rationalization Strategies and Applications 593