![A typical floor plan [8].](https://www.researchgate.net/profile/Patrick-Janssen-3/publication/261289549/figure/fig2/AS:1079789985247234@1634453402323/A-typical-floor-plan-8_Q320.jpg)
Content uploaded by Patrick Janssen
Author content
All content in this area was uploaded by Patrick Janssen on Oct 17, 2021
Content may be subject to copyright.
555Simulation, Prediction, and Evaluation - Volume 1 - eCAADe 30 |
INTRODUCTION
Visual Dataow Modelling (VDM) (Janssen and Chen
2011) has becoming increasingly popular within the
design community, as it can accelerate the iterative
design process, thereby allowing larger numbers
of design possibilities to be explored. Modelling
in a VDM system consists of creating dataow net-
works using nodes and links, where nodes can be
thought of as functions that perform actions, and
links connect the output of one function to the in-
put of another function. VDM is now also becoming
an important tool in performance-based design ap-
proaches, since it may potentially enable designers
to explore and rene design possibilities through
an iterative process of parametric variation cou-
pled with performance simulation (Shea et. al. 2005,
Coenders 2007, Lagios et. al 2010, Toth et. al. 2011,
Janssen et. al. 2011).
In order for the process of iterative renement
to be eective, it is critical to set appropriate trade-
Iterative Renement Through Simulation
Exploring trade-os between speed and accuracy
Patrick Janssen1, Vignesh Kaushik2
National University of Singapore
1patrick@janssen.name, 2vigneshkaushik@gmail.com
Abstract. Performance-based design approaches typically use iterative simulation as
a way of exploring design variants. For such approaches, the speed of execution of the
simulations is critical to enabling a uid and interactive design process. This research
proposes an iterative simulation design method where simulations are congured to run
in two modes: in fast mode, simulations produce less accurate results but due to their
speed can be applied successfully within an iterative renement process; in slow mode,
the simulations produce more accurate results that can be used to verify the performance
improvements achieved using the iterative renement process. A case study is presented
where the proposed method is used to explore performance improvements related to levels
of incident illuminance and incident irradiance on windows.
Keywords. Iterative; design; renement; simulation; Radiance.
os between simulation speed and simulation ac-
curacy. In order for simulations to be used uidly
and interactively, execution time must be kept to a
minimum. However, the accuracy of the simulation
is often inversely related to the speed of execution.
Fast simulations produce low-accuracy results, while
slow simulations produce high-accuracy results.
This paper proposes an iterative simulation
design method that overcomes this problem by
calibrating simulations to run both in a fast and
less accurate mode and in a slow and more accu-
rate mode. The fast mode simulations are used to
enable designers to apply iterative renement in a
uid and interactive manner, while the slow mode
simulations are used to verify the performance im-
provements achieved using the iterative renement
process.
In order to demonstrate the proposed method, a
case-study experiment has been conducted, where
556 | eCAADe 30 - Volume 1 - Simulation, Prediction, and Evaluation
the method is used to explore design variants for a
large residential project consisting of over a thou-
sand units. Design variants are evaluated based on
visible daylight and radiant heat (including sunlight)
incident on the surface of the windows of residential
units.
SIMULATION NODES
Visible daylight is measured as illuminance, which is
the visible light incident on the surface, and is meas-
ured in Lumens/m2 or Lux. Radiant heat is measured
as irradiance, which is the electromagnetic radiation
incident on the surface, measured in Watts/m2. Both
illuminance and irradiance are calculated using the
simulation program Radiance [1].
Radiance simulations
Radiance is a collection of programs that perform a
variety of related tasks. The main input le for Radi-
ance is the RAD le that describes the scene to be
simulated. Given a RAD le, the rst step is to con-
vert this into a dierent le format called an octree,
using a program called oconv. Using this octree le
as input, the user can specify sensor points in the
model and then use the rtrace program to measure
illuminance or irradiance at these points.
When the octree is generated using the oconv
program, the radiance description of the sky dome
can be included. Dierent skies need to be gener-
ated for the illuminance and irradiance simulations.
For the illuminance simulation, standard CIE over-
cast sky is required, as this is the worst case scenario
for calculating daylighting. This sky is not aected
by the position of the sun, and as a result it is time
independent. The illuminance that is calculated will
then represent the actual Lux value for the worst
case scenario.
For the irradiance simulation, rather than fo-
cusing on the worst case, a cumulative approach is
needed whereby the irradiance incident on a par-
ticular point throughout the year is added up. In ad-
dition, irradiance is of course time dependent as it is
aected by the position of the sun. One option for
calculating cumulative irradiance would be to gen-
erate skies for multiple points in time and then to
run multiple simulations. However, a more ecient
approach is to create a cumulative annual sky from
a climate le (Robinson and Stone 2004). The irradi-
ance results from a single simulation run using such
a cumulative sky would then represent cumulative
irradiance for the whole year.
For generating the standard CIE overcast sky
for the illuminance simulations, Radiance includes
a program called gensky[2]. For generating the cu-
mulative annual sky, a program called GenCumula-
tiveSky[3] is used. This program produces cumulative
annual skies in Radiance format from EnergyPlus
weather les [4]. The sky is discretized into 145
patches using a method devised by Tregenza (1987)
and the Perez luminance/radiance distribution
model (Perez et. al. 1993) is used to determine the
radiance of each patch, according to the information
from the climate le. Figure 1 shows both the stand-
ard CIE overcast sky and the cumulative annual sky
used in this research.
Figure 1
Falsecolor image of the stand-
ard CIE overcast sky generated
by gensky (left) and the cumu-
lative annual sky generated
by GenCumulativeSky (right)
using the EnergyPlus weather
le for Singapore.
557Simulation, Prediction, and Evaluation - Volume 1 - eCAADe 30 |
Radiance VDM nodes
In order to support a user-friendly integration of
Radiance into the design workow, VDM nodes
were created for an advanced procedural modelling
system called SideFX Houdini [5]. These nodes link
Houdini with the various Radiance programs and
the GenCumulativeSkyprogram. For more informa-
tion on the development of these nodes, see Jans-
sen et. al. (2011). For this research, the nodes were
further developed to allow users to create a sky via
three methods: by using the gensky program, by us-
ing the GenCumulativeSky program, or by loading a
sky description le.
The main simulation node executing oconv and
rtrace has two inputs: one for the model geometry
and one for sensor grids. The rst input is for in-
putting the geometry. The node will translate each
Houdini polygon to a Radiance primitive of type pol-
ygon. The polygons expected to have custom attrib-
utes to dene the material, and the node will extract
the values of these attributes when generating the
Radiance input le.
The second input of the Radiance node is for
inputting sensor grids. When the rtrace simulation
is run, the simulation results will be copied to the
sensor points within Houdini as attributes. This then
means that the results from an rtrace simulation can
be graphically displayed inside Houdini, using col-
oured surfaces.
Simulation parameters
In terms of speed, rtrace will tend to take signicant-
ly more time to execute than the other programs
such as oconv, gensky, and GenCumulativeSky. The
settings for the rtrace parameters are therefore criti-
cal when it comes to the trade-o between speed
and accuracy.
The Houdini node provides parameters for spec-
ifying certain key rtrace parameters relating to ambi-
ent lighting, namely [6]:
• Ambient bounces (ab): the maximum number
of diuse bounces computed by the indirect
calculation. A value of zero implies no indirect
calculation
• Ambient accuracy (aa): the approximate error
from indirect illuminance interpolation. A value
of zero implies no interpolation
• Ambient resolution (ar): The maximum density
of ambient values used in interpolation. Errors
will start to increase on surfaces spaced closer
than the scene size divided by the ambient
resolution
• Ambient divisions (ad): The error in the Monte
Carlo calculation of indirect illuminance will be
inversely proportional to the square root of this
number. A value of zero implies no indirect cal-
culation
• Ambient super-samples (as): Super-samples
are applied only to the ambient divisions which
show a signicant change
A detailed explanation of these parameters is be-
yond the scope of this paper (see the Radiance
documentation for more information). However, it
should be noted that perhaps the most important
parameter is the ambient bounces. The number of
bounces may vary from 0 to 8, with higher num-
ber of bounces producing more accurate results
but also in much higher computation times. Since
the sky is modelled as ‘glow’, it will only take part
in Radiance’s indirect lighting calculation, which is
only performed when ambient bounces is set to 1
or more. Since the only light is coming from the sky,
then a value of 1 is equivalent to calculating only
direct and diuse light, but ignoring any reected
indirect light.
For the ambient accuracy parameter, a lower
value produces more accurate results with slower
execution times. However, note that if no reected
indirect lighting is calculated (as is the case when
the only light is coming from the sky dome and the
ambient bounces parameter is set to 1), then this pa-
rameter can be set to 0, as it will not have any eect.
For the nal three parameters, higher values will
generally produce more accurate results but with
slower execution times.
558 | eCAADe 30 - Volume 1 - Simulation, Prediction, and Evaluation
CASE STUDY EXPERIMENT
The case study experiment will focus on the Inter-
lace, a large residential project designed by OMA
[7] and currently under construction in Singapore.
The design consists of thirty-one apartment blocks,
each six stories tall. The blocks are stacked in an in-
terlocking brick pattern, with voids between the
bricks. Each stack of blocks is rotated around a set of
vertical axes, thereby creating a complex interlock-
ing conguration. An example is shown in Figure
2, where 6 blocks are stacked and rotated to form a
hexagonal conguration.
Design exploration task
For this research, an exploration task has been de-
ned, in which the designer is required to minimize
the number of windows receiving either low illu-
minance or high irradiance. The designer will carry
out this exploration task via a process of iterative re-
nement, whereby a parametric model is built and
the parameters in the model are gradually adjusted
in order to try and improve performance. Each it-
eration of parametric adjustment by the designer
is followed by a simulation of design performance,
and if performance improves then the parametric
changes are kept. Using this approach the designer
may gradually be able to improve performance of
the design.
For this task, the parametric changes that can
be made by the designer have been constrained to
the rotation of the blocks and the addition of sun
shades. Block rotation a change that aects global
conguration, while sun shading is seen as a change
that aects only local conguration. In order to con-
strain the task, other possible changes, such as the
stacking of the blocks and the position and size of
balconies were not considered. However, it is noted
that the iterative approach used in this research
could also be expanded to include such parameters.
In order to test the iterative approach, a Houdini
model of the design was built that included all sig-
nicant exterior features including walls, windows,
inset balconies and protruding balconies. On the
interior, most of the detail was omitted and only
unit walls were included. The resulting model had
a total of approximately 47.5 thousand polygons, of
Figure 2
The process of rotating the
brick pattern. The diagram
on the left shows 6 blocks
arranged in a straight line,
while the diagram on the right
shows the same six blocks
folded into a hexagonal
pattern.
Figure 3
A typical oor plan [8].
Each block is approximately 70 meters long by 16.5
meters wide, with two vertical axes of rotation spaced
45 meters apart. The axes of rotation coincide with
the location of the vertical cores of the building,
thereby allowing for a single vertical core to connect
blocks at dierent levels. The blocks are almost totally
glazed, with large windows on all four facades. In ad-
dition, blocks also have a series of balconies, both
projecting out from the facade and inset into the fa-
cade. A typical oor plan is shown in Figure 3.
The OMA design has stacked the 31 blocks into
22 stacks of varying height, and has then rotated the
stacks into a hexagonal pattern constrained within
the site boundaries. At the highest point, the blocks
are stacked four high.
559Simulation, Prediction, and Evaluation - Volume 1 - eCAADe 30 |
which about 7800 were windows. These windows
were grouped into four types: living room windows,
bedroom windows, kitchen windows, and utility
windows. For the performance exploration, it was
decided to focus on the living rooms and bedroom
windows only, which totalled 5250 windows. The il-
luminance and irradiance incident on each window
was measured at just one point in the centre of the
window. Therefore, for each iteration, illuminance
and irradiance was to be measured at 5250 points in
the model.
For the exploration task, target thresholds were
set for both illuminance and irradiance. Windows
falling either below the illuminance threshold or
above the irradiance threshold were considered to
be undesirable, and therefore in need of improve-
ment. The aim of the exploration task was to reduce
the total number of undesirable windows. These
thresholds were mainly used as a simple way of
summarizing relative performance, so that the de-
signer was able to quickly assess whether improve-
ments has been made.
Parameterisation of the model
In order to allow the designer to uidly and inter-
actively make changes to the rotation angles of the
stacks of blocks, the blocks need to be parametrical-
ly linked. Looking at the arrangement of the blocks
in plan in Figure 4, it is evident that the congura-
tion is actually a branching hierarchical structure,
with a central root and three branches.
This type of branching structure can be modelled
within animation tools such as Houdini using ob-
jects that have parent-child relationships. In the plan
in Figure 4, the root node is indicated by the larger
dot and is the parent of three block stacks: s1, s5 and
s10. Each of these three stacked blocks is the start
of one branch. The parent-child linking relationship
means that any transformation applied to an object
will automatically also be applied to all the descend-
ants. The designer can therefore freely explore dier-
ent rotation combinations without having to worry
about the stacked blocks becoming disconnected.
Iterative simulation design method
The key step in the iterative simulation design meth-
od was the executions of the simulations. Calculat-
ing the illuminance and irradiance at a high level of
accuracy can be very time consuming, and therefore
very disruptive for the designer.
For obtaining accurate results, the following Ra-
diance ambient settings were used: ab=4, aa=0.15,
ar=2048, ad=516, and as=516. Using these settings,
the illuminance simulation took 8 hours and 30 min-
utes and the irradiance simulation took 13 hours 50
minutes making a total of 22 hours and 20 minutes.
The computer being used for running the simula-
tions was a typical oce computer: a 2.4GHz dual-
core processor with 8GB RAM running 64 bit Win-
dows.
Figure 4
The original design. The plan
on the left shows the root
node and the branching
structure. The model on the
right shows windows with low
illuminance in dark grey, and
windows with high irradiance
in light grey.
560 | eCAADe 30 - Volume 1 - Simulation, Prediction, and Evaluation
The simulation results showed that within the ex-
isting design, a signicant portion of windows had
either low illuminance or high irradiance. The il-
luminance and irradiance patterns on the facade
were also seen to be very varied and hard to pre-
dict due the complex massing of the building, and
also due the eects of protruding balconies shad-
ing windows below. The iterative simulation design
method was therefore deemed to be appropriate for
exploring options with fewer undesirable windows.
However, due to the excessive simulation time, an
iterative simulation design method was developed
where each simulation was congured to run in fast
mode and in slow mode. The aim was to reduce the
total simulation time of the fast version to below
two minutes, but to ensure that the results from the
fast and the slow mode simulations still correlated
reasonably well. This would then allow the fast simu-
lation to be used as a driver for the exploration pro-
cess.
The iterative simulation design method was di-
vided into three main phases: calibration, iteration,
and verication. In the calibration phase, the fast
mode simulations are set up and congured in order
to ensure that appropriate trade-os are achieved
between speed and accuracy. In the iteration phase,
the fast mode simulations are used within the itera-
tive renement process in order to explore design
variants with improved performance. Finally, in the
verication stage, both the initial design and the -
nal design from the iterative process are evaluated
using the slow mode simulations in order to verify
the performance improvements. The three phases of
the iterative simulation design method are shown in
Figure 5.
The calibration phase
For the calibration phase, a series of Radiance simu-
lations were executed with parameter settings that
favoured speed over accuracy. In all cases, the ambi-
ent bounces parameter was set to 1 and the ambi-
ent accuracy parameter was set to 0. This therefore
meant that no indirect reections were calculated
which signicantly reduced the execution time. For
each of these simulations, Microsoft Excel was then
used to plot the trend-line between the fast and
slow mode simulation results, and to calculate the
R2 correlation coecient (or the coecient of deter-
mination). Table 1 shows the results for these experi-
ments.
Figure 5
The three phases of the
iterative simulation design
method.
561Simulation, Prediction, and Evaluation - Volume 1 - eCAADe 30 |
Based on the execution time and R2 correlation
results, it was decided that for both the fast illumi-
nance simulation and the fast irradiance simula-
tions, the second set of settings from table 1 would
be used. These settings allow the simulations to be
executed in under 1 minute each, and also maintain
an R2 correlation of close to 0.9.
The nal step in setting up the fast simulations
was to map the results from the fast simulation us-
ing the linear trend-line equation. Microsoft Excel
was used to obtain the linear trend-line equation,
which was then transferred back to Houdini, where
it was used to map the results from the fast simula-
tion. This option for mapping the simulation results
was provided as part of the Houdini node. In eect,
this mapping of the fast simulation results adjusts
the trend line so that it passes through the graph
origin at 45 degrees.
The iteration phase
Within the Houdini environment, the total number
of undesirable windows for both illuminance and
irradiance were continuously displayed to the de-
signer as both numeric totals, and as coloured poly-
gons within the three-dimensional model. Once the
designer had made a set of changes to the model,
they were then able to trigger the simulations to
re-execute. After two minutes, once the simulations
completed executing, both the numeric totals and
the coloured polygons would be automatically up-
dated, thereby giving fast feedback to the designer
as to whether their changes resulted in better per-
formance.
The exploration process was set up as a two
stage process. In the rst stage, the rotation parame-
ters were iteratively explored. For each iteration, the
designer would identify a particular cluster of win-
dows with low illuminance, and would then make a
Table 1
Table showing the execution
time (T) and R2 correlation for
a range of dierent ambient
light settings for the Radiance
rtrace simulation.
Figure 6
The design modied in order
to reduce the number of win-
dows with low illuminance.
The plan on the left shows
how the branching structure
has been modied to try
and increase the openness
between the branches. The
model on the right shows win-
dows with low illuminance in
dark grey, and windows with
high irradiance in light grey.
Radiance rtrace ambient settings
Illuminance
Irradiance
T
R2
T
R2
ab=1, aa=0, ar=2048, ad=512, as=512 92s 0.8892 88s 0.8841
ab=1, aa=0, ar=1024, ad=256, as=256 49s 0.8892 53s 0.8839
ab=1, aa=0, ar=512, ad=128, as=128 32s 0.8875 36s 0.8796
562 | eCAADe 30 - Volume 1 - Simulation, Prediction, and Evaluation
small number of changes in order to try to reduce
the obstructions for those windows. In some cases,
such changes would indeed improve the situation,
but in other cases, the changes would cause dete-
rioration in performance in some other part of the
design. At this stage, the focus was on reducing the
number of windows with low illuminance, as this
was deemed to be a more challenging task. How-
ever, the designer also kept a check on the number
of windows with high irradiance, since changes that
improved illuminance often also resulted in higher
levels of irradiance. The nal design for stage 1 is
shown in Figure 6.
In the second stage, the best solution from the
rst stage was selected and the addition of solar
shading devices was then explored with the aim of
reducing the number of windows with high irradi-
ance. For the rotation parameters, the changes were
applied manually, since there were only 22 set of
stacked blocks. However, for the windows, the man-
ual approach could not be used since there were
thousands of windows. An automated approach was
therefore created within Houdini whereby shading
devices were parametrically generated for the win-
dows with high levels of irradiance. The depth of
the shading devices was varied in relation to level
of irradiance on the window. In this case, the itera-
tive process was used to explore the relationships
between the depth of the shading devices and the
level of irradiance on the window. As with stage 1,
the designer also kept a check on the number of
windows with low illuminance, since the addition of
solar shading devices reduced illuminance levels in
some cases.
It was found that in the rst stage, the reduction
of the number of windows with low illuminance was
dicult to achieve. The number of low illuminance
windows was reduced through an iterative rene-
ment process consisting of 18 iterative steps. In the
second stage, the windows with high irradiance
were more easily solved using additional sun shad-
ing devices. During this stage, the number of high
irradiance windows was reduced in 6 iterative steps.
The verication phase
In order to verify that performance had indeed been
improved, the initial design and the nal design
were evaluated using the slow mode simulations
and the results were compared. Note that the goal
of this verication was not to compare the results
from the fast mode simulations with those from
the slow mode simulations, but rather to measure
the actual performance improvements that were
achieved through the iterative renement process.
Despite good R2 correlations of close to 0.9, the re-
sults from the fast mode simulations could not be
used as an objective measure of performance. In-
stead, the fast simulation modes were used only as a
way of measuring relative performance, and within
the iterative phase were used as a driver for the ex-
ploration process.
The slow mode simulation results show that the
total number of windows with low illuminance and
high irradiance have been reduced by 8% and 32%
respectively. This conrms that the performance was
successfully improved using the iterative simulation
design method.
CONCLUSIONS
This research aimed to explore the trade-o be-
tween speed and accuracy when applying iterative
simulation approaches to complex designs where
the size of the digital models typically becomes
large, and as a result execution times for simulations
may become prohibitively slow.
This research has explored an approach in which
simulations are run in two modes: fast mode and
slow mode. An iterative simulation design method
has been developed consisting of three phases:
in the rst phase, fast mode simulations are cali-
brated by setting appropriate trade-os between
speed and accuracy; in the second phase, the fast
mode simulations are used to iteratively rene the
design in response to performance feedback; lastly,
in the third phase, the performance improvements
achieved through the iterative renement process
are veried. The application of the proposed meth-
563Simulation, Prediction, and Evaluation - Volume 1 - eCAADe 30 |
od to a complex case study of a large residential
design demonstrates the feasibility of the approach.
Future research will focus on further exploring
how the proposed approach can be applied to a
wider range of simulation tools, including structural
simulations and energy simulations.
REFERENCES
Coenders, JL 2007, Interfacing between parametric asso-
ciative and structural software.In Proceedings of the 4th
International Conference on Structural and Construction
Engineering, Melbourne, Australia, 26–28 September.
Janssen, PHT, Chen, KW and Basol, C 2011, Iterative Virtual
Prototyping: Performance Based Design Exploration. In
Proceedings of The International Conference on Educa-
tion and research in Computer Aided Architectural De-
sign in Europe (eCAADe ‘11), pp. 253–260.
Janssen, PHT and Chen, KW 2011, Visual Dataow Model-
ling: A Comparison of Three Systems, in Proceedings of
the CAAD Futures ‘11, pp. 801–816.
Kolarevic, B and Malkawi, A 2005, Operative Performativity
(panel discussion), in Performative Architecture: Beyond
Instrumentality, New York :Spon Press, 2005, pp.239–
246.
Lagios, K, Niemasz, J, and Reinhart, CF 2010, Animated
Building Performance Simulation (ABPS) – Linking Rhi-
noceros/Grasshopper with Radiance/Daysim, in Pro-
ceedings of SimBuild 2010, New York City, August 2010
Robinson, D and Stone, A 2004, Irradiation modelling made
simple: the cumulative sky approach and its applica-
tions, InPlea2004 – The 21st Conference on passive and
low Energy Architecture, Eindhoven, The Netherlands.
Tregenza, P 1987, Subdivision of the Sky Hemisphere for
Luminance Measurements, Lighting Research and Tech-
nology, Vol 19, pp. 13–14.
Toth, B, Salim, F, Frazer, J, Drogemuller, R, Burry, J, and Burry,
M 2011, Energy-oriented design tools for collaboration
in the cloud. International Journal of Architectural Com-
puting, 4(9): 339–359.
Shea, K, Aish, R, and Gourtovaia, M 2005, Towards integrat-
ed performance–driven generative design tools. Auto-
mation in Construction. March 2005, 14(2): 253–264.
[1] http://radsite.lbl.gov/radiance/
[2] http://radsite.lbl.gov/radiance/man_html/gensky.1.html
[3] http://diva4rhino.com/
[4] http://apps1.eere.energy.gov/buildings/energyplus/
cfm/weather_data.cfm
[5] http://www.sidefx.com/
[6] http://radsite.lbl.gov/radiance/man_html/rtrace.1.html
[7] http://oma.eu/projects/2009/the-interlace
[8] http://www.theinterlace.com.sg/
564 | eCAADe 30 - Volume 1 - Simulation, Prediction, and Evaluation
