PERFORMANCE DRIVEN DESIGN OPTIMISATION WITH
SCIENTIFIC WORKFLOW SYSTEM
CHEN KIAN WEE
Singapore ETH Centre – Future
National University of Singapore
Institute of Technology in
Architecture, Department of
Architecture, ETH Zürich
This paper proposes the use of a scientific workflow system to facilitate the
implementation of a performance driven design process. The use of a scientific
workflow system would enable the architect to customise and automate his/her
own workflow for each project. Through this the architect will be able to
manipulate the data generated from a performance driven design process
efficiently and explore more design options. A simple example is presented to
demonstrate the design process proposed. The scientifc workflow system,
Kepler is used to oversee the design process and it is linked with the 3D
modelling program, SideFX Houdini 3D. Lighting simulation program,
Radiance and lastly Evolutionary Algorithm (EA) is used for the optimisation
of the design.
Keywords: Performance driven design, Scientific workflow system,
Evolutionary algorithm, Parametric design, Design process
The integration of environmental simulation and Computer-Aided Design
(CAD) programs into a common platform is essential for a performance oriented
building design. This integrated environment facilitates a performance driven
design process by providing feedback on the architect's design decision.
Furthermore, the automation of the design performance assessment and the
employment of optimisation algorithms enable the architect to generate a
multitude of possible design variants and thus explore more design possibilities.
Research shows that the majority of architects and engineers use simulation
programs for validation and not for exploration of alternative designs (Flager
and Haymaker 2007). One of the reasons is that the data from most of the CAD
program are not sufficient and often not compatible with simulation programs.
This greatly impairs the flow of the design process, as architects then have to
manage and error check the various domain specific data exchange, instead of
spending time on developing the design.
There are basically three level of integration of CAD and simulation
programs into the design process. At the first level, a standard CAD system is
linked to a set of simulation programs so that one can easily evaluate design
variants that are generated manually. There are various research efforts in this
field to achieve this smooth transition from the CAD system to each domain
specific simulation program. (Citherlet, Clarke et al. 2001; Malkawi 2004) One
of the example is the post processing and mapping of Building Information
Modeling (BIM) 3D model to domain specific simulation program (Sanguinetti,
Abdelmohsen et al. 2012), it tries to achieve a smooth transition between design
and evaluation by processing the data from the BIM model and preparing it for
the different simulation programs. Another example is The Design Analysis
Integration (DAI) initiative.(Augenbroe, Wilde et al. 2003; Augenbroe, Wilde et
al. 2004) It uses a workbench to manage the process and at the same time make
use of existing file formats like Industry Foundation Class (IFC) for the
information exchange. DAI do not prescript the work flow for the user but give
the user certain amount of flexibility to customise their workflow.
At the second level, a parametric CAD system such as Grasshopper or
Houdini3D is linked to a set of simulation programs so that design variants can
be more easily generated and evaluated without manually remodeling each
design from scratch. Parametric modeling refers to the process in which the
architect defines the design with a set of parameters. For example a set of
parameters to control the height, breath and length of a rectangular block tower.
This can speed up the process of design generation as various different designs
can be generated by varying the parameters. As a result more “what if” scenarios
and designs could be explored and evaluated(Shea, Aish et al. 2005).
Lastly, an optimisation algorithm is used to close the loop by linking the
simulation results back to the parametric CAD system, so that one can more
easily explore the performance tradeoffs of a large number of design variants
without the need to manually define the parameters for each design
variant.(Caldas 2006; Shea, Sedgwick et al. 2006; Flager and Haymaker 2007;
Flager, Welle et al. 2009). The use of optimisation algorithms such as Genetic
Algorithm (GA) and Ant Colony Algorithm is usually a more bottom up
approach, where the architect sets up rules, constraints and boundary conditions
for the generation of design alternatives. It is more of an exploratory nature in
which the algorithm takes a more active role in the design process and might
generates unexpected design alternatives.
This paper proposes the adaptation of a scientific workflow system as an
environment for the architect to set up design loops for performance driven
design optimisation. The use of a workflow system in the scientific community
is common practice, as most of an experiment involves repetitive cycles of
simulation, analysis and management of results. The use of a scientific
workflow system enables the automation of the cycles and at the same time
allows the scientist to concentrate on the research and not the computation
management. There is an array of available workflow system for tasks of
different purposes (V. Curcin 2008; Deelmana, Gannonb et al. 2009). This paper
presents the initial approach of the adaptation of the tool for the architecture
SCIENTIFIC WORKFLOW SYSTEM
A workflow is an abstract description of the steps needed and the
information flow between the steps for executing a specific process. Each step is
made up of a series of activities. The activities include processing of data for the
next step, execution of a simulation with the provided data or the analysis of the
given data. The steps are executed in order from start to finish and most of the
time it is repeated in cycles with variation of the value of the information, but
not the nature of the information. The workflow is usually constructed with a
visual front-end or hand-coded.
The visual front-end of a workflow system is usually a graph illustrating the
order of the execution of steps needed to complete a specific process. The
connecting wire between each node is representation of the flow of data between
the different steps. The users can usually alter the workflow by manipulating the
graph, it varies depending on different systems.
One key advantages of using a workflow system is that it serves as a form
of documentation for one's design process. As alterations to the design are
common during the architectural design process, a well documented design
process facilitates making changes to the design. The workflow could be reused,
improved or altered after each project. It also enables the sharing of workflow
between collaborators. Some workflow systems allow the nesting of sub-
workflow within a workflow. This means as long as the inputs and the outputs
of different steps are well defined, nested workflows could be distributed for
collaborative purposes. This results in a modular workflow system which is
highly flexible and customisable.
Figure 1 Possible modular workflow with well-defined data exchange
The diagram in Figure 1 conceptually illustrates a modular workflow
system with well defined data exchanged between each sub-workflow. In this
case, each sub-workflow could be authored by different professionals. For
example, the architect might author the CAD sub-workflow while the
mechanical engineer authors the SIM1 sub-workflow. This could be possible by
the usage of common file exchange such as Industry Foundation Class (IFC) or
an agreed upon data format between the two professionals.
THE WORKFLOW SETUP
A workflow which uses Evolutionary Algorithm (EA) for its feedback
mechanism was setup. The workflow is made up of two main components; the
feedback and evaluation component. The Kepler system (Pennington, Higgins et
al. 2007) was used for managing the workflow of the design process. The Kepler
system uses a Director and Actor modelling paradigm in which the Actors are
the workflow components while the Director is in charge of the overall
workflow orchestration. SideFX Houdini3D was the 3d modelling program used
for the parameterising and generation of the design schema. For the evaluation
tasks the workflow is able to link the Radiance lighting software for simulating
solar irradiation and Python scripts were written for the calculation of the RETV
value and for the implementation of EA in Kepler system.
Figure 2 Parameters and sub-workflow
Figure 2 shows the workflow. Each node in the graph is a nested workflow.
The parameters shown in Figure 2 are the necessary inputs for running the
workflow. It includes information for running the EA, to generate the designs
and also the location of files necessary for running the workflow. These
parameters feed the sub-workflows with the information for running the
Figure 3 Inside the "Feedback" sub-workflow
At the start of the workflow is the “Init and Feedback” sub-
workflow.(Figure 3) This sub-workflow is in charge of doing the EA feedback
mechanism. It generates new designs either randomly in the first cycle or by
reproduction at the subsequent cycles. The sub-workflow is made up of an EA
environment node written in Python script. The node takes in the Genotype
Meta settings and produce the parameters for generating the design variants. It
takes in the Score Meta settings for manipulating the evaluation scores produced
by the evaluation task, whether to minimise or maximise is specified at the
“min_max_list” parameter. Lastly, it takes in the “live_file” and “dead_file”
parameters for writing the results.
Figure 4 Inside the "Evaluation" sub-workflow
The design variants are send to the “Evaluation Task” as a array of
parameters.(Figure 4) The sub-workflow includes generating a 3D model by
accessing the 3D software Houdini3D, extracting the geometry data and sending
it for evaluation tasks such as solar irradiation or RETV evaluation. This is done
by using the Python Houdini3D Advance Programming Interface (API). A
Python script was written to access Houdini3D and the parameters are passed to
the script as variables for generating a design variant. The geometry of the
design variant is then extracted and processed for the relevant simulations. All
these are done through the use of a library of Python codes and API written by
The parameters and the scores of each design variant is sent to the
“WriteToLiveFile” sub-workflow in which the designs are written to a .csv file.
The cycle repeats itself at the “Init and Feedback” sub-workflow where it will
read the .csv file produce from the previous cycle and perform EA feedback
APPLICATION AND CASE STUDY
A simple case study of a mix-used development in Singapore was done to
demonstrate the workflow setup. It is parameterised as follows:
The two storeys commercial mall is not altered during the design
generation. The residential tower above the mall varies with the
parameters X and Y which determine the footprint of the
residential tower. This in turn will affect the height of the tower as
The orientation and location of the residential blocks.
The height of the residential block exceeds 60m, instead of one
60m tower it will generate two 30m towers.
Figure 5 Design schema and parameters
The environmental strategy is to maximise the amount of solar irradiation
falling on the building envelope while minimising solar heat gain. This
corresponds to the conflicting objectives of generating as much electricity from
the installation of Building Integrated PhotoVoltaics (BIPV) while at the same
time reducing the amount of solar heat gain. It is an attempt to source for
possible energy source on site and still reduce the demand of cooling energy in
The evaluation methods used are the Radiance ray-tracing lighting
simulation for the measure of solar irradiation and Residential Envelope
Transmittance Value (RETV) (Chua and Chou 2010) for assessing the solar heat
gain of the envelope. RETV is a measure of the solar heat gain from the building
envelope, it is developed by the Building and Construction Authority (BCA) of
The design schema is optimised using an EA. At each cycle :
1. A population of 100 design variants are randomly generated by
assigning parameter values to the four parameters described above.
2. Each design variant is evaluated and assigned scores for the two
3. A sub-population of 50 design variants are randomly selected from
the main population and pareto ranked according to their scores.
4. 30 design variants at the bottom of the ranking will be “killed” and
20 new design variants are reproduced from the “stronger” design
variants, the reproduction process is done through crossover
techniques and mutation occurs at a rate of 0.01.
5. The cycle is repeated from step 2 with the “fitter” design variants.
At the end of each cycle the design variants that are “alive” or
“dead” are written to two .csv files live.csv and dead.csv
respectively. The architect will be able to check the progress by
reading the .csv file.
As a result, the population of designs gradually evolves after each cycle.
The optimisation loop stops as specified by the user at the start or anytime
during the process. It is an performance driven design exploration of the
possible design variants.
The EA process ran for 160 generations and 4000 designs were produced.
The graph in Figure 9 shows the design variants produced at each generation.
One can observed that after each generation the design variants are moving
towards bottom right corner of the graph forming the pareto front, where it is of
high solar irradiation and low solar heat gain. Thus, achieving the aim of
increasing the generation of electricity with BIPV while still keeping the cooling
demand as low as possible.
Figure 6 Graph of EA results
Figure 7 Design 4209 (top) Design 4116 (bottom) on the pareto front
The case study illustrates a basic workflow set up with Kepler. This
workflow or its sub-workflow can be packaged and shared as a template for
future use. It can be reused or altered to fit the specific project. For example,
another user might not be proficient in using Houdini3D and is only proficient in
Rhinoceros3D. The Houdini3D actor node could be replace with a
Rhinoceros3D actor node. This applies to the evaluation actor node too. As more
users are willing to contribute , it will expand the capability of the workflow
system and accommodate more digital tools for design purposes. It is also
possible for the designer to first do a series of non-looped experiments where
the designer 'learns' what needs to be optimised. The designer could then
incorporate an optimisation algorithm to optimise the key aspects. By having
well orchestrated workflow, it is possible for architects to “plug in” design into
the workflow to be evaluated and optimised for its environmental performance.
[Augenbroe, G., P. d. Wilde, et al. (2004). "An interoperability workbench for design analysis
integration." Energy and Buildings 36(8): 737-748.
Augenbroe, G., P. d. Wilde, et al. (2003). THE DESIGN ANALYSIS INTEGRATION (DAI)
INITIATIVE. Eighth International IBPSA Conference. Eindhoven, Netherlands: 79-86.
Caldas, L. (2006). GENE_ARCH: An Evolution-Based Generative Design System for Sustainable
Architecture Intelligent Computing in Engineering and Architecture. I. Smith, Springer
Berlin / Heidelberg. 4200: 109-118.
Chua, K. J. and S. K. Chou (2010). "An ETTV-based approach to improving the energy performance
of commercial buildings." Energy and Buildings 42(4): 491-499.
Citherlet, S., J. A. Clarke, et al. (2001). "Integration in building physics simulation." Energy and
Buildings 33(5): 451-461.
Deelmana, E., D. Gannonb, et al. (2009). "Workflows and e-Science: An overview of workflow
system features and capabilities." Future Generation Computer Systems 25(5): 528 - 540.
Flager, F. and J. Haymaker (2007). A comparison of multidisciplinary design, analysis and
optimisation processed in the building construction and aerospace industries 24th W78
Conference Maribor 2007, Bringing ITC knowledge to work. Maribor, scix.net: 625-630.
Flager, F., B. Welle, et al. (2009). "Multidisciplinary Process Integration and Design Optimisation of
a Classroom Building." Journal of Information Technology in Construction 14: 595-612.
Malkawi, A. M. (2004). "Developments in environmental performance simulation." Automation in
Construction 13: 437-445.
Pennington, D. D., D. Higgins, et al. (2007). Ecological Niche Modeling Using the Kepler Workflow
System Workflows for e-Science. I. J. Taylor, E. Deelman, D. B. Gannon and M. Shields,
Springer London: 91-108.
Sanguinetti, P., S. Abdelmohsen, et al. (2012). "General system architecture for BIM: An integrated
approach for design and analysis." Advanced Engineering Informatics 26(2): 317-333.
Shea, K., R. Aish, et al. (2005). "Towards integrated performance-driven generative design tools."
Automation in Construction 14(2): 253-264.
Shea, K., A. Sedgwick, et al. (2006). Multicriteria Optimization of Paneled Building Envelopes
Using Ant Colony Optimization. Intelligent Computing in Engineering and Architecture
13th EG-ICE Workshop 2006. I. F. C. Smith. Ascona, Switzerland, SpringerLink. 4200:
V. Curcin, M. G. (2008). Scientific workflow systems - can one size fit all? Biomedical Engineering
Conference, 2008. CIBEC 2008. Cairo International Cairo