Font Size: 

Introduction

Computational design approaches employing parametric models are increasingly used in design processes because of the designers’ ability to embed behaviors in the model and explore variations and variants by changing parameters without spending inordinate efforts on modifying the model or even on rebuilding it entirely.

It is possible to incorporate building performance analysis methods into the model so that performance metrics for each variant of the model can be computed and thus better performing variants of the design can be selected for further development, based on the popular assumption “fail often, fail fast, fail cheap” (quoted from Jim Estill, 2009, Blogging Innovation, without citation there) doing so with semi-automatically generated and evaluated solution in form of low cost digital models rather than with real buildings.  Parametric models permit to generate more potential solutions and with embedded analyses designers can identify better performing solutions, determine factors that influence the success of those solutions and continue to develop the design toward even better performance.

The ultimate assumption is that the result of this innovative computational design approach is a better design outcome when compared to traditional design processes, given all other constraints like time, budget, or effort remaining equal.  This type of work flow requires translation of the client’s and design team’s goals into a parametrically controlled model including types of analysis that reflect applicable goals and make them measureable.  Because it has been an innovative work flow in architectural design various researchers have investigated several aspects of learning that is involved in acquiring a working skill level of parametric design and its extensions.  Topics investigated have been for example how parametric design can be made accessible through different types of views (Aish and Woodbury 2005), how recasting of parametric models into meaningful pieces, i.e. “design patterns” may help ease their construction (Woodbury et al. 2007; Qian 2009), how sensitivity analysis unlocks the understanding how specific parameters influence behavior of the parametric model (Hopfe et al. 2007; Hopfe 2009), or what challenges analysis integration into early design processes may pose (Bleil de Souza 2009, 2012).

This paper reports about a conceptual parametric design process as a case study to illustrate yet different aspects of learning that are involved or even required for a successful computational design process integrating various building performance metrics as design criteria through multi-objective optimization.  In multi-objective optimization a parametric model is utilized to generate in automated fashion many candidate solutions.  The integrated analyses compute for each solution the corresponding metrics which may be subject to additional computation in order to match more closely the design goals.  These values are routed as objective functions to the optimization algorithm which uses them to determine the next set of candidate solutions.  For various types of optimization algorithms the details may vary.  Crucial is that with the search for the set of best performing solutions having been automated there is an additional level of due diligence necessary to assure that this automated search is leading in the correct direction.

Methodology

The methodology used is an ethnographic approach focusing on participant observation amended by discussions replacing qualified key informant interviews.  The material is presented in form of a case study even though similar observations have been made in similar cases; however, there has not been an attempt to formally report on those other cases.  It is based on observations of learning processes which occurred during a two-day workshop about parametric design with integrated analysis and multi-objective optimization at a conference for computer aided architectural design.  Most participants were designers or researchers with experience in parametric modeling, analysis embedded in design processes, and optimization.

The Case Study

The workshop participants chose as format a collaborative approach similar to a design team working on a single project.  A parametric design tool with scripting and graphic programming capabilities was used combined with an optimization module using a genetic algorithm to generate the individuals in the populations used for the optimization process.

The workshop started with a simple mass model using four simplified analysis methods as proxies to evaluate states of the parametric model for energy efficiency, day lighting potential, economic performance, and construction costs.  The assumed correlated project goals would be a high level of environmental performance as measured by energy usage for climate control and lighting, as well as a high level of economic return combined with a low level cost of first ownership.

The kick-off step was for the team members to familiarize themselves with the model by evaluating the model’s behavior, determining whether parameters controlled what their names claimed to control, and to validate that the model’s optimization behavior did actually optimize for the project goals.  Iterative modification and exploration of the model supported by multi-objective optimization first remedied and then amended the model to achieve more consistent behavior in order to generate improved results in the semi-automated model exploration process.  These iterations helped the team tune the model behavior to match better the expectations set by the project goals.  What has been described so far encompasses an initial level of learning in a practice-oriented computational design process: learning about the model behavior in order to ascertain that it meets the design intent.

After validating the design model’s behavior multi-objective optimization reveals what the trade-offs are between the various optimization criteria which represent the project goals.  Based on this next level of learning about the design project which is a uniquely enabled by multi-objective optimization the design team can formulate what decisions the client needs to make in order to select either a single solution or update the strategy what design trajectory to pursue further.  In the workshop an additional iteration added as a project goal an aesthetic criterion for which a computational proxy was defined as a test of how such apparently non-computable design goals could potentially be encoded in a way to make them computable for inclusion in the overall computational design process while not losing sight of their non-computable qualities.

Results

The examination of this case study reveals that there are different levels of learning about a computational design model: first, the design team needs to ascertain that the model behaves in a way that actually conforms with the requirements in order to reach the project goals; second, after this validation of the computational model, the design team can start to understand what potential trade-offs exist between different project goals, and thus understand the decisions that need to be made, or additional problems that need to be solved in order to arrive at a better design solution.  This entire process in itself is iterative.

Discussion

Multi-objective optimization automates the generation of design solutions within the solution space given by the type and range of parameters of the parametric design model.  Optimization algorithms use various techniques to generate and evaluate solutions.  Mostly these are iterative processes where the computed performance metrics of a previous solution set provides the algorithm with the information it needs to generated solutions for the next iteration.  Optimization algorithms use diverse strategies to converge on better solutions.

Multi-objective optimization algorithms, however, in most cases cannot converge on a single “best” (optimal) solution because they are commonly employed when there are potentially conflicting design goals, for example maximizing usable floor area while minimizing construction cost.  Therefore, multi-objective optimization algorithms result in a set of solutions that are better in one or another metric and not outperformed (not dominated) by other solutions as initially defined by Vilfredo Pareto in 1906 in economics and since then adapted to evaluation of engineering solutions that need to meet several criteria.  A strategy for evaluation are Pareto plots with distinct design goal metrics spanning the plot area and solutions’ performance metrics presented as scatter plots.  After an optimization run such Pareto plots reveal Pareto frontiers formed by those solutions which in either plot axis toward improved performance are not performing worse than other solutions.

In case of the design tool used in the workshop, dots in the Pareto scatter plot displayed all Pareto optimal solutions in the set in all criteria dimensions with those dots on the Pareto frontier for the respective pairwise criteria emphasized.  Selecting any dot in the plot displays the performance metrics and the geometry of the corresponding solution.  This way the design team can explore the set of solutions in order to learn how specific performance values are related also to the geometry that is generated for these solutions.  The difference between the design team’s expectations and the performance of the solutions gives rise to the opportunity for the two types of learning described in this paper.

Conclusion

Parametric modeling with integrated analysis and multi-objective optimization is not a trivial matter.  A potential pitfall is that design teams are missing to identify gaps in their concepts about computational design and rely on algorithms to bridge those gaps while they may only be hiding them.  Therefore, it is necessary to identify where these gaps may occur and raise awareness of these gaps through education, instructional materials, and other means.  Nevertheless, computational design offers new opportunities for improving the results of design processes based on measurable criteria.  Given the possible breadth of such design investigations the expectation is that given the same amount of effort, this type of computational design process can result in better designs, ultimately in better performing buildings.

simulation and modeling; generative systems; parametric design; multi-objective optimization; computational design learning