Generative Design & Metric space analysisA use case of Multi-Objective Genetic Algorithm in space planningSrihari PramodBlockedUnblockFollowFollowingMar 2Computers had a dramatic impact on architectural practice.
Although this transformation increased the efficiency and productivity incredibly, CAD was just the computerized version of the fundamental workflow where the design decisions developed through what are typically called ‘rule of thumbs’ and human design intuition.
For example the use of symmetrical layouts with a focus on the geometrical centre of the layout, and clustering programs together or placing them in proximity to access points.
Although this traditional approach is efficient, when applied to complex problems they generally do not yield to the best overall solutions.
Especially when there are more objectives the design has to be optimised for and all the processes to be streamlined, which is what a modern design asks for, the traditional design processes often fail to impress.
As the constraints get increasing, it becomes difficult to get the optimized solution within the high-dimensional space of all possible designs.
Generative design integrates artificial intelligence into the design process by using metaheuristic search algorithms to discover novel and high-performing results within a given design system.
They go beyond basic automation to create an expanded role for the human designer and a more dynamic and collaborative interaction between computer design software and humans.
Parametric vs Generative DesignAlthough the parametric approach has broadened the possibilities of design and increased flexibility, the exploration of the design space is still limited by the abilities of the human designer.
Only the constraints found explicitly in the design configurations can be taken care of, for the most part, the human designer must investigate different design options by manually varying individual parameters and evaluating each option using their own intuition and past experience which, in a way is not much different than any traditional design methods.
Whereas on the other hand, what a generative approach does is tasking a computer with exploring the design space semi-autonomously, and then reporting back to the designer which options it considers promising for further analysis.
This process potentially can reveal interesting parts of the design space and discover novel design solutions that would otherwise be hidden.
The Generative Process:Generative design is essentially a multi-stage process which broadly consists of :The Geometric model :The design of a geometric model which can create many design variations.
This requires constructing a robust parametric configuration such that unique parameters define the geometric variations that the design can possibly have.
Design | Performance Metrics :The design of a series of performance metrics is of the core importance in the generative design process.
Since the computer does not have any inherent intuition about design, the human designer must explicitly describe to the computer how to determine which designs perform better than others.
The tricky part is to convert all the quantifiable metrics into numeric for evaluation.
The Solver :The model needs to be connected to a search algorithm that can control the input parameters of the model, get feedback from the metrics, and intelligently tune the parameters to find the high performing designs.
Investigation of the resulting design data through statistical analysis.
Design ProjectThe purpose of this hypothetical project is to demonstrate the generative design in spatial planning using an evolutionary solver (Galapagos) and Multi-Objective Genetic Algorithm.
A simple city block layout has been considered and the design problem is to generate an efficient commercial block — road network.
The Geometric model:A Complete parametric model has been designed with the following conditions:The total layout has been divided into neighbourhoods by means of roads with a fixed width.
While only the road width has been fixed, the configurations are controlled by unique parameters that let the geometry iterate within the layout.
Commercial plots are divided into 4 pockets and each of them is glued along the roads.
The area of the pockets is controlled by unique parameters.
The central area which forms the intersection of roads is left out devoid of commercial plots in order to avoid traffic congestions.
Design metrics:Four design metrics have been considered to optimize the design:Commercial Plots as 4 pocketsThe total area of the commercial plots:Buzz represents the traffic congestion at junctionsSpread describes the distance between the pocketsThe major goal is to achieve a maximum footprint of commercial plots (With no less than 20% of the total layout area).
Buzz: This design metric is used to calculate the traffic congestion for a particular road network.
In this case, it has been abstracted to just the distance between junctions which ideally should be maximum.
Spread: This metric calculates the spread of the commercial lots across the layout.
The area between the centres of the lots has been considered and to be maximized.
Road area: Total circulation areaRoad area: The final metric is to achieve the maximum circulation area (With no less than 10% of the total area)Galapagos:Now that the goals are explicitly defined, time to use the search algorithm for simulation.
Initially, the model is tested in Galapagos.
Since it is a single objective evolutionary solver, the weighted average of all the metrics are fed to it before the simulation.
The simulation resulted in producing few non-reliable design configurations which are not even closer to the desired output.
(This simulation is dead slow! Only a partial simulation has been carried out just to understand the variance of the metric data and their mutual relationship)Moreover, this being a more deterministic algorithm (evolutionary in this case), most of the results produced are local optimums rather than global optimums.
Multi-Objective Genetic Algorithm:MOGA is considered to be one of the successful optimization techniques when it comes to space planning & optimization because it can optimize designs along any number of output metrics.
Furthermore, the user does not need to prioritize or weight the individual metrics beforehand.
This is because the MOGA determines relative performance based on the idea of dominance rather than the absolute difference in metric values.
Basically, it uses the concepts found in natural evolution to generate new designs based on the input parameters (genome) of previous high performing designs, thus gradually promoting the best options (survival of the fittest).
The most important feature of this technique is its ability to consider any number of objectives and constraints of various data types.
The inputs can be numeric, categorical (To be converted to discrete numeric) and sequences.
Created using the Discover Framework by Danil NagyThe job is set up by explicitly defining the inputs, constraints and objectives along with other metrics that define the structure of the simulation (Population and generation count).
This works by recombining high-performing designs from different areas of the design space, and slightly mutating designs over time, the algorithm iterates over a hyperdimensional space(16 dimensions in this particular case) to produce all possible design outcomes.
Screengrab from the actual simulationThis particular use case had generations of 100 designs each and ran the process for 50 generations creating 5,000 designs.
The starting population of 100 designs was generated by randomly sampling from the design space and used the previous result as the starting population for the next generation with 0.
2% mutation.
The entire process ran over 1.
5 days on a Windows 10 workstation with a 2.
60GHz Intel Core i7 processor and 8 GB RAM.
Exploring the resultsDataset produced as the outcome of simulationsThis process generated a data set containing 5,000 designs, including the input values for each design and its score along the six metrics.
One approach at this stage would be to filter the dataset by the metric scores and directly select a few high-performing designs for further analysis.
But more can be learned about the nature of the problem as a whole.
In order to investigate this process and gain a deeper understanding of the design space, we can study how the designs perform along the four performance metrics by plotting them using scatter plots.
Metric space analysisNot just a single best design but a range of equally high performing designs can be found along the trade-off between competing metrics.
Scatter plot of PlotArea vs Spread (Created using Explore Javascript framework)This interface allows us to create the best-possible two-dimensional projection of the high-dimensional design space and see how the sampled designs are organized within that space.
Scatter plot of PlotArea vs Buzz (The spread metric distribution is represented through colour variation)Such tools can help us understand the design problem in general and reveal potential design strategies, rather than simply picking the single best design.
ConclusionThe demonstration discussed in this write-up is fairly abstract and simplified, providing only rough geometry in a hypothetical case.
The post Generative design process consists of a deep analysis of various high performing designs and their trade-offs.
Still, there is a scope of further refinement and preparation before the design can get materialised.
More than anything else this aids the designer by suggesting potential design strategies that he can further explore to achieve a final best design.
The workflow can be improved by integrating other types of modelling, particularly machine learning, for quantifying aspects of the designs that are difficult or impossible to compute through direct calculation or also difficult to relate to a computer.
This is both challenging and interesting because it might allow the computer to develop knowledge of various intangible design aspects such as comfort, aesthetics and other spatial experiences which can potentially transform the living spaces.
Get the project repo here: https://github.
com/SrihariPramod/Generative_Design_MOGAREFERENCESTerzidis K.
, “Permutation Design”Generative Design for Architectural Space Planning: The Case of the Autodesk University 2017 Layout Lorenzo Villaggi & Danil Nagy, The Living, an Autodesk StudioTerzidis K.
, “AAD- Algorithms-Aided Design.
”Pramod is an Architectural Technologist and a data science enthusiast.
Follow his Instagram and LinkedIn handles for his works and updates.
.. More details