GENETIC MODELLING The geometrics of genetic models
Experience it first : http://digitaldesignacademy.com/processing/applet_genetic_model/index.html
What is genetic modelling? Genetic Modelling is a way of creating design representations to facilitate design exploration. In our case, it means creating design representations within CAD environments. In the context of CAD, genetic modelling it can also be called parametric variational modelling, implying that design variations are created by varying a set of design parameters of a CAD representation of a design. For the purpose of design exploration, let us consider creating design representations that hold a range of design possibilities instead a single instance of it. In other words, the design representations here represents a region in design space, instead of representing a single instance of design (represents by a single point in design space. We can explore how these design representations may be used for exploring design possibilities – as in sketching, usually at the initial stage of any design. Design possibilities are mostly explored through sketching. Sketches hold within them, multiple possibilities of a design which are either reinformed or discarded during its subsequent development. It is this ability of holding multiple representations that we wish to capture through genetic models. But in early stages of design, the design representation entails ambiguity, it also lacks complete information. Therefore, we are unable to represent this in computers – which are primarily used to represent designs at the tail end of a design process. We need to explore ways of sketching in computers. Genetic models allow us to do just that.
Another way to look at genetic modelling is to view it as a “natural” way of modelling. It is the right way and should have been the logical way but we misuse computers to represent finalised designs and thus, deprive design representations of variability. We create a pre‐formulated shell, devoid of life, flexibility and intelligence. An ideal genetic model An ideal genetic model is a representation of design that covers large areas of design space which are of interest to designers. We have to be careful here, as to how we define this; because we know that designers change their minds all the time during design development. Genetic models that create surprising and unexpected forms may be of value in stimulating designer’s search for suitable forms. On the other hand, if designs represented by a genetic model are largely undesirable or unviable, it would burden designers as they have to wade through thousands of useless designs. Hence, a genetic model 3
should have a fine balance between representing a large design space capable of exceeding designers’ imagination but avoiding regions (not totally) that are unviable or undesirable. The exploration regions can somehow be constrained by setting limits on parameters, but we are discussing how to structure a design representation well before the space is parametrically defined. So, we need to think of design spaces independent of parameters – based on their geometric variability. Genetic representations that can represent a more diverse range of designs are preferred over the ones that can only represent a limited amount of design variations as they help designers explore larger design spaces. In other words, genetic models need to be “expressive” – that is, that they should be able to represent a large number of designs that are very different from each other. Here the geometric difference between the models represents the extent of the design space. For now, we will assess this visually, for we do not have yardsticks yet to measure distances in design space. We are primarily discussing designer‐driven generative design processes which are different from automated optimisation processes (which also require variable models). When we structure models for designer‐led exploration, we need to structure them in such a way so that we can navigate search space; whereas in an automated design process it is unnecessary to do so – as the search process is automated. A well structured genetic model will not only cover large tracts of design space, it will also make its navigation easier. Let’s start with a 2D Model
We use a simple 2D model to represent a few important issues in creating genetic models. The model shown is made of 3 parameters (R1,R2 & R3) that represent the distance from the centre. By varying these distances, we can create different shapes of designs.
The variability of the model depends on the range set on parameters
R is set 60~90%
R is set 0~100%
You can observe that exploration limits set on the driving parameters (R) having a direct impact on the range of design variations. Higher number of parameters create noisy designs
You may notice that the designs become somewhat similar when the number of parameters increases. Design distinction is created when the parameters are less. It is virtually impossible to visually classify 5
designs that appear noisy. In using genetic models (for designer‐driven design exploration) we need to have a sense of where we are in the design space, when we look at a particular design. Since we cannot explore the entire design space (that contain trillions of designs), we end up exploring it based on design instances that are sufficiently different from each other. The notion of families helps here. We can use a limited number of design instances representing families of designs that are somewhat similar. It is also important to remember that designer‐led exploration may not be neither systematic nor comprehensive due to limitations of human attention spans. It is bound to be an ad‐hock process, driven by a designer’s hunches – especially in early stage design where the design is not fully formed. Classification of designs can only be done when the designs are distinctive. Hence, we need to develop a sense of “seeing through the noise” and develop the ability to see core geometries. Those who study signal processing have managed to do this. They use something called Fourier transforms to break a complex signal into its constituent parts. Lessons from Signal Processing The square wave shown in red can be broken down into component waves. The component waves have higher frequency and lower magnitudes than the fundamental wave that is dominant. The geometries of objects too can be broken down along such lines.
From : http://referencedesigner.com/books/si/fourier‐transform.php
Frequencies in genetic models Genetic models may also be seen to be composed of low and high frequency geometry. The base model outlined in red is of a low frequency which is then modified by the addition of a higher frequency signal shown in blue. By breaking down the geometry into low and high frequency components, we are able to create distinct geometries that have complex shapes instead of creating noisy shapes that have no distinct form or character.
Lessons from evolution The evolutionary tree presents another way to look at design classification. It solves the problem of classification through the tracing of the common ancestry. This allowed biologist to group species in meaningful ways, according to shared attributes. This gave rise to the concept of the “common ancestor” at the very bottom of the tree. This commonality is more pronounced in developmental stages.
Biological designs have more commonality in the developmental stages and diverge as they develop. It is the development process that creates the significant differences that we see in life forms. Can we use such a strategy to help us classify designs during the developmental process? If yes, then it would help us navigate vast search spaces, because we then can trace the developmental points at which the species diverge. Biological developmental processes are full of amazing strategies from which we can learn. But can we apply what we learn from it, in developing CAD based genetic models?
Developmental form When building genetic models, it is often helpful to imagine a developmental process to be made of a combination of moulding and sculpting steps (as in foetus development). But these steps need to be implemented using the native functions of the CAD system such as extrude, cut, pattern, and revolve. Most CAD systems come with what is called build history (sometimes exposed and sometimes hidden) – which records the sequence of actions required to create forms. This is a useful feature that we can also use for structuring the developmental form of CAD‐based genetic models. While we do not really know how to structure developmental processes in optimum ways, we can certainly work towards it. By experimenting with various ways (other than the obvious straight forward way of constructing CAD models), we can develop models that are expressive – covering a larger design space. Layered Geometry It will help us to think of geometry in terms of layers that are applied on top of each other; each layer depending on the layer below. This way, when we change the bottom layers, the top layers are automatically updated. This is what good modelling is all about anyway. CAD systems now enable this through relationship management tools. CAD systems with kernels allow designers to build complex relationships graphically. Those without geometric kernels require designers to explicitly define relationships. Build history is CAD’s way of noting down the layering involved in constructing geometry. The art of building genetic models in CAD is the art of extending this capability to create highly variable forms. This requires the capturing of the geometric login within the design representation, purely through the way the design is constructed in CAD without the use of external logic models or programs to maintain the geometric logic of the model. 8
Most advanced CAD systems are built around a geometric kernel. This kernel maintains the logic of the geometry. CAD systems with kernels therefore have an “awareness” of geometry, whereas the quick and zippy CAD systems without kernels do not have an inbuilt awareness of geometry. For example, if we draw two boxes within each other, a CAD system with a geometric kernel will know that dimension y is 5, because it will have an internal understanding of the concept of square, parallel lines and other geometric concepts.
While geometric kernels make parametric designs so much easier, they pose a problem ‐ CAD kernels are different to each other (ever wondered why it is difficult to take designs from one CAD package to the other?). On the other hand, CAD kernels can be very helpful. If this were to be a site layout problem and the building is to be placed two meters from the site boundary, it can be implemented directly through the dimension system. The other advantage is that CAD kernels are able to alert us when the geometry gets to a ridiculous state. They will flag a build error whereas those without a kernel will continue to construct the most ridiculous geometries at great speeds as they act without geometric intelligence. CAD kernels certainly make building logical models that can be constrained to behave within desirable limits, easier. But if the models are to be shared across CAD platforms, then it is important to structure them in ways that they are kernel independent. But by doing so we may lose the use of some of the facilities in CAD packages that can maintain geometric logic. Hence, kernel independent genetic models come with a price, because every geometric action then needs to be explicitly defined (e.g. If you place a point in a line, you need to indicate that that point sits in that line). A way out of this would be to separate the structure of the genetic model in a graphic representational form and then implement it in CAD packages in the most appropriate way, fully exploiting its native capabilities. This way we can share the structure of the data in a useful sharable form even though we may not be able to transport the representation to other CAD packages. Hence it would be a good practice to define genetic geometries in diagrammatic form without reliance on kernels. CAD kernels are very useful in creating complex layered geometries. While it is possible to manage kernel free genetic models within CAD engines powered by kernels, the reverse is not possible. Design Skeletons 9
Another way of capturing and maintaining the geometric logic of artefacts is through the use of geometric skeletons. Designers sometimes create skeletal sketches on which their designs hang. These representations are mostly used for the construction purpose only. These skeletons can be hidden but can be used for generative purposes. You can see how nature modifies biological skeletons to achieve vast design variations. Nature’s super sophisticated ways of structuring genetic information allows it to cover vast amounts of design spaces using shared genetic models and build procedures that are common across species.
From : http://www.ekcsk12.org/faculty/jbuckley/leclass/evolutionqz1.html
While the bottom layer of the build process is shared, the top layer is not. That explains why we do not end up looking like fruit flies with which we share great many build processes. Creating Genetic Models in CAD Good CAD modelling practice is also about layering geometry in a structured way, but this is rarely done in practice as most of the exploration happens off CAD. Currently, CAD is used only to create the final representation of the design. Creating genetic models in CAD enforces good modelling practices requiring designers to capture the geometric rational of the model. History‐based CAD allows just that. It records and replays a logical build process through which complex forms are authored. We need to get good at this. Building genetic models is about capturing the geometric spirit of objects and not about creating shells. By capturing the logic of designs we are inevitably capturing geometric intelligence. That is what genetic models are all about. We need to view genetically structured designs not in its final form but through its developmental stages; the early stages being the more important ones where the defining characteristics emerge. This is akin to rough sketching, where the most important design decisions are made. We need to develop genetic models that can explore design spaces at various stages of the build process (not only at the final stage), it is only by giving your genetic model a good workout at every stage of the developmental process can you build genetic models capable of creating great diversity. Geometric Workout The purpose of creating genetic models is to empower you to explore the far corners of the design space while maintaining the logic of your design. Like any form of good exercise, this will stretch your models to the limits and it should. Models that are poorly constructed will fail; only those that are well structured will survive. By fail, we mean the collapse of the intent of the model or simple geometric failure. A good genetic model will maintain high levels of design intent throughout the workout session. They will remain 10
stable, like good gymnasts who remain stable (and even smile) while they subject their bodies to extreme contortions. Structuring genetic models is also about anticipating the kind of geometric torture that the model will be subject to and ensuring that it will hold onto its geometric intent. Hence, genetic models have to be built, exercised, re‐built, re‐exercised untill they display stable behaviour during generative design processes. Hence, building genetic models is an art that requires higher levels of understanding of how geometry can be structured to enable high levels of meaningful variations. But then, if the model behaves entirely in very predictable ways, then there is something wrong – because it will be at the expense of wider exploration. A good genetic model will always be able to produce surprising results, but at the same time produce a large range of viable designs. This is a balance that only experience will help develop. Nature seems to have figured out how best to balance it, as it mutates only a small percentage to reach out into the vastness of design space. Capturing Commonality From birth, we develop our abilities to identify what is common in all things around us. We learn to identify plant, animal, bird forms by what is common in them. We do the same in architecture through the study of historic styles. We need to intensify this ability of observing the commonality in designs until it becomes an unconscious ability. For instance, if you are shown various types, sizes and vastly different shapes of dogs, you can effortless see the underlying commonality. We are able to see this in biological creations but not in human created objects and this is the challenge you need to prepare yourself for. Understanding commonality is in a way a reverse design skill. Designers are trained to see mainly the differences in order to facilitate their own abilities to create differentiated designs. The centre of designs spaces are full of common and known designs. The training of designers is about how to move away from this centre – so that they may claim as an unexplored space in design space, by creating desing in that space. The work process developed to achieve this is what design training is all about. Designers are taught how to get to the outer edges of the design space through their own creative imagination and work processes. This needs to be reversed. And this is the hard part. Generative design exploration is also about exploring the outer edges of the design space – where interesting possibilities lie. But we need to start from the centre; because we use the sheer dumbness of computers to reach out to the outer edges of the design space. The dumbness of computers becomes a virtue here, as it will get us to the far corners of the design space without difficulty. Now this has a very important implication on how we should structure genetic models.
We should structure genetic models to be in the centre of the design space – in that, it should represent the most common form of design. It need not be creative. It should not be, because if it does it is not in the centre.
Generated spoons If we plot instances of existing designs, we will find a dense cloud of points at the centre, representing the vast majority of designs that share similar attributes. In the centre, you will find well worn designs that satisfy all the constraints imposed by reality. The centre is a result of the deliberation of multitudes of designers who have laboured to find a safe ground and it is from here our exploration should commence. Computational methods in design should help us reach much further than what designers can manage with their limited creative capabilities. The future of design will belong to those who master this ability.
Mantras – for structuring genetic models
Structure it to explore design space Make it highly variable Conceive geometry in layers Less parameters are more expressive Cover a good part of all viable designs Some should surely be weird Start with the commonest configuration Encapsulate design intent Embody geometric logic Learn by trial and error