The Architectural User experience is a complex intervention conceived from numerous relationships between the context and user behaviors. This project was a collaborative development with The Madras Office for Architects and Designers, Chennai. We find as the scale of the projects increases the architectural interventions become generic as the time taken to manage all the parameters becomes mundane reducing it to mere construction. The project Ugly Design aims to create a generative system that creates feasible assemblages. Design is a process of synthesis, some architects use their intuition while others go through a linear process of design. This process is faster for people who try to repeat, tough for people who would like to innovate, and extremely hard for people like us who would go through a series of iterative processes before arriving at a better outcome. We aim to have computational tools to handle this complexity and also to minimize the usage of resources during and after the synthesis of the design.
Generative systems can be a set of rules, transformations, or a computational model that can understand the relationships provided by the users and adapt them to the designer’s intent. William Mitchell in his book “The Logic of Architecture” gives us functional grammars that form the basis of generative systems. The functional grammars are made up of terminal vocabulary(have known geometries and behavior) and markers(have the essence of the function). The generative systems either create markers and replace them with terminals or create terminals directly onto the digital space with the given set of conditions.
In practice, Architects tend to create floor plans through an iterative process by considering certain contextual and user parameters. The Contextual parameters include Site conditions, Climate, and Privacy. The ByLaws provide the rules that define the height and volume that can be built. The User requirements show us the types and number of rooms with respective functional standards. The Architects also need to consider the availability of the resources and the ability to construct. The overall process can be split into Organizations( the floor plans are generated only based on the functional standards), Zoning(the organizations are optimized based on the Climate and other contextual parameters), and Layouts(the organizations are optimized based on the resources available and for structural stability). We analyzed the existing products in the market with their level of output and we found ourselves in need of a set of tools that aid us in all these stages separately and also coherently work together. This leads to a diffused process where the architects can intervene at any stage and mend their outcomes with respect to their intent. The tools can be reframed in any order to create different forms of process and allow the user with greater freedom to explore.
Rectangular Dual Graph
In the preliminary stage, I used an adapted model inspired by Krishnendra Shekhawat, Nitant Upasani, Sumit Bisht, and Rahil Jain which applies a series of transformations on an adjacency graph that represents the space syntax. The relationships are given in terms of an adjacency graph. The input graph is triangulated then the edges are labeled with horizontal and vertical directions. The resulting graph is split into vertical and horizontal graphs which are converted into rectangular dual graphs and merged to get the resultant dual graph. The doors are placed with the original connections. The model generates markers in the form of rectangles that are replaced with rooms later and also generates markers for door positions. The details are added later to the selected organization. The model is inefficient in terms of mapping the relationships from the parameters, like climate and resources used, to the markers and also doesn’t have the functionalities to generate complex terminals. The optimization time to generate and select the variants is huge thus making it inefficient in terms of interaction and limiting our imagination.
Packing models
The dual graph models struggle to address the variations in the terminal. In general, all shapes can be dealt with as an array of positional attributes. Thus I used boundaries as markers for the terminal. The model takes in spaces of different boundaries and packs them in a given overall volume. Each boundary is taken in and the Minkowski difference is applied with the Site shape and the Minkowski Sum is applied to the placed boundary to find the available positions within the volume and placed on the site. The outputs are then evaluated with climate and contextual parameters. Due to the character of the model, it always finds minimal coverage required. The terminals and markers are not differentiable in such models as it takes the boundary of the shape directly to compute. This makes the model heavy in terms of computation and denies real-time collaborations. The terminals are detailed before the process thus it fails to aid in generating design variants.
Discrete growth
The discrete growth concept was used to generate Single-unit organizations. The modules were defined by the input parameters like room dimensions. The Adjacencies were also provided for each room. The discrete growth algorithm takes in these modules and organizes them with the site's shape. The model sequentially places these modules. The Organizations are then optimized with Coverage and climate in the subsequent stages.
Ray tracing
An adjacency tree that defines the rooms and the circulation is determined by the input parameters. Each node is given a weight that represents the functional area that is required for the room. The adjacency tree is then pruned by adding the weights of the branches to its predecessors. At each step, the details are added by navigating to each branch. For example, the weight of the Bedroom and its toilet is stored in the living room. Then during the process of organization, the Area is allocated for the whole bunch, and in subsequent stages, this area is further divided into smaller areas for bedrooms and toilets. This model initiates random points within the shape of the overall volume and then shoots rays on all sides to scan its environment. The Points where the ray hits the site are stored and their bounds are calculated. Within the bounds, a rectangle is drawn with the area required for the room.
In general, the clustering model uses Euclidean distance for its process. However, the cluster outcomes are mostly in the form of spheres as they measure the radial distance for two points. A sphere formed using the Chebyshev distance as a metric is a cube with each face perpendicular to one of the coordinate axes, and a sphere formed using the Manhattan distance is an octahedron. This concept was used to create varied shape clusters with the requirement.
Chev By Chev Clustering
The previous models discussed above gave freedom us to define our terminals. However, the models were not able to bring in more relationships we had to rely on evolutionary optimizers to evaluate the outcomes. To have an efficient system to generate we need to cater to varied markers for varied terminals or that can generate terminals directly. The model must be able to function at an atomic level to create markers of any terminal. The points form the basic unit of a cartesian space and all the terminals invariable to the character can be converted to point clouds. Thus we need a generative system that can manipulate the points to create detailed markers on all scales from furniture to rooms to buildings. The adjacency matrix is efficient in addressing the relationships between the terminals while it doesn’t include relationships from the context. The problem was the graph’s dimension. I had to add some weightage to each node representing the component to capture relationships on multiple levels. These weights and the relationships defined by the inputs can be used to manipulate the points. Kohnen’s self-organizing map is a classic algorithm that clusters weighted points that can be adapted to organize our terminals. The Euclidean distance in clustering was replaced with Chev by Chev distance to create rectangular markers if needed. This model can be used in all scales but we had to adjust their tolerance level respectively. Each node takes in information about the standards required for the function to happen, their neighbors, and weights representing their behavior to the context. These wights also determine whether they are open or closed spaces as well. To reduce the computational time the relationships are converted into a large weighted adjacency graph and then a downward pruning is done adding the weights to the resultant branch. This process allows us to get a basic structure of the whole and each branch is used in successive levels of generation. Thus we can use the same method to address different scales and typologies.
The above-mentioned chev chev model can be used to generate toilet floor plans that vary in boundaries. The model tries to adapt the given relationships concerning the given context. The minimum and maximum data required for each component are predefined along with their adjacencies as the Wash areas should be placed towards the entry, and The bath areas to be placed at the end with the WC in the middle. The point cloud is generated from the overall volume and the spaces are clustered. After the clusters are optimized they are converted into meshes of the same volume. In the next stage the components respective to the Organizations are oriented within the volumes and the Layouts are generated. The generation can be used at different scales and different shapes of the toilet.
The same model can be used to generate markers that represent layouts for a single unit of house or a single unit of apartment. The model has predefined values of the minimum and maximum area required for each room for the function to happen and the adjacencies are defined based on circulation and privacy. The Overall volume is defined by the contextual parameters, ByLaws, and other micro conditions. Since the algorithm works primarily with point clouds the tool can be used to generate layouts of varied typologies and varied site boundaries. The input volumes can be changed at any point in the process and the outputs can adapt to the present condition.
The multiple-unit models show the most complexity in organizations. The model is predefined with areas and adjacencies for each floor. The units are then optimized with respect to the climate and privacy within the units. The multi-unit model can also apply to larger neighborhoods without loss in complexity. However, the computational time required to optimize such a model is huge and requires larger resources to arrive at better solutions. This issue can be rectified with the use of Supervised models that are prioritized with the characters. The apartment model can vary with the number of cores increasing the privacy or number of units. The Multi-unit model can address units of varied inputs simultaneously.
The models that have been discussed have their advantages. But they lack in translating all the relationships that make up a terminal. As designers, we need to have a diffused model that gives us space to steer the process of generation rather than having a black box that just optimizes the inputs that we provide. When comparing the previous models we can find that models that generate markers are efficient in generating as they take less time to create but take more time to optimize for the requirement and models that generate terminals directly are efficient in collaborating as the users have the freedom to freeze the details before generation. Replacing the parametric models with a series of supervised models is effective in diffusing the process. After analysis, for such a process of generation, we require three types of models. The primary model that can understand the relationships from the input to create the terminal vocabulary. The secondary model creates markers and organizes them. The tertiary model translates them as layouts. The dataset we use forms the base for the terminal vocabulary. The models are interlinked such that the outputs can be changed at these junctions for the designers to drive the generative process.
Even though we were able to generate volumes based on circulation and functions the chev by chev model was inefficient in utilizing other important parameters like Climate. GANs are perfectly suited for such workflows as they build accurate relationships within the input and output data. The current model is inspired by a cancer detection model that can map multiple channels of images that represent layers of tissue. We used the same concept where the building is represented as layers of floors. The GAN model takes in inputs of size 256x256x22 and maps them to the output of the same size. The dataset comprises a collection of 1000 floor plans taken predominantly from the Indian context and of varied scales. The floorplans gathered are then augmented into the required tensor with 22 channels. Each channel comprises information about the contextual and user requirements with the respective output. The collected floor plans range from 25 sqm to 60000 sqm in area, with the number of floors ranging from 1 to 10. The collected floor plans were usually single-unit modules and we generalised them to be used with multiple units. The data gathered are then stored in a generic format that can be used to train both the models that generate the overall volume as well as the model that generates the organization. The collected floor plans were predominantly within the site area ranging from 200 sqm to 6000 sqm and of double story height. The floor plans of each level were color-coded with respect to their function and aligned vertically. Later on, each floor was captured as an image of size 256x256 with an added value representing its scale with respect to the original area. The pixel values were also carefully chosen such that the information is not lost in the latent space. The same dataset is converted into specific formats to train each model.
Model 1 generates the overall volume that is required to build within the context for the input features. The model is trained with different typologies of houses. The overall volume generated is then connected to model 2 for further detailing. Model 1 also generates vehicular circulation and vertical circulation within the plot. The model further optimises coverage to minimize resources and also can create volumes that are better for the given climate.
The Multi-channel GAN discussed above is trained to generate layouts of single-unit 2-floor houses of varied boundaries. The model is tested with inputs of varied shapes and conditions. The model adapts the organizations with respect to the vehicular circulation and the vertical circulation provided with limited computational time and power. The model also works with other typologies like houses with courtyards and with different inner volumes. Such a model could address higher complexities when applied to multiple sites without the loss of detail.
The Pix2Pix GAN model was first tested for generating toilet layouts for residential spaces. The GAN model takes in images size 256x256x3 as input and maps it to the output of the same shape. The GAN model is trained with 400 toilet layouts color-coded to represent the volume and adjacencies of each component and their relationships with respect to the toilet boundary. The Dataset varied in shape, and areas and was predominantly filled with orthogonal Layouts. The outcomes from the model are precise enough to understand the scale, As with outputs 5 and 7 the area is allocated only for bathing or WC as only either can happen.
This project introduces a Denoising Diffusion Probabilistic Model (DDPM) designed to generate a diverse array of architectural floor plans from a single conditional input image. Representing the third generation of our design synthesis research, this DDPM advances beyond the limitations of our initial mathematical models and a subsequent Generative Adversarial Network (GAN). While the GAN was effective at learning a direct, one-to-one mapping, the diffusion model excels at generalization and creative exploration, producing multiple unique and viable design options that all satisfy the initial constraints.