Scene Description: By introducing AI technology into the architectural design process, researchers from Harvard University provided more reasonable and diversified choices for the generation of architectural drawings and planning, bringing further attempts and explorations for the development of artificial intelligence in the architectural design industry.
Key words: GANs architectural design discipline integration
This article is reprinted from the public account “Core reading” ID: AI_Discovery
Writing | Stanislas Chaillou (harvard design school)
Compile | Martin culvert, Lu Jiaqi
The full text is expected to take 13 minutes
Master plan diagram generated by GAN
As a discipline, AI has penetrated countless fields and can be applied to challenges that industries have previously failed to solve. As I wrote earlier, the use of AI in architecture is still in its infancy, but it promises to reshape the discipline.
So, is this statement true? This paper intends to apply AI to the built environment to confirm the accuracy of this prediction.
Specifically, we intend to apply AI to the analysis and generation of building plans. The final goal is divided into three aspects:
1. Generating architectural plans, such as optimizing the generation of large and highly diverse floor plans;
2. Select qualified building plans, such as providing appropriate classification methods;
3. Allow the user to “browse” the generated design options.
Our approach is based on two intuitions: 1. Creating a floor plan is an extraordinary technical challenge, even though it involves standard optimization techniques; 2. Spatial design is a continuous process, requiring continuous design steps across different scales (city scale, building scale and unit scale).
To make use of the above two aspects, we selected the embedded Generative Adversarial Neural Networks. This model allows us to capture more complexity in the floor plans we encounter and break it down by solving the problem in successive steps. Each step corresponds to a given model and is trained specifically for that particular task. The whole process can finally prove the possibility of human-computer interaction.
Plan is actually a high dimensional problem, at the intersection of quantifiable technology and qualitative attributes. Studying architectural precedents is often a dangerous process, with less rigorous analysis that ignores the richness of the amount of resources available. Our approach is inspired by current data science methods and aims to determine quality floor plans. By creating six indicators, we propose a framework that captures architectural parameters related to floor plans. On the one hand, footprint, orientation, thickness and material can be used to capture the essence of a given floor plan style. On the other hand, planning, connectivity, and fluidity aim to describe the nature of plan organization.
In short, machines, once an extension of pencils, are now used to map architectural knowledge and, through training, to help us create viable design choices.
A framework.
We work at the intersection of architecture and artificial intelligence. The former is the topic, the latter is the method. Both have been reduced to clear and actionable categories.
“Architecture” here can be understood as the intersection of style and organization. On the one hand, we see buildings as vehicles of cultural significance, expressing a style through geometry, taxonomy, typology and decoration — Baroque, Roman, Gothic, modern and contemporary. A careful study of floor plans reveals many architectural styles. Buildings, on the other hand, are products of engineering and science, following strict frameworks and rules — building codes, ergonomics, energy efficiency, exits and procedures — that can be found when we read floor plans. This kind of organizational requirement will help us complete the definition of “architecture” and advance our research.
The framework matrix
We will use Analytics and generative adversarial networks, two major areas of AI research, as research tools.
First, we’ll delve into the topic of generation. We will use GANs to apply our own artificial intelligence systems to architectural design. We hypothesize that the use of artificial intelligence can enhance the practice of architectural disciplines. The field is entirely new, experimental, and has already yielded surprising results. We want to be able to train artificial intelligence to draw real floor plans.
We then propose a powerful analytical framework for selecting and classifying the resulting floor plans. Ultimately, our goal is to organize the results of GANs so that users can seamlessly navigate the various design options created. To that end, the number and ubiquity of tools that data science has to offer will prove valuable.
Through this dual design, at the intersection of style and organization, selection and generation, we developed a framework to arrange the meeting of architecture and artificial intelligence.
Generated by two.
Designing building plans is central to architectural practice. Mastering floor plan design is the gold standard in the discipline. Practitioners are working overtime and constantly trying to improve the practice through technology. In the first part, we delve into the potential of artificial intelligence applied to floor plan generation to further advance the field.
In order to address the style and organization of floor plans, we will use a framework to explore the potential of space planning based on artificial intelligence. Our goal is to provide a robust set of tools to demonstrate the potential of this approach and test our hypotheses.
There are three challenges: 1. Choosing the right toolset; 2. Separate the correct phenomena and show them to the machine; 3. Make sure the machine “learns” correctly.
Artificial Intelligence and generative adversarial Neural Networks
Generative adversarial neural networks (GANs) are our weapon of choice. Neural network is an important research field in artificial intelligence. Recently, the emergence of generative adversarial neural networks has demonstrated the creative power of this model. As a machine learning model, GAN can learn statistically important phenomena from given data. However, their structure represents a major breakthrough: GAN consists of two key models, generators and discriminators, and the feedback loop between the two models can be exploited to improve the ability to generate relevant images. Discriminators are used to identify images from data. With proper training, the model can distinguish between real images extracted from the dataset and “fake” images unfamiliar to the dataset. Generators, however, are used to create images that are similar to images from the same data set. As the generator generates an image, the discriminator gives it feedback on the quality of its output. In response, the generator takes this feedback to produce a more realistic image. Through this feedback loop, GAN slowly builds up the ability to create relevant composite images, taking into account phenomena found in the data.
Generate structures for adversarial neural networks
Presentation and Learning
Knowing what to show them is crucial if GANs presents us with great opportunities. We had the opportunity to have the model learn directly from the architectural plan. By formatting the image, we can control the type of information the model learns. For example, showing our model the shape of the lot and the associated building footprint gives us a new model that creates a typical building footprint based on the shape of the lot. To ensure the quality of the output, we will use our “architectural sense” to organize the content of the training set: each model is only as good as the data provided by the architect.
A. Style change
The transition from modern to Baroque
In architectural plan, the “style” of the wall can be judged by its shape and graphic plane. While typical Baroque churches feature heavy columns and circular indents, Mies van der Rohe’s modern villas feature thin, flat walls. GAN can recognize zigzag shapes on wall surfaces. By showing these two images and using one as a floor plan wireframe and the other as the actual wall structure, we can establish a certain amount of machine intuition based on the architectural style.
This section presents a diagram of the results of model learning in baroque style. We continue the style transformation by grinding down the thick walls of A given building plan (A) and then giving them A new wall style (B).
The result of style change: The transformation of the apartment units from modern to Baroque
B. Layout assistance
Layout AIDS in a step-by-step process
In this section, we have prepared a multi-step flow chart that contains all the necessary steps to draw an architectural plan. It mimics the architect’s plans for different building scales and compresses each step into a specific model for performing a given operation. From plot to architectural map, from architectural map to room map with walls and Windows, from room map to hardcover floor plan, every step has been carefully designed, trained and tested.
Generation process (Model I to III)
At the same time, the system allows for user interference between each model by dividing the entire process into several separate steps. By selecting and editing model outputs, users can control the design process before the outputs go to the next model. Its input determines the decisions of the model so that the expected human-computer interaction can be realized.
1. Covers an area of
Surrounding | block (input) | generated covering (output)
The first step of the process solves the challenge of creating an appropriate building occupancy map based on a given plot geometry. To train the model, we used an extensive database of Boston building occupancy maps and created a series of models, each matching a specific property type: commercial, residential, condominium, industrial, etc.
Each model creates a set of related architectural maps for a given plot, similar in size and style to the types in training. Here are 9 examples using the housing model:
Result: Generated footprint (residential)
2. Room division and window opening
Covers an area of | opening & balcony (input) | plan & window (output)
The next natural step is for the building to dominate the layout of the rooms on the map. It was a challenge to be able to segment a given floor plan while retaining meaningful articulation, normal room dimensions, and proper windowing. GANs was able to work it out, with surprising results.
Using a dataset of about 700 annotated floor plans, we were able to train models. Each is oriented towards a specific number of rooms, and can produce unexpected and relevant results when used on a blank building map. The graph below shows some typical results:
Result: Generate plan & open window
3. Furniture
Plan (input, option 1) (input, option 2) | | furniture position furniture configuration unit (output)
The final step takes the generative principle down to its most subtle level: adding furniture to the space. To do this, we first trained the model to complete the furniture configuration of the entire apartment at once.
The network can learn based on the plan of each room, the relative arrangement of furniture in the space and the size of each element. The results are shown below:
Result: furnished units
These results give us an idea of possible furniture layouts, but the quality of the resulting drawings is still too vague. To further improve the output quality, we trained an additional set of models for each room type (living room, bedroom, kitchen, etc.). Each model can only transform the color blocks on the floor plan into carefully drawn furniture, with furniture types numbered by color codes. The results of each model are shown in the figure:
The result of the configuration model room furniture | bathroom/kitchen/living room/bedroom
4. Take it one step further
If the technology can be used to create a standard apartment, the next natural step is to push the limits of the model further. In fact, GANs has remarkable flexibility to solve seemingly difficult problems. In the layout of the architectural plan, it is a very challenging process to divide and decorate the space manually due to the varying size and shape of the building on the map. In this case, our model is very “smart” and can adapt to changing constraints, proving as follows:
The spatial layout under GAN
We can control the entrances and Windows of the building units, and the model is very flexible, so we can go beyond the individual building units to solve the spatial planning problems on a larger scale. In the example below, we can extend the application of technology to the entire building.
The experimental master plan diagram generated by GAN
Three. Select qualified floor plans
Not being able to name it adds to the chaos.
– Albert Camus
In order to balance our ability to generate architectural plans, it was critical to find the appropriate framework to organize, order and categorize the generated design options. Building plans are only as good as our ability to manipulate a database of generated options. By borrowing architectural concepts, we hope to transform common architectural phrases into quantifiable indicators.
To this end, we established six key indicators to describe the six important aspects of floor plan design: footprint, plan, orientation, thickness and texture, connectivity and fluidity.
Six indicators
These indicators form a comprehensive framework that addresses the stylistic and organizational issues of the floor plan. Each metric is an algorithm and has been thoroughly tested.
A. covers an area of
Building shape is the simplest and most intuitive index used to determine building style. “Footprint” is a metric that analyzes the shape around a building’s floor plan and converts it into a histogram.
The descriptors encode a building’s shape, converting architects’ commonly used adjectives – such as “thin”, “bulky” and “symmetrical” – into digital information that can be communicated to a computer.
Of the map
Technically, it uses extreme convexity to transform a given profile into a set of discrete values (vectors) that are then compared with other building plans. We use an array of polar lines from the center of a planar graph to extract planar regions captured by spatial slices. As you can see in the figure below, this approach has proven to yield good results. The technique can also be used to determine the shape of interior Spaces and surrounding buildings.
Typical building plans using footprint indicators (left: Query, right: Results)
Plan B.
The building plan, in other words the type of rooms contained, is the main driver of the internal organization. Understanding this is critical to our approach. To describe the combination of rooms, we coded the rooms on a given floor plan with color. These color combinations become indicators that describe building plans. It acts as a template, blending the number of rooms on the floor plan with the quality of the plan. This metric may seem intuitive to humans, but it can also be translated into reliable coding techniques that machines can understand.
Typical building plans using planning indicators (left: Query, right: Results)
Technically, by using color combinations, we can calculate plan similarities and differences between any given floor plan. To visualize the results, each plan is both a colored architectural plan and a one-dimensional color vector of the plan.
Typical architectural plans using plan indicators (left: Query, right: Results, bottom: Results’ Program)
C. toward
The wall orientation in the plan can provide important information. It can describe both the closure of the floor plan (the spatial closure caused by the orientation of the walls) and the style. In fact, we can easily distinguish modern houses from Gothic cathedrals by extracting the histogram of the wall orientation.
Toward the figure
Technically, orientation extracts walls from a given building plan and sums up their length along various spatial orientations (from 0 to 360 degrees). Evaluate the overall orientation of the floor plan and arrive at a set of values. This index can be used to obtain a single descriptor or as a vector to compare planar graphs.
Typical architectural plans using orientation indicators (left: Plan, right: orientation)
D. Thickness and material
The index “thickness and material” determines the “fat and thin” of the floor plan: wall thickness and thickness variation. Wall thickness and wall material vary greatly in different styles of floor plans. Thick columns and jagged walls can be seen in the halls of the academic style, but in the villas of Mies van der Rohe only thin walls are seen in straight lines. This metric can easily grasp these characteristics (as shown in the figure below) :
Thickness and material
Technically, this metric can isolate all the walls in a given plan and output a histogram of the wall thickness. The algorithm can also calculate changes in wall thickness to better describe wall materials (such as flat walls and mullions).
Typical Building Diagrams with Thickness and Material Indicators (LEFT: Plan, right: Resulting Diagrams)
E. connectivity
The “connectivity” metric solves the problem of contiguous rooms. The proximity of rooms to each other is an important indicator in architectural plans. In addition, the rooms are connected by doors and corridors, which determine the connection between the rooms. Connectivity uses room connections as a standard chart to study the quantity and quality of connections.
Connectivity and adjacency matrix
Technically, by opening Windows in the plan, we can infer the existing relationship between rooms. Connectivity then creates a matrix graph to represent these connections. The result is a diagram of the room connections. With this diagram, we can compare the different building plans and consider the similarity of their room connections.
Analysis of typical building plans using Connectivity Indicators (left: Connectivity Graph, right: Plan Adjacencies)
F. liquidity
Fluidity in the floor plan describes the movement of people through space. By extracting the skeleton of the circulation, in other words the wire frame of the circulation network, we can quantify and determine the movement of people on the floor plan.
liquidity
Technically, ‘mobility’ extracts the degree of circulation of a given floor plan and sums up its length along all directions of the space (from 0 to 360 degrees). The resulting histogram evaluates the shape of the circulation network and can be used to compare the fluidity of different building plans.
Analysis of typical building plan using circulation index (left: Circulation Graph, right: Diagram)
Draw and browse
A similar plan obtained by comparing plans
Reviewing our GAN model, each model outputs multiple options at various steps in the generation process. Then the designer needs to “pick” the desired option and improve it as needed before moving on to the next step. However, browsing through the generation options can actually be an uncomfortable and time-consuming process. To that end, the six metrics described in the section “Selecting a Qualified floor plan” can be used to their full potential here to complement our generation process. Users can use them as filters to narrow down their choices and find relevant design options in seconds. The duality of generating filters is further demonstrated by the value of our work: we provide a complete framework and leverage ARTIFICIAL intelligence within the reach of the average user.
Once the model is filtered according to the given criteria (footprint, plan, orientation, thickness vs. material, connectivity vs. flow), we provide the user with a tree to show his/her choices. The selected options are in the middle of the tree, surrounded by the options closest to the user’s selection criteria. Users can then narrow their search to find the most desirable design option, or recalculate the graph by selecting other options in the tree.
A similar tree diagram of a building plan
In live.
Artificial intelligence could soon help architects in their everyday practices. With this potential about to be proven, we participated in a proof of concept, and our framework provided an opportunity for discussion. We invited architects to participate in AI and suggested that data scientists should consider architecture as a field of study. But our work today can be summarized in the following four aspects:
First, conceptually, we believe that statistical approaches to design concepts determine the potential of AI in architecture. Its uncertainty and comprehensiveness certainly create opportunities for our field. Relying on them to extract important qualities and mimic them throughout the design process, rather than using machines to optimize variables, is a paradigm shift.
Second, we firmly believe that the ability to design processes correctly will drive AI as the new construction tool. We chose the “grey box test” introduced by Professor Andrew Witt in the Log, which is likely to yield the best potential results. “Gray box testing” is the opposite of “black box testing,” which only allows users to enter information in advance and only get design options at the end of the process, with no control over successive generation steps. Instead, by breaking the process down into individual steps, “gray box testing” allows users to intervene at any time. Strict control of the machine is the ultimate guarantee of the quality of the design process.
Third, on a technical level, we believe that the sequentiality of applications improves manageability and promotes growth. It is important to be able to intervene throughout the generation process: each step of the process represents a different block of architectural expertise, and each model can be trained independently, enabling significant improvements and experimentation in the future. In practice, improving the entire process from start to finish can be a long and tedious task, but incremental improvement is a manageable process and is within the reach of most architects and engineers in the industry.
Finally, we want our framework to address the infinite breadth and complexity of models, whether they require training or are part of the generation process. We believe that, among the many options, dealing with lots — floorspaces — room division, as we did, is the way to go. Principles, not methods, are the key to outlining the necessary steps of spatial planning. As building data becomes easier to access, we encourage people to open their minds to further research and experimentation.
Rather than seeing AI as a new dogma in architecture, we see it as a new challenge full of potential and promise. It is possible for us to achieve fruitful results here, which will enrich our practice and solve some blind spots in our discipline.
