What are the possibilities for Python and deep learning to connect with architectural design?

I've been paying attention to this problem for almost a week, and so far I haven't found any big surprises. I feel that the architectural design community still needs to learn one. Don't see that deep learning is very popular, just make a big news and move this thing to architectural design.

In fact, how can deep learning be linked to architectural design? If we only refer to "deep learning", then my understanding is that many layers of artificial neural networks are applied. What role can this technology play in architectural design? I really haven't found an example of directly using deep learning technology to assist architectural design. However, if the concept of the subject question is looser and becomes "how to use machine learning and other algorithms to help architectural design", then I think there are still better examples.

machine learning technology is a technology used to improve the running performance of programs with the accumulation of input and time. For example, if you write a program to predict the house price of a villa, the function of this program is to predict the house price according to the parameters such as [city, street, location, area, apartment type, ...] of the input data. In order to improve the accuracy of the program prediction, you need to feed some data with the evaluation results to the program first, and then the program can predict the house price by itself after learning certain data. So, what is the connection between this way of program work and our workflow of architectural design? When we do design, we also investigate and refer to a large number of cases of similar buildings first, and then we can start designing by ourselves after accumulating a certain amount. Knowing this, we can imagine an idea of using machine learning to assist architectural design: let the program learn the previous architectural design scheme first, and then the program can do the design itself!

this 21 paper, Computer-generated residential building layouts, can be said to be a model of the above ideas. The author claims to have "designed a tool that can automatically create the architectural layout with very good visual effect by using data-driven method", but I think the quality of the paper is far beyond such a humble description, because the idea of realizing the paper greatly draws lessons from the design process of architects in reality, and uses Bayesian network, a very beautiful mathematical model, to describe the distribution of an architectural program in space, and all the training data of Bayesian network comes from the design scheme of real architects. Personally, I think it is more correct to use machine learning to deal with architectural layout problems than Shape Grammar's Procedural Modeling idea in the past ten years.

how did the paper unfold? First of all, the author summarizes the predecessors' work, and shows that the previous algorithms based on exhaustion are not feasible. It is almost as difficult to exhaust such a complicated space allocation problem in 3D space as it is for a monkey to get a Shakespeare work by tapping the keyboard at will. Then, the author criticizes that the Shape Grammar developed by Muller in the 199s is no good, because Shape Grammar is graphic grammar. As we know, programming languages are generated based on a grammar set of finite rules, and such rules and clauses in architectural design are difficult to describe formally, and tend to make the number of grammars infinite. Incidentally, decades ago, when computer scientists tackled the problems of speech recognition and machine translation, they also thought that natural language was generated entirely based on limited rules, but it was later found that it was not feasible, and it was not until later that the research was based on statistics that a breakthrough was made. Then another point that shape grammar is not suitable for architectural design is that architectural design is not a game of playing with graphic transformation, and every architectural space has considerations based on functions, psychology and utility. There have been papers on analyzing the grammatical rules of Wright's fireplace house with shape grammar before, but the application scope is too narrow. So this road is impassable.

The author thinks that making a plan must be based on people's comfort, psychological needs, habits and social relations, and then a more reasonable result will be obtained after comprehensive consideration. For example, the shape of the room is better to be convex rather than concave, because in lighting, furniture placement and formal shape above the line of sight are better; Each function of the building should form an openness-privacy gradient, because this reflects the way the building is used. In order to study a better method, the author went to the architects of an architectural firm to ask them about the working methods of professional architects, and got a particularly useful conclusion: after consulting with customers at the initial stage, architects usually use bubble diagrams to think about problems before drawing detailed plans, while bubble diagrams reflect highly concentrated information of an architectural scheme, including privacy, room adjacency, lighting, owner's preferences and cultural customs.

Bubble diagram is the first step of scheme design in the view of architects, but it is a complex data structure-diagram in the view of computer scientists. Drawing a bubble diagram from the user or the task book is the embodiment of the architect's professional ability. No matter whether this ability is based on long-term training or improvisation, we know that this diagram contains a lot of information. The question now is, how can a computer generate such a bubble diagram?

At this time, it's the turn of the machine learning algorithm that started with this answer. The author uses a probability diagram model-Bayesian network to describe the bubble diagram made by architects. In human terms, it is used to describe the probability distribution of each part in the architectural scheme in space. It seems that it is still not easy to understand, but in a word, we know that this model can let the computer know that the living room, bedroom and corridor are more likely to appear in some positions than in others, and the high probability position is the reasonable position, which makes it easier to generate a good plan.

for the training of Bayesian network, the author made 12 residential building schemes as training data to feed to the program. Now that the program finally knows the basic method of architectural scheme, the following is the bubble diagram generated based on Bayesian network:

After the bubble diagram is done, how can the architectural plane be generated from this thing? After many twists and turns, the author finally used the Metropolis algorithm to solve it. The specific description of the generation process is written in the paper, so I won't write it in detail (in fact, I can't understand it. The next step is to generate a 3D model. The author proudly said that this algorithm he worked on really realized the generation of a multi-storey building scheme, while the previous one could only generate a single-storey plane, with limited effect. The following are some achievements: after the bubble diagram is done, how can the architectural plane be generated from this thing? After many twists and turns, the author finally used the Metropolis algorithm to solve it. The specific description of the generation process is written in the paper, so I won't write it in detail (in fact, I can't understand it. The next step is to generate a 3D model. The author proudly said that this algorithm he worked on really realized the generation of a multi-storey building scheme, while the previous one could only generate a single-storey plane, with limited effect. The following are some achievements:

Some of the six planes are made by people, and the rest are made by machines. You can guess which ones are made by machines. The answer is in the paper. Some of the six planes are made by people, and the rest are made by machines. You can guess which ones are made by machines. The answer is in the paper.

finally, the generated 3D model, and the furniture is arranged by hand. Still quite real!

The author finally mentioned that he is going to apply this thing to more types of buildings, such as office buildings. If this thing is mature, there is no need to worry that architects will not lose their jobs, even now!

finally, let me talk about my evaluation of this thing. In recent decades, the automation of knowledge work has spread all over the world in full swing. Many people are worried about whether low-end labor positions will be replaced, such as driving cars, express delivery, writing press releases, translation, shorthand and so on. This kind of worry is really not without reason! When elegant and concise mathematical models meet the computer, a machine that can calculate violently, they combine to create a series of intelligent machines that can be copied on a large scale. It is still scary to think about it. However, don't worry too much, that is, the number of people in the architectural design industry is still very large. Many times, when Party A looks for someone to do the design, it pays attention to the designer himself. Besides professional ability, communication, personality charm, background, license and other endorsements are also very important. However, it is unclear whether this long-standing industry model in the architectural industry will be subverted in the future. In addition, the program also needs to eat a lot of training data to improve its design accomplishment. At present, all colleagues in architectural design hide their collected schemes as treasures, and it is difficult to guarantee the large and complete amount of data. I wonder if the popularity of BIM will improve this aspect in the future.