AI and Machine Learning in Architecture

In recent decades, computation has given architecture a greater edge in understanding structure, easily solving complexities, and making better use of resources.
The evolution of the use of computers or computational design in architecture is divided into eras: the first era is that of 2D drawing, the second of 3D modeling, the third era is that of the BIM, the fourth of the design based on algorithms and finally, the fifth era of the Machine Learning (ML) and of theArtificial intelligence (AI).

Before understanding the impact of ML e AI, it is important to understand the evolution of the application of calculation in simpler terms. 2D drawing (software like AutoCAD, Sketchpad, etc.) involves drawing lines with the help of calculation, in 3D modeling (software like Blender, Maya, etc.) surfaces, volumes, NURBS and meshes come into play. When it comes to BIM (it is a workflow, but the most important software that helps in BIM are Revit, Tekla, ArchiCAD, etc.), an object can store data that can be used productively, for example, a roof once made, data such as its area, volume, physical properties, materials, texture, etc. are stored inside it, they can be used in BOQ, structural analysis, etc.
Algorithm-based design (software like Grasshopper 3D, Dynamo, etc.) came into play after the popularization of the works of Zaha Hadid and parametric movement.
In algorithm-based design, a design is made with some input parameters and let the computer provide the design output according to the direction encoded by the designer.
And finally, the era of machine learning which concept comes out from the late 90s, but its application is happening now, software like Lunchbox, Finch 3D, Sketch Graphs, etc. are developed based on machine learning.

Machine learning (ML), also known as statistical learning, is a type of artificial intelligence that uses a dataset to predict the outcome with a certain percentage of accuracy. It requires a training dataset (the larger the dataset, the more accurate the outcome), based on which ML provides the output. Different mathematical models such as artificial neural networks (ANNs), support vector machines (SVMs), Bayesian networks, Radial Basis Function networks (RBFs), etc. are used in ML to get different types of results.

Here is the list of activities with mathematical models that will generate outputs.

1. Generate the project concept from the client's request

Machine learning will be able to generate the building concept from the description provided by the client as input. The ML model that helps to get such output is Artificial Neural Networks (ANN) that will capture the data patterns and correlations hidden in the data.
ANN will provide many conceptual ideas based on the description and different contexts related to the project. This will allow architects to explore more concepts and find the best one without spending much time on searching and context-based information such as buildings in the site. In just a few years, we may find machine learning models that will generate from the 3D model information that will further increase the exploration and exposure of architectural solutions.
When the architect uses ML to generate concepts, ML also allows the user to get advice in the design process. ML will help the architect clarify his design intentions and recommend choosing the structure types, interior distribution and facade treatment.

2. Mass customization and urban planning

In the case of urban planning and mass customization systems, the style transfer technique using convolutional neural networks (CNN) is useful. CNN is mainly used to analyze visual imagery and uses neural connections, such as our brain's image processing technique from the visual cortex.
To use this technique in the field of urban design, we need to produce the building styles we need, a basic design on which the variations will be performed, and a previously trained panel of neural networks.
Architects need to have control over the output and use ML only as a tool, as this system can produce unwanted variations.

3. Program summary

Algorithms are needed to produce the design at various stages of an architectural project. The order to create a program that solves a particular design intent requires programming knowledge and in some cases, it is difficult to find Application Programming Interface (API) experts to execute the idea. In some recent developments, an API expert is created who will guide, advise and predict the step of the designer who uses it. This type of system can be seen in the auto-completion of texts while using G-mail. The ML-built API expert will suggest to the user which API is best suited for a particular problem, identify bugs and fix them and finally learn algorithms that solve a particular problem.

4. Analytical modeling

Modeling is now an integral part of architectural projects, it allows us to design a building in detail, run simulations with calculations, look for conflicts between the various parts and prepare the views before the building is built.
These processes take a long time when the geometry becomes complex and heavy for the file to process and run simulations. For this reason, we need to have low poly models of the heavy models prepared by us and reduce the size and number of geometries in the file for faster simulations. These processes could be done with artificial intelligence and this system will also isolate different components such as beams, columns, slabs, etc.
Similarly, geometric component extraction could be performed from satellite imagery for urban scale modeling and planning work.

5. 3D modeling, production and classification

When it comes to production and classification tasks, Generative Adversarial Network (GAN) model is the appropriate model to perform the task. GAN is a ML model, developed by Ian Goodfellow, which is capable of unsupervised, semi-supervised, fully supervised and reinforcement learning. The GAN model can be trained with data consisting of space names, their geometry, etc. based on which it will generate and classify appropriate spaces. Running this model with style transfer will generate a better model that will be easy to understand.

6. Find hidden correlations

While evaluating a design based on certain parameters using a ML model, it becomes important for the model to separate the data used for evaluation from the rest of the data, which will improve the efficiency of the system. Models like Surrogate Based Optimization (SBO), ANN, SVM, Gaussian Processes (GP), Radial Basis Functions (RBF) and some ready-made tools like Opossum, Lunchbox are used for data approximation, isolation and evaluation. This will allow faster and more meaningful simulations and can generate model solutions useful for design.

The application of machine learning and artificial intelligence is all around us and the applications in the field of architecture can be many. Objectives such as developing an adaptive control system of the facade, creating the desired materials and developing more intelligent systems can be achieved with machine learning.
The introduction of these systems will not sacrifice the creativity of the architect but will allow to explore many more solutions, find the best and fastest simulations, automatic organization of files and much more. Now it is only a matter of time, then we will only need to ask the right question to get the best out of machine learning.

With all this technology, what are the disadvantages?

The main disadvantage of applying machine learning in architecture is that the instruction set of the systems must be continuously updated, otherwise it will always generate a similar result. Unfortunately, we do not have a large training set available in architecture. An unbalanced dataset may make the ML biased towards “peculiar” decisions and research to minimize these kinds of decisions is ongoing. There are several anomalies related to ML that will need to be solved and without these “adjustments” the application in architecture can be disastrous.

The document "The Role of Artificial Intelligence in Architectural Design: A Conversation with Designers and Researchers” by Giuseppe Gallo, Giovanni Francesco Tuzzolino and Fulvio Wirz includes interviews with leaders in the architecture industry and attempts to understand the impact of AI in architecture. It is quite interesting to see that out of 10 leaders interviewed, 4 industry leaders put Machine Learning and other computational methods and Digital Manufacturing at the top of their list of technologies that will prove useful in the architecture field in the next 10 years.
Other technologies reported are Building Information Modeling (BIM), Internet of Things (IoT), Augmented Reality (AR) and Virtual Reality (VR). So, it is quite clear that in the field of architecture the impact that machine learning and artificial intelligence will have in the next 10 years will be very strong.

There are many architecture tools that use ML and AI technology to some extent and have become part of our daily life. Google Maps, Google Email Filters, Linked In, Google Search Algorithm, etc. In architecture, some of the software that use ML and AI are Unity 3D (which uses AI to find the shortest distance of fire exits), Lunchbox (uses general ML for regression analysis, clustering, and networks), Opossum (uses ML for functional evaluation, near-optimal solutions, and simulations), Autodesk’s Project Dreamcatcher, etc.

Being surrounded by technology and foreseeing the possibilities of the future, the question “what is the next step?” remains the same throughout the centuries and this is the difference between AI systems and humans.