Hao Zheng

Architecture + AI: where the future stands

1. Why architecture needs Artificial Intelligence

1.1 From discreteness to parametric design

Architecture is known as the crystallization of human civilization. In the long history, architects have been looking for a design method to summarize and practice architectural design strategies. From a historical point of view, design rules have become more and more complex, and more and more computing power is required to support the execution of design rules. But in fact, the core concept of architectural design is the same: it is discrete and can be expressed by states and rules. The state is the measurement in the building, and the rule describes the method of transforming one state into another. The final architectural design plan is represented by a collection of states, which shows the metric data in the building.

With the rapid development of computer science, the concept of discrete architecture has been widely used in practice. In the field of architecture, with the digitization of architectural data, design methods based on two-dimensional hand-drawn drawings have gradually evolved into three-dimensional digital modeling. Subsequently, the algorithmic design that appeared in the 1990s opened the door to digital architecture. Architects can complete more complex designs in a shorter time by converting design rules and states into programming codes and data.

Left: The Universal Constructor, Middle: Evolution of Tuscan columns, Right: Morphogenetic Design Experiment
1.2 Decisions by designers VS decisions by AI

As Nicholas Negroponte defined in his book “The Architecture Machine”, the machine design assistant should be able to automate current processes, convert compatible processes into designs, and iteratively calculate the design data. There is no doubt that the current design algorithm has reached the standard of a design assistant. However, when we discuss the changes in the design industry brought about by computing power, computers are still only regarded as calculators, which only generate designs based on the rules given by designers, and do not have their own design logic. Designers still rely on their instincts to lead a design.

In recent years, with the development of artificial intelligence (AI), computers have been widely used to find solutions to real-world problems, map information from the real world to quantitative data, and find the relationship between them. Although the problem-solving process of artificial intelligence is different from the traditional perceptual understanding of architectural design, their starting point is the same, they summarize and apply rules.

Therefore, in this era of artificial intelligence and big data, we discuss possible integration points in the architectural design industry. In architectural design involving artificial intelligence, designers provide cases to represent the initial and final state of the design, and computers build machine learning models to fit design rules, and then apply the trained models to the generation of new designs. Computers will not only be design assistants, but also design collaborators, helping architects make decisions based on design rules.

Computer-Aided Design VS Computer-Decided Design
1.3 The breakthrough point of architecture + artificial intelligence

In the field of architectural design, the application scope of artificial intelligence includes three aspects: design cognition, design generation and design acceleration. The complexity of each design is different. Therefore, when the difficulty of design tasks increases, the main difference between artificial intelligence and human designers becomes prominent. Artificial intelligence has the powerful ability to learn and achieve complex design goals, but human designers may not be able to navigate beyond their knowledge and imagination.

For example, in engineering-related problems, there is a descriptive and relatively simple objective function that can be easily generalized and summarized as a design strategy by both human designers and artificial intelligence. In the field of architectural design, the objective function is still descriptive, but it is more complex than the objective function in the engineering field. Faced with the complex design problem of ambiguous objective function, human designers have limitations. Finally, when a human designer encounters a complex logic that is not enough to create simple rules, artificial intelligence can produce design cases beyond the capabilities of human designers after training on design data, and generate design cases that human designers can hardly describe. The powerful learning ability of artificial intelligence will bring new breakthroughs to the design industry, making design a descriptive, controllable, and optimizable process.

2. How artificial intelligence solves architecture problems

2.1 Spatial layout problems

Considering the specific application scenarios of artificial intelligence in the construction industry, the first thing worth mentioning is the issue of spatial layout. When architects arrange the space, they often design relatively well-founded designs based on certain industry consensus. Therefore, most of this “layout” task is supported by a logic behind it, and the architect completes it according to a well-documented rule. Therefore, the task of designing a floor plan is very suitable for the participation of artificial intelligence. In one of our research project in 2017, we implemented the use of image-to-image neural networks (GAN) to train the artificial intelligence model, allowing it to automatically generate internal layout drawings based on the boundaries of the floor plan.

Machine learning from floor plan boundary to design drawings [1]

However, the shortcomings of GAN are also obvious. This kind of “one-to-one” neural network can only provide one plan view based on a plan boundary, but architects can often design a variety of plans for clients. In order to solve this problem, we have carried out vectorization of the plan drawings, reconstructing the plan drawings data from the perspective of graph rather than image, so that the expression of the data itself is closer to the essential abstract logic.

Vectorization of apartment floor plan [2]

Based on the vectorized data, we took the generation of apartment flop plans as an example to train a hybrid neural network. Our artificial intelligence model no longer only understands a plan from the image, but really learns the vector logic behind the plan, and generates a variety of plan layouts in a mode closer to the thinking of architects.

Multi-solution of the plan layout problem [3]

Under the usage of this hybrid neural network, the program can automatically generate up to 200 different floor plans for a user-inputted apartment boundary, which covers various configurations, such as one bedroom and one living room, or two bedrooms and one living room, and so on. There are also some unique solutions, such as widened balcony space or larger kitchen, etc., which can fully meet the different needs of users.

Generating up to 200 different apartment plans for an input boundary [3]

For the spatial layout problem, what we have given above is a solution based on big data and machine learning. By replacing different data, such as the floor plan of a commercial building, we can train artificial intelligence to learn and generate different types of building plans.

In addition, another solution is based on evaluation functions and a multi-objective optimization algorithm. Here we introduce this kind of solution with the current popular residential layout problem as a case. The residential layout problem can be understood as placing multiple buildings in a space so that they meet a series of mandatory requirements while also bringing the greatest advantage. According to the specific needs of users, we write these mandatory requirements and evaluation of advantages into mathematical formulas, and use artificial intelligence optimization algorithms to find the optimal solution among the formulas, and then generate a layout plan.

Another solution of space layout problem: optimization based on evaluation function [4]

As we can see from the optimization process shown in the animation below, the program first reads the site and building information preset by the user. The algorithm gradually converges to several feasible solutions after randomly trying a variety of strong ranking possibilities, and finally finds the most optimal solution. This optimal solution takes into account the landscape advantages brought by the surrounding rivers and the requirements of daylighting distance, etc., forming a layout plan.

The optimization process of the layout in the residential area [4]

In terms of spatial layout, data-based and algorithm-based solutions adapt to different conditions. When we choose a solution, we should evaluate the difficulty of data acquisition and the difficulty of algorithm design, and then customize the technical means to adapt to different tasks.

2.2 Urban prediction model

Also based on the machine learning of floor plans, the second type of application focuses on urban information systems, training artificial intelligence to predict various indicators in the city, and assisting designers or city operators to improve the urban environment. First of all, in the age of advanced GIS systems and information explosion, collecting city data is no longer a difficult task. If the big data is not filtered, irrelevant data will be mixed into the artificial intelligence training data, which will directly lead to a decrease in the efficiency of training and a decrease in the accuracy of prediction. Therefore, screening the data that needs to be collected with expert knowledge from human designers is the first step in training AI in urban design.

Urban data collection and visualization [5]

The artificial intelligence model with completed training process has predictive capabilities, and what it can predict is the type of data we filter for it during training. For example, in our latest study, we used the positioning function of the mobile app to collect a large number of users’ walking and cycling data. We presented the data on a flat map through the heat map model, and trained AI to predict the number of times users pass by each area based on the city map. The larger the predicted value is, the more people will pass by walking or cycling, which brings more vitality to the community.

City vitality prediction model: Is your community alive? [6]

Based on the same technology, we can also accurately predict the urban crime rate and determine the probability of crime in each street or even each building. Through the analysis of all the alarm record data in the past ten years, we can generate a crime rate heat map, which intuitively represents the total number of crimes in each community in the city. When designing new cities or transforming old cities, artificial intelligence models can predict the crime rate of current urban design based on the results of big data learning, and provide suggestions for improvements that can reduce the crime rate from a design perspective.

Urban crime rate prediction model: Is your community safe? [7, 8]

Under the combined effect of the two artificial intelligence models, we have established a set of prediction systems to help users when deciding where to live (such as buying a house), using the power of AI to find the most suitable residential area to ensure that the crime rate is low while the vitality of the community is higher. When we train AI with different types of data, this technology can be customized to predict any city indicators to assist users in decision-making.

City vitality and crime rate prediction model: calculating the best place to live [6-8]
2.3 Form-finding task

Artificial intelligence can learn not only two-dimensional image data, but also three-dimensional model data to assist designers in three-dimensional creative tasks. In the machine learning of 3D data, our main concern is the stylized generation of architectural forms. First, we tried a learning model based on big data, collected nearly 200 sets of architectural digital models, and vectorized the data to form a point cloud model.

Collection and vectorization of architectural form data [9]

We then trained a set of neural networks to generate the three-dimensional shape of the building based on the boundary of the building site input by the user and the desired architect style. Users can freely choose or add new building data to improve the artificial intelligence model, allowing AI to learn and generate the architectural form of their favorite style.

Rapid stylized generation of architectural forms [9]
Architectural form-finding: results of DF2019 workshop of Tongji University

Style transfer is another artificial intelligence algorithm that reconstructs design styles. It uses pre-trained models to stylize the input design data and generate new design schemes. This is also a kind of artificial intelligence algorithm that designers can use conveniently to find inspiration.

Style transfer: reshaping the architect’s inspiration [10]
Style Transfer: results of DF2020 workshop of Tongji University
2.4 Other tools

In addition to the above three common architecture + AI applications, there are other interesting artificial intelligence models that are involved in architecture-related tasks at different levels. One type of artificial intelligence model focuses on the optimization of the computing efficiency. Here we take a structural optimization algorithm as a case to briefly introduce the application of this type of artificial intelligence model. When we perform traditional structural calculations (such as finite element calculations), we often consume a lot of computing time. The intervention of the artificial intelligence model can fit a small number of calculation samples to the overall solution space, replacing the original calculation model, which can greatly improve the calculation efficiency and realize fast optimization calculations.

Neural network surrogate model: accelerated structural optimization calculation [11]

Another interesting experiment is a tentative development in our laboratory. The goal is to expect that the artificial intelligence model can interact with humans and learn human aesthetic tendencies through human feedback data, thereby helping designers conduct more effective design evaluations. First, let an experimenter observe several sets of patterns and choose one of his favorites. Then we repeat this experiment, by continuously collecting the forms selected by the experimenter, and building a database to train the artificial intelligence model to learn the logic when the experimenter chooses the form. The trained artificial intelligence model can automatically determine whether the shape will be liked by the experimenter according to the subsequently generated shape, and then iteratively find the experimenter’s favorite shape as a recommendation system. In the future, if we replace the morphological data with more general architectural renderings or floor plan data, this system will be able to learn the aesthetic inclination of the design evaluator (such as the clients), and then invest in its preferred generation plan.

Recommendation system: What kind of form your clients like most? [12]

3. What can we do in the future?

3.1 Customized problem solving

Above, we have introduced the possible applications of artificial intelligence in various fields related to architecture. In fact, we can see that artificial intelligence has prompting solutions in various sub-areas, but there is still no method that can completely solve all the problems in the architecture field. Rome was not built in a day. There is still a long way to go for artificial intelligence to become an ideal “master architect”.

However, this does not prevent artificial intelligence from gradually intervening in the architecture industry. Customized artificial intelligence solutions for each problem can have an immediate effect within the scope of application, quickly improve the original process model, and improve efficiency. It is conceivable that the best landing scenario for artificial intelligence in the field of architecture at the moment is customized development for special needs, rather than shouting of the so-called slogan of subverting the industry.

3.2 On this basis, design is improved by human-machine collaboration

Therefore, under the influence of artificial intelligence, the role of human designers in the future is changing, thus forming a new relationship of cooperation and coexistence. Artificial intelligence will replace some functions, such as assistant positions such as secretary, draftsman, and analyst, but it places higher demands on human designers, requiring future human designers to be more decisive and creative, becoming an organizer, decision maker, and creator. Artificial intelligence and human designers cooperate with each other to promote the evolution of human designers to achieve a higher level of design intelligence and design thinking. The ultimate goal of this partnership is to continuously improve the overall level of architectural design, thereby simplifying the original complex architectural design, or exploring a new architectural model to adapt to complex buildings.

3.3 Frontiers of Academic and Development

In order to break through the development of intelligent design in the future, our Architectural Intelligence Group (AIG) aims to deepen the research and application of emerging technologies of artificial intelligence in the fields of architecture, urban design, landscape, and construction. Docking with cutting-edge innovative academic achievements, technical concepts and commercial projects, we are committed to finding a common breakthrough in academic research and technical applications, lead the new technological transformation of AI intelligent design with research, and summary methods to guide practice.

The contents of our concentrated development are mainly divided into the following four types of research: 1) Architectural Intelligence; 2) City Intelligence; 3) Landscape Intelligence; 4) Construction Intelligence.

Focus of AIG

We take into account the characteristics of laboratory (academic research) and brain tank (project applications), and integrate innovative ideas of artificial intelligence technology across multiple fields such as architecture, urban planning, interiors, human-computer interaction, computation, and big data.

Hao Zheng is a Ph.D. researcher at the University of Pennsylvania, Stuart Weitzman School of Design, specializing in machine learning, digital fabrication, mixed reality, and generative design. He holds a Master of Architecture degree from the University of California, Berkeley, and Bachelor of Architecture and Arts degrees from Shanghai Jiao Tong University. Previously, Hao worked as a research assistant at Tsinghua University and UC Berkeley with a concentration on the robotic assembly, machine learning, and bio-inspired 3D printing. His teaching experience includes: workshop tutor at Tongji University; lecturer at Macau University of Science and Technology; lecturer at the University of Pennsylvania; teaching fellow at Shanghai Jiao Tong University. Also, Hao serves as the co-organizer and reviewer for international conferences of ACADIA, CAADRIA, CDRF, and SCI/SSCI journals. His publication includes around 20 papers in top international conferences and SCI journals.

The members of AIG have many years of experience in design, research and development. They are scholars from University of Pennsylvania, University of Oxford, University of California, Berkeley, University College London, National University of Singapore, Delft University of Technology and others with outstanding global education background, and with working backgrounds such as Facebook software engineers, Lenovo software engineers, Microsoft Design Institute assistant researchers, Johnson & Johnson data scientist.

AIG team members

Our academic achievements in different fields mainly involve artificial intelligence and digital design, architectural plan generation, form finding, plan optimization; urban data analysis, urban prediction models, planning optimization; creative plan generation, landscape big data, scene recognition; construction path optimization, computer vision, and virtual reality. Past research papers were published in SCI journals and top international conferences, and included in CAADRIA, ACADIA, eCAADe, and many other international academic associations.

AIG academic achievements

We also provide business consulting and application development related to artificial intelligence technology. Our clients include Baidu, Alibaba, Country Garden, Longfor Group, Glodon, Landleaf Technology, Insome, Tianhua Design, Shanghai Jiao Tong University Design and Research Institute, Kreysler & Associates, SpaceMaker and Brainpool.

AIG business consulting and application development

AIG has deeply cultivated core emerging technologies and innovative academic ideas in related fields such as AI+Design. You are welcome to establish academic discussion and business cooperation with us, and follow our Wechat account to receive our news!

Contact AIG

References:

1.            Zheng, H., Drawing with Bots Human-computer Collaborative Drawing Experiments, in Proceedings of the 23rd International Conference on Computer-Aided Architectural Design Research in Asia. 2018: Beijing, China.

2.            Zheng, H. and Y. Ren, Machine Learning Neural Networks Construction and Analysis in Vectorized Design Drawings, in Proceedings of the 25th International Conference on Computer-Aided Architectural Design Research in Asia (CAADRIA). 2020: Bangkok, Thailand. p. 709-718.

3.            Zheng, H., A Hybrid Machine Learning Method for Generating Multi-Solution Architectural Plan Drawings. Under Review, 2021.

4.            Zheng, H. and Y. Ren, Architectural Layout Design through Simulated Annealing Algorithm, in Proceedings of the 25th International Conference on Computer-Aided Architectural Design Research in Asia (CAADRIA). 2020: Bangkok, Thailand. p. 275-284.

5.            Liu, X. and H. Zheng, Applications of Machine Learning Methods for Predicting the Risk of COVID-19 in Beijing Based on Multifaceted Data, in Proceedings of the 12th annual Symposium on Simulation for Architecture and Urban Design (SimAUD). 2021: Los Angeles, USA.

6.            Sun, Y., L. Jiang, and H. Zheng, A Machine Learning Method of Predicting Behavior Vitality via Urban Forms, in Proceedings of the 40th Annual Conference of the Association for Computer Aided Design in Architecture (ACADIA). 2021: Philadelphia, USA.

7.            He, J. and H. Zheng, Crime Map Prediction through Generative Adversarial Networks. Under Review, 2021.

8.            Shou, X., P. Chen, and H. Zheng, Predicting the Heat Map of Street Vendors from Pedestrian Flow through Machine Learning, in Proceedings of the 26th International Conference on Computer-Aided Architectural Design Research in Asia (CAADRIA). 2021: Hong Kong, China.

9.            Zheng, H., A Generative Architectural Design Method through Artificial Neural Networks. Under Review, 2021.

10.          Zheng, H., DigitalFUTURES World 2020 Workshop 2020.

11.          Zheng, H., V. Moosavi, and M. Akbarzadeh, Machine learning assisted evaluations in structural design and construction. Automation in Construction, 2020. 119: p. 103346.

12.          Zheng, H., Form Finding and Evaluating through Machine Learning, in Proceedings of the 1st International Conference on Computational Design and Robotic Fabrication (CDRF) 2019. 2019: Shanghai, China.

END

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s