Machine Learning- Building Envelopes
UCLA IDEAS Technology Studio Fall 2018
Research Lead: Guvenc Ozel
Instructors: Guvenc Ozel, Benjamin Ennemoser, Gabby Shawcross
Technologies: machine learning, gaming engines
Students: Christina Charalampaki, Pratina Chadha, Bohan Cheng, Ruiyang Cheng, Kevin Clark, Haocheng Dai, Yutong Dai, Yue Di, Rui Ding, Lu Geng, Jorge Gutierrez, Vigil Escalera, Hoodeen Hakimian, Ying Huang, So Hyun Kim, Zhou Le, Luo Lei, Yunqi Lei, Zhengtao Liu, Qingyun Lu, Sanjeet Sanjay Mukadam, Narisu Narisu, Nidhi Parsana, Kshama Swamy, Hanshuang Tong, Yiliang Wang, Kan Yen Wu, Xuanyi Xu, Ce Yan, Yixuan Ye, Yicheng Zhao, Yiran Zhou
Data Architectures: Collaborating with AI for Data Based Design Methodologies
So far we have made a differentiation in the studio between the performative aspects of machine intelligence and its ability to automate form-making processes. We primarily (and intentionally) limited their application on the ability of design- form to respond to outside contexts such as the environment, human occupation and human psychology, and programmed behaviors for spaces to interactively respond to conditions as such through sensor interfaces working in collaboration with formal systems that are conducive to physical and ambient transformation through mechanics and media. We treated the accumulation of contemporary formal systems and styles as pedagogical and methodological repositories, which would act as inventories for further formal iteration. This approach yielded to productive results in understanding the human presence as it relates to form in motion, but has not resulted with alternative spatial languages that can be deployed and explored further in order to pose alternatives to familiar models of computational form making, which are heavily constricted by the limitations of parametric design tools. In many cases, the questions of enclosure as it relates to motion, novel material science experiments, and variable modulations to strike a balance between static and dynamic qualities of space were explored through investigating the historic evolution of such forms in architectural and industrial design. In order to overcome this problem, the studio will focus on the relationship between human agency and computational iteration by using machine learning as a design tool. We will be using basic machine learning tools in order to classify and iterate stylistic approaches that exist outside the discipline of architecture to allow for ML to design, coordinate, randomize and iterate qualities as they relate to pattern, color, proportion, hierarchy and language. The human engagement in this design process will be limited to the initial curation of input data that the ML system can learn from, and also in regulating and choosing the iterations as the final outputs of two-dimensional images such systems are capable of producing. An additional computational mediation process in the form of agent-based systems will be deployed in order convert two dimensional information into 3d geometry.
Process:
Documentation:
Through documenting selected civic buildings and their urban presence in Downtown Los Angeles, we plan to bypass discussions regarding urban mass, adjacency and other architectural concerns that fall beyond the scope of the studio. These selected buildings will serve as vessels for our machine learning operations. We will be using drone footage and photogrammetry techniques to interpret and translate 2D visual data into 3D.
Precedent Study:
We will be looking outside the discipline of architecture to mine for inspiration for new aesthetic paradigms. Looking into the world of film, photography, painting, fashion and design, the students will create a library of images from a particular body of work by an artist or designer.
Training Data:
The selected body of work will be used as “training data” for a machine learning algorithm. By feeding the image database created in the precedent study to the system, the algorithm
is “trained” to recognize and develop biases based on the formal qualities of the work regarding color, depth, pattern and other visual traits.
Style Transfer:
We will be using a machine learning technique called “style transfer” in order to apply the Training Data to the massing of the building as seen on the drone footage. This two dimensional yet dynamic footage will create a sketch and a concept for the new building mass.
Reconstruction:
Through procedural modeling techniques, photogrammetry, agent based systems and other contemporary 3D tools,
we will be turning the two dimensional information in the revised drone footage into a 3D mesh. This new 3D mesh will constitute the design sketch for the new building.
Composition:
After the mesh is refined, we will be creating hyperrealistic cinematic representations of the new designs by inserting them back into the drone footage through advanced video editing, compositing and rendering techniques.