MACHINE LEARNING REPORT B OLIVER BALDOCK. MARCH 2020. ARCH7043.
Challiou suggests that the ‘transition’ through modularity, computational design, and parametricism has led to where Architecture exists today with regards to Artificial Intelligence (Challiou, 2019). In “Architecture + AI: Towards a New Approach” the theoretical and practical innovations of these four periods are explored with regards to each other’s progress. This approach has been (re)interpreted to produce Figure 1 which looks at the specific influences that have placed us in the position to be writing this report today. In addition to this Figure 2 describes the fundamental differences between the processes mentioned above. According to the American Institute of Architects, ‘Artificial Intelligence is the application of data - data is what machines learn from’ (AIA, 2020). According to Brok Howard, an architect for dRofus, a data management and BIM collaboration software, architects are in the best position to use data to advance their profession. He states “Any data that you can collect today that can help you tomorrow is the data you should be collecting”(AIA, 2020). With this in mind, this report will examine a number of case studies which utilise data and advances in technology to improve both the design of proposals but also the workflows of the team. Case Study 01 The first is from chapmanbdsp, a MEP & Environment Consultant that Feilden Clegg Bradley Studios have worked with on a number of projects. The case study in question examines the facade of the recently completed Centre Building at the London School of Economics (Fig.3). Chapman describes the evolutionary design process of the facade which defines the aesthetic of the proposal. The design for the vertical fins, and the entire facade system has been driven by data modelled through a variety of plugins for Grasshopper, and additional external programmes. Vast amounts of data regarding daylighting, thermal properties and heat flows within the proposal, and how they differed according to the depth of the fins, were then produced. By applying weighting to each of these measured variables, a singular value, or output, is then produced for each depth of fin,or input(s). This can then be plugged into Galapagos, an evolutionary solver which attempts to maximise, or minimise, that output by altering the input(s). This process is demonstrated in Figure 4. This design process generates a larger amount of design options, in a shorter period of time, than a designer would be able to manage. It allows the designer to quickly understand how the input affects the output and how a design can be optimised within the given parameters. However, there is some debate as to whether the method described above can be considered under the umbrella of Artificial Intelligence. Galapagos implements a genetic multi-objective algorithm solver. This solver uses a pool of solutions, provided by a number of input parameters which have been set by
the designer, in Chapman’s example it is the depth of the vertical fins at each level of the building. This pool of solutions then undergoes a number of recombinations and mutations which produces new ‘children’. Each ‘child’ is assigned a fitness value, in our example, how close the iteration comes to the optimum comfort value for the occupants. The process is then repeated with the fitter values as the solver tends towards an optimum solution. As such, we observe that the solver ‘learns’ from its previous experience, replicating a form of intelligence and as such can be considered a form of supervised learning where the constraints of the machine’s learning have been defined by the data provided to it. Genetic solvers also sit very cleanly within Parametricism, providing the next logical step in pushing an algorithm to produce its optimum output. However, the optimum output is only as good as the inputs parameters allow it to be. And these are always defined by the designer. Case Study 02 In contrast to this, Chaillou suggests that with the introduction of Artificial Intelligence into architecture, architects have an opportunity for “reflexive empowerment”. He goes on to say that with the correct education, i.e. data curated by the design team, and a properly explained job, AI can become a “trustworthy assistant”. Chaillou’s research uses Generative Adversarial Neural Networks, or GANs, to develop a machine which designs under the guidance of an architect. (Challiou, 2019) GANs consist of two models, the first, the Generator, creates images which resemble images from a training dataset provided by the architect. The second, a Discriminator, is trained to recognise images from that training dataset. The Discriminator provides the Generator with feedback as to the quality of the images being produced, and in response, the Generator adapts. This feedback loop allows the machine to improve the “fitness” value of what it is producing. Chaillou combines three models to replicate the design process that architects go through. These are shown in Figure 5 as developing from initial massing on the site to apartment layouts and finally furnishing. Furthermore, by changing the training data provided to the models, in this case from Boston building footprints and their room layouts to Baroque housing, he was able to apply fundamental parameters of that style to new footprints, demonstrating the importance of data curation when training artificial intelligence. My Project Fig 6 demonstrates, at a high level, the type of data that I intend to extract from the surrounding context and the (hypothetical) occupants. The importance, and thus weighting, of this environmental data, will be decided by the priorities of those the dwellings are being designed for. The balance in the resultant feedback will then affect the design proposal. These changes are then weighted against an environmental and financial cost to produce a “score” that the evolutionary solver will attempt to optimise by adjusting the design parameters. This score is further balanced against the needs and desires of the other 49 proposed units within the scheme.
Subsequently, data from this testing can be extracted to inform a machine learning model that interrogates the correlation between user priorities and subsequent design changes, improving how both user input is interpreted but also how the scheme develops as a whole. Furthermore, whilst the area of the units I am proposing is fixed by the historic footprint of Pancras Arches, existing small housing schemes can be used to inform efficient space layouts. By feeding a machine learning model existing floor plans that balance efficient uses of space with desirable living environments it may be possible for the model to (re)produce options for these floor plans. However, since this design is site specific, understanding the effect of environmental factors is key to producing an effective result. There is also a case to made for the wider data applications. A lot of time was spent in the middle of first year manually analysing the rooftops of Camden to calculate the feasible area on which rooftop homes could be built. The analysis was simple enough, and could, with a detailed lidar model and data on site ownership to be implemented with a machine learning model. This could be employed similarly to WeWork’s site analysis, quickly identifying potential locations based on the size, age and orientation of existing buildings.
BIBLIOGRAPHY AIA (American Institute of Architects). (2020). Embracing artificial intelligence in architecture. [online] Available at: https://www.aia.org/articles/178511embracing-artificial-intelligence-in-archit:46 [Accessed 9 Feb. 2020]. Chaillou, S. (2019). AI + Architecture | Towards a New Approach. Harvard University, 188. Singhal, D., Calleja, H. and Lloyd, S. (2020). Road to Zero Carbon.
APPENDIX INVENTION & THEORY INNOVATION Walter Gropius: Baukastan Concept Computers for Architectural Design?
Walter Gropius Packaged House System Robert W McLaughlin Winslow Ames House
MODULARITY Luigi Moretti Parametric Architecture Invention
Le Corbusier Le Modulor Buckminster Fuller Dymaxion House
COMPUTATIONAL DESIGN
Dartmouth College Artificial Intelligence Invention
Patrick Hanratty PRONTO
Ivan Sutherland Sketchpad
Luigi Moretti Twelth Milan Triennial Architectura Parametrica
Christopher Alexander Notes on the Synthesis of Form Joseph Weizenbaum AI Lab, MIT ELIZA Christopher Alexander A Pattern Language Nicholas Negroponte The Architecture Machine
Moshe Safdie Habitat Yona Friedman Flatwriter
George Stiny Two Exercises in Formal Composition Autodesk AutoCAD
Cedric Price Generator
John Holland Genetic Algorithms and Adaptation Douglas Lenat Cyc Project Samuel Geisberg Pro/ENGINEER
PARAMETRICISM GENERATIVE DESIGN
Frank Gehry & Jim Glymph CATIA Zaha Hadid Vitra Fire Station
with Evolutionary Design Systems
Bruce Mau Incomplete Manifesto for Growth
Phil Bernstein REVIT Yoav Parish & Pascal Muller SIGGRAPH 2001 CityEngine & L-systems
Patrick Schumacher Parametricism: A New Global Style for Architecture & Urban Design
David Rutten Grasshopper
Paola Fusero et al. Parametric Urbanism: A New Frontier for Smart Cities Ian Goodfellow et al. Generative Adversarial Neural Networks Nathan Peters Enabling Alternative Architectures Stanislas Chaillou AI & Architecture Towards a New Approach
ARTIFICIAL INTELLIGENCE Autodesk MaRS Innovation District of Toronto SUPERVISED LEARNING CLASSIFICATION | REGRESSION UNSUPERVISED LEARNING CLUSTERING | ASSOCIATION | DIMENSION REDUCTION REINFORCEMENT LEARNING
ARCHITECTURE adapted from Towards a New Approach (Challiou, 2019)
FIGURE 01 - Architecture & AI Progression
COMPUTATIONAL
Designer
Design Proposal
PARAMETICISM
Designer
Parameters + Generative System
Design Proposal
Data
ARTIFICIAL INTELLIGENCE
Designer
Machine Learning
Parameters + Generative System
Design Proposal
FIGURE 02 - Computation, Parametricism, Artificial Intelligence
FIGURE 03 - Chapmanbdsp, Central Building LSE
REVIT BIM Model
XFlow CFD Analysis DIVA Daylight Analysis
Input (Geometry)
Sketchup 3D Model
Rhinoceros 3D 3D Model
Radiance Daylight & Solar Analysis
Ladybug Daylight & Solar Analysis
Grasshopper Algorithm
Honeybee Thermal Analysis
Galapagos Evolutionary Solver
ENERGYPLUS Thermal Analysis
EDSL TAS Advanced Thermal Analysis
AutoCAD 2D / 3D Drawings
FIGURE 04 - Chapmanbdsp, Central Building facade design workflow
FIGURE 05 - Challiou’s model progression
INPUTS Other Units
Acous�cs
OUTPUTS LIFECYCLE ANALYSIS
Light
Materiality
EMBODIED
COST
OPERATING
CARBON FOOTPRINT
OPTIMISE
ITERATION SCORE
Structure
Access
CONTEXTUAL ANALYSIS
USER PRIORITIES
DESIGN DEVELOPMENT
Grasshopper
User Consulta�on
Designer
DESIGN PROPOSAL
Privacy / Views
Thermal Spacial Requirements
FIGURE 06 - Proposed data source workflow
Galapagos
Output (Resolved Geometry)