Summary
In architecture, engineering and construction (AEC), designers are currently relying on time-consuming methods that require manual exploration of the design space. Tailormade AI-assisted tools can accelerate the design process by enabling architects and engineers to explore a broader set of solutions that satisfy technical constraints and requirements.
Introducing generative AI allows to invert the design process by leveraging machine learning (ML) models to provide a multitude of valid designs based on requested performance attributes. We introduce the AIXD (AI-eXtended Design) toolbox, which combines forward and inverse modeling to explore a diverse collection of possible designs to swiftly find suitable and more varied solutions. Developed based on real-world applications in architecture and civil engineering, the toolbox is designed for broad applicability across various design problems.
This is joint work with collaborators from Gramazio Kohler Research and kfmResearch at ETH Zurich.
Interactive design generation with inverse modeling
Before a building is constructed or a mechanical component is manufactured, designers follow standard design phases, which are monolithic steps relying on expert judgement and requiring many hours of work. Given time constraints, designers cannot delve into an exhaustive exploration of all possible solutions and therefore have to conform to solutions whose constraints or requirements can be guaranteed, even if they are suboptimal for the case.
During the different design phases, architects and engineers harness parametric models for forward design (Fig. 1), which provide a mapping between design parameters and the simulated performance attributes. By leveraging software for computer-aided design (CAD) and finite-element modeling (FEM), designers iteratively explore different parameters’ values to gain an intuition that allows them to refine the design. However, given that forward design does not allow to explore all viable parameters combinations, designers tend to rely on well-known solutions, design tables and their expertise.
Inverse design makes it possible to work in the opposite direction of forward design. It takes the target performance attributes as input and outputs the design parameters that match the requested attributes (Fig. 2). To carry out inverse design, we leverage machine learning models trained on pairs of design parameters and performance attributes generated through the parametric model. This enables a swifter and more comprehensive exploration of the solution space, as designers can now query the trained model to obtain several designs satisfying the requests.
A toolbox for AI-assisted design
In the AIXD toolbox we combine forward and inverse design by integrating ML models to facilitate the design process [1]. With AIXD, we propose a paradigm shift that allows designers to go beyond the traditional forward approach. Now, we divide the full workflow in two phases (Fig. 3). During an “offline phase”, the parametric model is used to generate design samples, characterized by design parameters and performance attributes to create a dataset, which is then used to train a neural network. The intervention of the user during this phase is minimal and mostly restricted to the configuration and assessment of the neural network model. Once a model with satisfactory performance is obtained, the designer can move to the “online phase” and interactively query the model to obtain satisfactory solutions. The implemented model is a conditional autoencoder, which enables both forward design using the encoder and inverse design using the decoder. More details on the model and its utilization can be found in https://www.datascience.ch/articles/ai-augmented-architectural-design.
The most useful ability of the autoencoder is to perform inverse design. The designer can easily interact with the inverse model, which can serve as a “co-pilot” accelerating the exploration of solutions, enhancing the understanding of the problem at hand and unveiling innovative designs. This is facilitated by the visualization tools integrated in the AIXD toolbox that make for a seamless exploration of the design space and a more thorough understanding of the problem, augmenting the designers’ intuition. We would like to highlight that inverse design does not reduce the importance of the parametric model but instead complements it by leveraging its data generation capabilities.
Real-world examples
AI-augmented vertical garden
A signature use case is presented with Semiramis, a vertical garden structure, consisting of several planted platforms, built in Zug, Switzerland. Inverse design was used to optimize performance attributes required by the landscape architect [2], who needed to control sun and rain exposition in each platform for the selection of the most adequate plants (Fig. 4). Reversing the conventional design process led to a broader set of new and surprising geometries and enabled a better understanding of the design task.
A pedestrian bridge in the trees
Another case study was conducted on the “Brücke über den Graben” ("bridge over the moat"), now called «Wiborada-Steg» in St. Gallen, Switzerland, designed by Basler & Hofmann, DGJ Landscapes and Nau2 [3,4]. The engineering design of the new pedestrian link through a park in the old town required preserving the historical ensemble while the bridge was not allowed to touch any of the protected trees in the park (Fig. 5). This application study showcased how the inverse model generates bridge designs suitable for the project constraints. Besides, it showed how efficient sensitivity analysis highlights key factors influencing structural safety, cost and sustainability [5].
Studying urban massing
Another architectural design problem, which can be addressed with the AIXD toolbox, is urban massing and how new buildings, especially in a dense urban fabric, impact the neighbouring existing buildings in terms of access to direct sunlight. Architects and urban planners use massing models, which are simplified 3D representations, to explore shape, form, and layout of a building project during early design stages.
In this use case we examined and optimized the exposure to sun caused by the construction of a new assembly of high-rise buildings. Here, the parameters to explore are the heights of the buildings in the middle while satisfying constraints related to the volume of the buildings and the amount of sunlight received by the apartment blocks behind the complex (Fig. 6).
The exploration was done fully in the CAD software Rhino/Grasshopper, using ARA, an interface built to interact with the AIXD toolbox backend [6]. This facilitates architects to seamlessly harness the AIXD features directly from their CAD environment, by providing a set of visual modules that integrate with the built-in Grasshopper tools.
Conclusion
Researchers today have access to various machine learning resources to advance design techniques in architecture, engineering and construction. To bridge the gap between AI methods and domain users we built the AIXD (AI-eXtended Design) toolbox. It combines forward design based on parametric modeling with inverse design capabilities offered by generative AI. Thanks to the connectivity to established software solutions, architects and engineers can quickly integrate it in the design process. AIXD can be broadly applied in various domains relying on parametric models, from architecture to civil and mechanical engineering. Thanks to the inverse design capabilities, AIXD enables designers to more comprehensively explore possible solutions, leading to new and innovative designs and augmenting the designer’s intuition.
Co-authors
- Dr. Aleksandra Anna Apolinarska, Gramazio Kohler Research and Swiss Data Science Center, ETH Zürich
- Sophia Kuhn, kfmResearch, Institute of Structural Engineering (IBK), ETH Zürich
- Alessandro Maissen, Swiss Data Science Center, ETH Zürich / EPFL
- Prof. Dr. Michael Kraus, former member of the Institute of Structural Engineering (IBK) at ETH Zurich, currently professor at TU Darmstadt
- Dr. Konstantinos Tatsis, Swiss Data Science Center, ETH Zürich / EPFL
- Gonzalo Casas, Gramazio Kohler Research, ETH Zürich
- Dr. Romana Rust, VYZN, Research & Innovation
- Rafael Bischof, former Junior Data scientist at the Swiss Data Science Center, currently PhD candidate at the Computational Design Laboratory, ETH Zürich
- Prof. Matthias Kohler, Gramazio Kohler Research, ETH Zürich
- Prof. Dr. Walter Kaufmann, kfmResearch, Institute of Structural Engineering (IBK), ETH Zürich
- Prof. Dr. Fernando Perez Cruz, Swiss Data Science Center, ETH Zürich / EPFL
References
- ETH Zurich & Swiss Data Science Center. (2024). AIXD: AI-eXtended Design. https://aixd.ethz.ch/docs/index.html
- Salamanca, L., Apolinarska, A. A., Pérez-Cruz, F., & Kohler, M. (2022, September). Augmented intelligence for architectural design with conditional autoencoders: Semiramis case study. In Design Modelling Symposium Berlin (pp. 108-121). Cham: Springer International Publishing https://doi.org/10.1007/978-3-031-13249-0_10
- Basler & Hofmann. (2022). Mit Künstlicher Intelligenz zur optimalen Brücke: Basler & Hofmann engagiert sich für Forschung der ETH Zürich. https://www.baslerhofmann.ch/aktuelles/details/mit-kuenstlicher-intelligenz-zur-optimalen-bruecke-basler-hofmann-engagiert-sich-fuer-forschung-der-eth-zuerich.html
- World architects. (n.d.). DGJ Paysages sàrl project page on footbrigde “Über den Graben” St. Gallen https://www.world-architects.com/en/dgj-paysages-zurich/project/footbridge-uber-den-graben-st-gallen
- Balmer, V., Kuhn, S. V., Bischof, R., Salamanca, L., Kaufmann, W., Perez-Cruz, F., & Kraus, M. A. (2024). Design space exploration and explanation via conditional variational autoencoders in meta-model-based conceptual design of pedestrian bridges. Automation in Construction, 163, 105411. https://doi.org/10.1016/j.autcon.2024.105411
- Apolinarska, A. A., Casas, G., Salamanca, L., & Kohler, M. (2024, August). ARA-Grasshopper Plugin for AI-Augmented Inverse Design. In Design Modelling Symposium Berlin (pp. 231-240). Cham: Springer Nature Switzerland. https://doi.org/10.1007/978-3-031-68275-9_19
About the author
