AI-Augmented Designing Process of a Pedestrian Bridge in Switzerland
Leveraging AI to enhance designing processes in the AEC Industry: A Case Study of a Pedestrian Bridge in Switzerland
The world is facing an unprecedented urbanisation and with it, a growing concern for the environment. The building sector plays a crucial role in this issue, as buildings consume a significant amount of resources. Yet, the current practice of architectural, engineering and construction (AEC) often leads to a disconnect between the conceptual design phase and later phases of the building process.
The conceptual design phase is a critical stage in the building process where important decisions are made that can impact the final structure. During this phase, the problem is defined and distilled into a set of design parameters, such as the type of bridge, material, and number of supports, that control the shape, building materials, and elements of the final structure.
Once the design parameters are selected, engineers calculate the performance attributes, such as costs, material consumption, and structural utilization, that determine the “goodness” of the design. However, the calculation of these attributes can be time-consuming, especially for complex structures. Moreover, the conventional software used in the design process often lacks feedback on how to modify the design parameters to optimise performance, leaving it up to the experience and intuition of the design team. This leads to a number of problems:
- The design process can be lengthy and iterative, requiring multiple rounds of tweaking and refinement before arriving at a final design with satisfactory performance.
- It can only be sped up by skilled engineers with prior knowledge, which can be hard to find, especially for complex structures and cutting-edge technologies.
- Relying on the experience and intuition of the design team can lead to biased designs and a limited exploration of the design space.
To address these problems, researchers from ETH Zürich developed a method that uses Artificial Intelligence (AI) to invert the traditional design process. Instead of starting with design parameters and then assessing the performance attributes, this tool maps from performance attributes to design parameters and allows engineers to select from a range of designs that already meet predefined performance attributes. It saves a considerable amount of time that would otherwise be spent on iterations, time which can now be invested in exploring more designs in a more holistic manner that also takes environmental aspects into account.
The AI-augmented approach requires the computation of considerably more examples and the additional training of a ML-model. However, these computations can be done during down-times, such as overnight or over the weekend. Once the generative ML-model has been set up, it can generate designs in real-time. This streamlines the designing process and allows for the exploration of a much larger variety of designs.
It is important to note that the predictions made by the AI tool are only approximations. Before making a final decision, engineers still need to run the final design through established CAD software to verify the results.
Case study: Pedestrian Bridge in Switzerland
In order to assess the effectiveness of this tool, the developers tested it on a real-life case study, a pedestrian bridge built by Basler & Hofmann AG in St. Gallen, Switzerland.
The pedestrian bridge will provide a pedestrian crossing from the exit of a parking garage, over a busy street, then through a park, to terminate next to a church.
Parametrisation and Dataset generation
The chosen parametrisation of the bridge consists of 6 design parameters (denoted features in the following figure) and 37 performance attributes (denoted performances), which include boundary conditions such as whether a bridge clips one of the trees in the park or not.
With the parametrisation in place, a large set of designs were sampled (18'000 examples), meaning that the design parameters were slightly varied for each bridge, and the corresponding performance attributes computed. The resulting data of pairs of design parameters and performance attributes constituted the dataset for training the ML-model.
Machine Learning Model
A Conditional Variational AutoEncoder (CVAE) was selected as the machine learning model. It has two main components: an Encoder and a Decoder. The Encoder takes in the design parameters and generates predictions for the performance attributes, as well as a latent vector z, which encodes additional information about the design that is not captured by the performance attributes. The Decoder takes the performance attributes and latent vector as inputs and reconstructs the original design. Note that the information contained in the latent vector z is not designed to be directly understandable by humans and typically lacks a clear interpretation or relationship to the original input data. Its purpose is to capture relevant patterns and relationships in the input data in a way that can be easily manipulated by the model for specific tasks, not to provide a human-readable representation of the data.
One advantage of this CVAE model is its ability to solve both forward and inverse problems simultaneously. The Encoder serves as a surrogate model for the forward problem by predicting performance attributes, while the Decoder solves the inverse problem by mapping from performance attributes to design parameters. After training the model, new designs can be generated by inputting desired performance attributes into the Decoder along with random values for the latent vector, resulting in a variety of different designs that respect, with a certain error-margin, the requested performance attributes.
Interpretability of the ML-model
Neural networks and CVAEs are amongst the most powerful generative models available to this day. However, the are often referred to as “black box” models.
A black box model is a type of model where the internal workings are not easily understandable or interpretable. In other words, it is difficult or impossible to understand how the model arrived at its predictions or decisions.
This lack of interpretability is a disadvantage in certain applications where transparency and interpretability are crucial for safety reasons, such as healthcare and civil engineering. This is why methods for increasing the interpretability of neural networks are a hot topic in the field of AI and several techniques have been proposed, such as the sensitivity analysis.
In the context of CVAEs, sensitivity analysis is performed by taking derivatives of the output with respect to the input. The derivatives provide information about the rate of change of the output with respect to changes in the input, or, in other words, the derivatives tell us which direction the input should be changed in order to increase or decrease the output. This provides insights into the relationship between features and can help to identify the most important factors that influence the model’s predictions.
This figure shows that the model found a positive correlation between the height and thickness of a girder and the total cost of the bridge. Using the sensitivity plots, engineers can estimate the robustness of the model by determining if the found relations make sense or not. In this particular case, the relation makes sense, as the volume of the girder (determined by its height and thickness) has a large influence on the amount of concrete necessary for building a bridge, which in turn influences the costs.
User Interface
Finally, the entire toolbox was equipped with an intuitive user interface that allows engineers to quickly generate and evaluate new designs.
Conclusion
In conclusion, this toolbox has the potential to revolutionise the design process in architecture and civil engineering. Its ability to efficiently estimate design performance and providing instant feedback and proposals to designers make the design process faster and more efficient. Furthermore, the efforts at improving the interpretability through methods like the sensitivity analysis enhance the overall design optimisation process and increase trust in the tool. The user feedback supports the conclusion that this tool has a wide range of potential applications in the AEC industry and research, serving as a valuable aid in conceptual design studies.
Thank you a lot for reading until the end of this article! You can find more information about the toolbox on the project page of ETH Zürich and about the lead researcher Dr. Michael Kraus on mkrausai.com, or about myself on rabischof.ch.
[1] Balmer, Vera M., Kuhn, Sophia V., et al. “Design Space Exploration and Explanation via Conditional Variational Autoencoders in Meta-model-based Conceptual Design of Pedestrian Bridges.” arXiv preprint arXiv:2211.16406 (2022).
[2] H. Woodtli and C. Svec, “Projektwettbewerb Brücke über den Graben
(Passerelle St. Gallen, Unterer Graben 25), Projektwettbewerb für Senn
Resources AG,” Mar 2021.
