Artificial intelligence helps with the design and maintenance of bridges

Images of a collapsed tram bridge over the River Elbe in Dresden were seen around the world in September 2024. It’s a miracle no one lost their life – unlike in the collapse of the motorway bridge in Genoa in 2018, which led to 43 deaths. Both disasters were caused not by external influences, but rather by damage processes associated with the age of the structures. These processes were not detected and rectified in time. “Switzerland is also facing a situation in which a considerable proportion of its infrastructure is nearing the end of its planned service life and must be inspected and strengthened if necessary,” says Sophia Kuhn. “We’re developing a tool that helps to keep bridges in operation for as long as possible and therefore to conserve resources without running a disproportionate risk of accident.”

Sophia Kuhn is a doctoral researcher in the group led by Walter Kaufmann, ETH Professor of Structural Engineering (Concrete Structures and Bridge Design). Her doctorate is co-supervised by Fernando Pérez-Cruz, ETH Professor of Computer Science, and Professor Michael Kraus from TU Darmstadt. Kuhn’s research focuses on the use of artificial intelligence in construction, particularly machine learning algorithms. In collaboration with colleague Marius Weber and the Swiss Federal Railways (SBB), she has developed an AI model for “rigid frame bridges” – simple railway bridges made of reinforced concrete, which are particularly common in Switzerland and allow railways to pass above or below roads or footpaths, for example. Practically at the touch of a button, the AI model provides an initial assessment of structural safety thereby predicting whether a bridge is potentially statically critical or not. “It’s therefore possible to prioritise which bridges should undergo structural assessment without delay and may require structural interventions,” says Kuhn. 

AI can assess whether the analyses will be effective

The model not only delivers a predicted value for structural safety but also indicates whether this value is reliable; in other words, it quantifies the uncertainty of the model. In particular, it also helps with the decision regarding how to proceed when conducting a structural assessment of a bridge. Engineers always carry out more or less complex calculations on a computer, but this can be done either using conventional methods, which deliver results with relatively little effort, or using refined analyses, which are much more intensive in terms of time and processing power and therefore more expensive, although they deliver more accurate and less conservative results. “Often, you don’t know whether it makes sense to perform these refined analyses or whether they’re just an unnecessary expense,” Kuhn explains. “Our AI tool can assess whether the analyses are likely to be effective and whether the cost involved is worthwhile.”

Simulation pipeline delivers additional data

As a basis for the model, the researchers used the portfolio of SBB rigid frame bridges. “We looked at lots of examples – how they’re built, how variable they are – and developed a parametric simulation pipeline based on them,” says the researcher. This generated virtual structures from various bridge parameters, calculated the extent of the structural capacity utilisation and thereby produced additional data.

The researchers built an artificial neural network, an algorithm that learns from the data in a similar way to our brain. This gave rise to a machine learning-based model that delivers the desired predictions for many existing rigid frame bridges, even if these have not been calculated by experts or by the simulation pipeline. “We validated our model on a test dataset and evaluated it with real bridge examples,” Kuhn explains. “The model exhibits good alignment and the level of precision needed for SBB. We have therefore developed an initial prototype.” The next step involves working together with SBB to ensure that bridge engineers can apply the model in practice – and then facilitate broader applicability of the model.

AI assistant inverts design process

In a second project from the Kaufmann chair, Sophia Kuhn worked with Professor Michael Kraus and the Swiss Data Science Center on the design of new bridges. “Our aim was to develop an AI assistant that actively helps the team of engineers design the bridge and leads to cost-efficient structures that are as sustainable as possible without impairing safety,” Kuhn explains. Traditionally, engineers draft a bridge design and then use conventional calculation software to determine the structural safety, serviceability, costs and other characteristics. If these values do not meet the specifications, the team changes the design until the project objectives are met – a lengthy process in which often a great deal of potential goes unharnessed.

“Actually, what is preferred is to invert this process, but that isn’t possible with conventional calculation software,” says the researcher. “What one wants is to input the project objectives and boundary conditions and then receive proposed designs that meet these specifications without the need for laborious iterations.” The AI assistant developed by the researchers, which uses “generative” AI algorithms, allows precisely that. It not only speeds up the forward approach by assessing various designs almost in real time, but also proactively generates designs that meet the defined constraints and objectives.

As a case study for developing their AI assistant, the researchers, in collaboration with colleague Vera Balmer, used the project of a pedestrian bridge in St. Gallen designed by the engineering company Basler & Hofmann together with Nau2 and dgj Landscapes. This bridge, known as the Wiborada pedestrian bridge, runs through a park in the old town and should avoid touching any of the protected trees if possible. During their work on this project, the ETH researchers were in contact with the engineering company, which was impressed with the presentation of the results. The AI assistant delivered various possible bridge examples and also performed a “sensitivity analysis” that indicated which parameters have the greatest influence on structural safety in accordance with standards, or on the estimated costs or sustainability.

“The AI assistant therefore supports engineers but does not replace them”, Kuhn emphasizes. For example, if the AI assistant proposes a design that, although it is unexpected, meets the specifications in terms of structural safety and environmental compatibility, the engineers must still assess whether it’s possible to build such a bridge and whether it will be durable. “We’re not providing a one-click solution. It always involves an interaction between the engineer and the AI,” says the researcher.

Toolkit for adapted AI models

Bridge construction isn’t the only potential application of these advanced machine learning techniques. Together with other ETH researchers from the Swiss Data Science Center and the architecture chair Gramazio Kohler Research, the research group from the Kaufmann chair developed a toolkit that also made AI algorithms accessible to other engineers and architects without the need for extensive programming skills.

“With just a few lines of code, our open-source toolkit allows users to build both forward models and generative models that can be used to solve complex, high-dimensional problems in architecture, the construction industry and beyond,” explains Kuhn. This is intended to provide broad-based support for economic and sustainable planning in construction. “In the construction sector, these approaches are less widespread than in other industries such as mechanical engineering,” says the researcher. “There’s still considerable potential for greater efficiency and sustainability using data-driven methods – and that’s our objective.”