Predictive Structural Assessment of Concrete Frame Bridges using Machine Learning – Concrete E-Learning Platform

Link zur deutschen Version: Wie maschinelles Lernen die statische Überprüfung bestehender Bauwerke unterstützen kann

Many bridges across Europe and beyond are reaching the end of their original design lifespan. This presents a major challenge, as the safety of thousands of ageing structures has to be assessed in a short amount of time. Asset owners have to strategically allocate resources for maintenance while balancing budget constraints, environmental considerations, and limited personnel capacity. However, traditional structural assessment methods are time-consuming, expensive, and therefore do not scale efficiently to large portfolios.

Recent advances in machine learning (ML) offer promising solutions. By creating data-based predictive models, we can accelerate initial assessments, prioritize structural interventions, and provide decision support for the structural assessment process.

In this blog post, we introduce a ML-based pre-assessment tool developed for reinforced concrete frame bridges – one of the most common bridge types in Switzerland. This prototype was created as part of an ongoing research project at our chair in collaboration with the Swiss Federal Railways (SBB). In the following sections, we demonstrate the tool’s application on a real-world example and discuss its practical implications and development potential. For those interested in the model’s creation, insights into the underlying methodology are provided at the end.

Application Example from the SBB Portfolio

To demonstrate how this tool can be applied in practice, we tested our ML-based pre-assessment tool on a real Swiss railway underpass from the SBB bridge portfolio. To estimate the structural compliance of an existing frame bridge, users can input or import the relevant bridge parameters into the web tool and generate predictions (see Figure 1) . To provide an initial estimate, the model only requires a few high-level geometric dimensions as input: span, plate thickness, and the bridge’s height and width. To achieve higher prediction certainty, users can also include additional structural details such as reinforcement layout, material properties, and loading conditions.

Figure 1: Application example of the ML-based web tool for the pre-assessment of a concrete frame structure situated in Switzerland.

Figure 2 shows the results predicted by the ML model for the application example. The output consists of three predictive distributions representing the estimated compliance factors η for the inputted bridge structure, which indicate whether a bridge meets safety requirements according to the Swiss structural codes for ultimate limit state. The underlying data-driven model was trained and validated to generate predictions that closely align with detailed mechanical structural analyses.

Figure 2: Output of the ML model for the example bridge [1]: Predicted minimal compliance factors η for the bridge structure. These factors correspond specifically to the bending verification of the reinforcement (η*_M,s,min*), the bending verification of the concrete cross-section (η*_M,c,min*), and the shear verification (η*_V,min*). The predicted distribution is depicted in grey with its mean prediction μ (= best estimate) and its standard derivation σ multiplied with the calibration factor κ (estimating the prediction uncertainty).

For this specific bridge, the ML model predicts that both bending verifications meet required compliance factors with a high level of confidence (η >> 1). However, for the shear verification, while the best estimate (i.e., the mean of the predictive distribution μ) remains above the threshold for structural sufficiency (η = 1), the model also indicates a degree of uncertainty. This is evident from the spread of the predictive distribution with a small part of the predictive distribution also falling below the threshold of 1. For this application example the tool, therefore, identifies the shear verification as critical. Given this uncertainty and the predicted value being close to the threshold of 1, the tool recommends the utilisation of detailed analysis methods to check structural safety, rather than applying simplified analysis methods, which may be too conservative to verify this structure.

The tool’s recommendation for a refined structural assessment was confirmed by a subsequent non-linear finite element analysis (FEA) of the structure, which validated the accuracy of the ML model’s compliance factor predictions for this example application. As a reference, the bridge had been previously assessed by an engineering firm using conventional linear FEA. This assessment resulted in insufficient compliance factors, leading to a recommendation for strengthening measures at an estimated cost of CHF 300,000. This shows the advantage gained with the prediction tool in this example. By using the ML-based pre-assessment tool, the appropriate refined analysis method was selected from the start. The non-linear FEA results ultimately confirmed structural sufficiency, preventing unnecessary, costly, and resource-intensive strengthening measures.

Practical Implications and Development Potential

Such a ML-based predictive assessment tool is not meant to replace detailed structural analysis but to serve as an efficient and economical first screening method. The application example showed that the model’s predictions provide decision support, helping asset owners determine the next steps for the structures within their bridge portfolio. It also helps structural engineers determine the appropriate level of refinement for the structural assessment of a specific structure. It is important to note that the tool’s predictions must always be interpreted with engineering judgment.

This prototype currently is applicable to simple concrete frame bridges. However, the method can be extended to also cover other bridge types, other countries’ structural codes and even additional factors such as damage assessments. Further development of the tool aims at efficient prioritisation of bridge assessment across large portfolios, efficiently allocate resources where they are needed.

How the ML model was developed

Based on the large frame bridge portfolio of the Swiss Federal Railways (SBB), we parameterised common concrete frame bridge structures, developed a parametric structural assessment pipeline and ran detailed non-linear FEA simulations, to create a dataset of structural compliance factors. Based on this dataset, we then trained a ML model (specifically, a Bayesian Neural Network [1]) to learn the mapping between bridge parameters and compliance factors (see Figure 3). This allows the model to interpolate the data in the high-dimensional parameter space and to make predictions for bridges that have not yet been analysed. Unlike standard neural networks, the Bayesian approach also estimates how uncertain a prediction is. Both the parametric pipeline used for data generation and the trained ML model, were validated on parameter studies as well as on existing structures (as seen in the application example above). The model’s predictive accuracy was tested on unseen data, and its uncertainty estimates were calibrated to ensure accurate confidence intervals.

Figure 3: *Left:* Simplified ML framework. The ML-model takes bridge parameters x as input and predicts compliance factors η, while also providing an estimate of the prediction uncertainty. Right: Conceptual representation of the surrogate model (grey surface), which approximates the response surface of a mechanical model. The approximation is based on a set of point evaluations (black points) from the parameter space (blue space), defined by structural characteristics such as span and plate thickness. (Figure was adapted from [1].)

Want to Learn More?

Are you interested to know more? Do you have questions or comments to this work? Feel free to reach out to Sophia Kuhn.

Sophia Kuhn

Literatur

[1] S. V. Kuhn, M. Weber, A. Binggeli, M. A. Kraus, F. Perez-Cruz, W. Kaufmann, Predictive structural assessment of concrete frame bridges with bayesian deep learning, manuscript to be submitted for publication (2025).