Constrained Bayesian optimization for engineering bridge design

Heine Røstum, Sébastien Gros, Ketil Aas-JakobsenStructural 2025 · DOI

TL;DR

Bridge design typically relies on "black-box" simulations that are too slow for traditional optimization. This paper demonstrates that constrained Bayesian Optimization (cBO) can find optimal, low-carbon bridge designs in just six iterations by using probabilistic surrogate models to intelligently navigate the design space.

What Problem Does This Paper Address?

Structural engineers face a "black-box" problem: designing a bridge requires balancing dozens of variables (geometry, material strength, reinforcement) against strict safety codes. To check if a design works, they use Finite Element Method (FEM) software. However, FEM simulations are computationally expensive, and commercial software rarely provides the "gradients" (slopes) needed for standard numerical optimization.

Consequently, engineers often rely on intuition or limited "surrogate models" that approximate the FEM results. The gap the authors identify is that even when surrogate models are used, engineers rarely exploit the uncertainty of those models. By ignoring what the model doesn't know, traditional methods either sample too many points or get stuck in local sub-optima.

Methods

The authors implement a constrained Bayesian Optimization framework. The core components include:

  • Gaussian Process Regression (GPR): Instead of a simple curve-fit, GPR acts as a probabilistic surrogate. For any given design, it provides both a predicted value (e.g., "this bridge will emit 165 tonnes of CO2") and a measure of uncertainty (e.g., "plus or minus 5 tonnes").
  • Acquisition Functions: These are the "decision-makers" that choose the next design to test. The paper compares two state-of-the-art types:
  • Constrained Expected Improvement (cEI): Looks for designs that offer the biggest potential improvement over the current best, weighted by the probability that the design actually meets all safety constraints.
  • Constrained Max-value Entropy Search (cMES): An information-theoretic approach that chooses points to maximize the "information gain" about where the optimal design is located.
  • Baseline Comparison: They compare these against a standard GPR Mean (GPRM) approach, which only looks at the predicted average and ignores uncertainty (pure exploitation).

Datasets

The study doesn't use a static machine learning dataset; instead, it generates data through a structural case study:

  • The Subject: A three-span (20m–30m–20m) post-tensioned concrete road bridge girder.
  • Environmental Data: CO2-equivalent factors for concrete, reinforcement steel, and tendons were sourced from the vegLCA tool provided by the Norwegian Public Road Administration.
  • Simulation Tool: Structural constraints (strain limits) were evaluated using NovaFrame, a commercial FEM software suite.

Key Insights and Ideas

  • Uncertainty is a Resource: The "aha moment" here is that knowing where your model is "dumb" (high variance) is just as important as knowing where it thinks the optimum is. Bayesian Optimization uses this uncertainty to prevent the search from getting stuck.
  • Information vs. Value: The paper highlights a subtle difference between acquisition functions. While cEI is "aggressive" and hunts for the best value, cMES is more "inquisitive," seeking to reduce the overall entropy of the search space.
  • The Feasibility Trade-off: The authors found that cEI often pushes designs right to the edge of failure (the constraint boundary) to get better performance. While this results in slightly better CO2 numbers, it results in a lower percentage of "feasible" designs during the search process compared to the "safer" cMES.

Key Results

  • Rapid Convergence: The Bayesian schemes (cEI and cMES) converged in roughly 6 iterations after the initial sampling. This is significantly faster than traditional surrogate-based methods.
  • Consistency: In 40 repeated trials, the Bayesian methods showed very tight confidence intervals. This means an engineer can trust the tool to find a near-optimal result in a single run, rather than needing to restart the process multiple times.
  • Superiority over Baseline: The standard GPRM approach (ignoring uncertainty) was slower to converge and much more likely to recommend unfeasible designs in the early stages.
  • Concrete Improvement: The optimized designs achieved a relative improvement in CO2 emissions of approximately 8-9% compared to the best initial random samples.

Limitations and Future Work

  • Computational Complexity: GPR models scale cubically ($O(n^3)$) with the number of data points. While perfect for the small-to-medium datasets in this paper, they may struggle with massive, high-dimensional structural problems involving thousands of nodes.
  • Constraint Complexity: The study focused on Ultimate Limit State (ULS) capacity. Real-world bridge design requires checking Serviceability (SLS) and Accidental (ALS) states, which would introduce many more constraints for the GPR to model.
  • Future Directions: The authors suggest exploring Sparse Gaussian Processes or Deep GPs to handle higher dimensionality and larger datasets, potentially allowing for the optimization of the entire bridge structure rather than just a single girder section.

Summary updated

Abstract

Abstract Designing a bridge is a complex endeavour, involving multiple variables, limitations and requirements. The design process often includes high-fidelity analyses that are computationally expensive, and the internal working tools of the analysis software are often unknown. This limits the applicability of conventional numerical optimization, especially due to time constraints. As a mean to reduce the computational burden, surrogate modelling may be applied. Surrogate models are constructed on the basis of observed results from the computationally expensive high-fidelity analyses, and serves as a fast approximation of unobserved regions in the design space. If probabilistic surrogates are applied, the probabilistic element may be exploited in the optimization phase, resulting in a scheme known as Bayesian optimization. In this article, it is described how to derive a constrained Bayesian optimization scheme in the process of bridge design, where both the goal and constraints are approximated using probabilistic surrogates. The article also presents a case study where constrained Bayesian optimization is applied to a three-span post-tensioned concrete girder and the results are compared to conventional surrogate-based optimization. The results from the case study show that the Bayesian optimization schemes converge after about six iterations, significantly faster than the conventional surrogate optimization scheme, with a consistently higher relative improvement—providing a faster and more confident process for surrogate-based optimization.
BibTeX
@inproceedings{rstum2025constrained,
  author = {Heine Røstum and Sébastien Gros and Ketil Aas-Jakobsen},
  title = {{Constrained Bayesian optimization for engineering bridge design}},
  booktitle = {Structural},
  year = {2025},
  doi = {10.1007/s00158-024-03951-3},
}
Figure