Efficient multi-objective constrained Bayesian optimization of bridge girder

Heine Havneraas Røstum, Joseph Morlier, Sébastien Gros, Ketil Aas-Jakobsen

TL;DR

Researchers have developed a highly efficient Bayesian optimization framework that slashes both the financial and environmental costs of bridge construction. By combining dimensionality reduction (POD) and specialized surrogate modeling (KPLS), the method identifies optimal designs 10–20% more efficient than traditional engineering "rules of thumb" in just a few dozen simulation runs.

What Problem Does This Paper Address?

The construction sector is a climate heavyweight, accounting for roughly 37% of global greenhouse gas emissions—with cement production alone responsible for 7%. While bridge engineers want to reduce this footprint, the traditional design process is often manual, experience-based, and focused primarily on minimizing financial costs.

Mathematically, bridge design is a "black-box" optimization problem. We use Finite Element Method (FEM) simulations to check if a bridge is safe, but these simulations are computationally expensive. Furthermore, we aren't just looking for the cheapest bridge; we want a "multi-objective" solution that balances cost and carbon. Standard optimization algorithms often require thousands of simulations, which is impractical when a single run takes significant time and memory.

Methods

The authors propose a framework called PTMOO (Preferred Trade-off Multi-Objective Optimization). Its architecture relies on three pillars to overcome the "curse of dimensionality":

  1. Bayesian Optimization (BO): Instead of blind searching, BO builds a probabilistic "surrogate model" (a Gaussian Process) of the bridge's performance. It uses an acquisition function (Constrained Expected Improvement) to decide exactly which design to simulate next to get the most "bang for its buck."
  2. Proper Orthogonal Decomposition (POD): Bridge simulations produce high-dimensional data (e.g., strain values at hundreds of points along the girder). POD acts like Principal Component Analysis (PCA) for these simulation fields, reducing massive datasets into a few essential "modes" that the surrogate model can actually handle.
  3. Kriging Partial Least Squares (KPLS): Standard Gaussian Processes struggle when you have many input variables (the authors used 15). KPLS uses Partial Least Squares to find the most influential directions in the input space, drastically reducing the number of hyperparameters the model needs to learn.

The "Secret Sauce": Unlike traditional multi-objective methods (like cEHVI) that try to map every possible trade-off between cost and carbon (the Pareto front), PTMOO targets a specific, user-defined trade-off. This "targeted" approach reaches an optimal solution much faster.

Datasets

The study doesn't use a static dataset; it generates data on-the-fly using NovaFrame, an in-house FEM software.

  • Initial Sampling: 30 bridge designs were generated using Latin Hypercube Sampling (LHS) to "prime" the model.
  • Case Study: A two-lane, three-span post-tensioned concrete bridge girder (70 meters total length).
  • Variables: 15 design variables, including girder height, reinforcement areas, and tendon positions.
  • Constraints: 9 high-dimensional constraints based on Eurocode standards (e.g., crack width, concrete compression, and tendon strain).

Key Insights and Ideas

  • Efficiency over Exploration: Mapping the entire Pareto front is a luxury engineers rarely need. By focusing on a "preferred" trade-off, the authors showed they could find better designs with significantly fewer iterations than standard multi-objective BO.
  • KPLS is a Game Changer: The authors found that standard Gaussian Process Regression (GPR) had high variance and was prone to getting stuck in local optima. KPLS stabilized the process, providing much more consistent results across different runs.
  • Higher Utilization = Better Design: Traditional "Rule of Thumb" (ROT) designs often have very low "utilization ratios" (meaning the material isn't being pushed to its safe limit). The optimization process naturally pushes these ratios higher, ensuring every kilogram of concrete and steel is working as hard as possible, thus reducing waste.

Key Results

  • Environmental Impact: The optimized designs achieved a ~20% reduction in CO2-equivalent emissions compared to the benchmark traditional design.
  • Financial Impact: Financial costs were reduced by 10% to 15%.
  • Convergence Speed: The KPLS-enhanced model reached a stable plateau in roughly 20 iterations, whereas standard methods were still struggling to converge at 50.
  • Safety: All optimized designs strictly adhered to the nine structural integrity constraints, proving that "greener" doesn't mean "weaker."

Limitations and Future Work

  • Regulatory Hurdles: The optimization currently doesn't "hard-code" every single Eurocode requirement (like minimum reinforcement ratios), requiring minor manual adjustments post-optimization.
  • The "Preferred" Trade-off: The method assumes the designer knows the relative importance of cost vs. carbon upfront. If that preference changes, the optimization needs to be re-run.
  • Future Directions: The authors suggest exploring Deep Gaussian Processes to see if they can achieve even higher accuracy with even fewer data points, further reducing the computational burden of bridge engineering.

Summary updated

Abstract

Abstract The buildings and construction sector is a significant source of greenhouse gas emissions, with cement production alone contributing 7 % of global emissions and the industry as a whole accounting for approximately 37 %. Reducing emissions by optimizing structural design can achieve significant global benefits. This article introduces an efficient multi-objective constrained Bayesian optimization approach to address this challenge. Rather than attempting to determine the full set of non-dominated solutions with arbitrary trade-offs, the approach searches for a solution matching a specified trade-off. Structural design is typically conducted using computationally expensive finite element simulations, whereas Bayesian optimization offers an efficient approach for optimizing problems that involve such high-cost simulations. The proposed method integrates proper orthogonal decomposition for dimensionality reduction of simulation results with Kriging partial least squares to enhance efficiency. Constrained expected improvement is used as an acquisition function for Bayesian optimization. The approach is demonstrated through a case study of a two-lane, three-span post-tensioned concrete bridge girder, incorporating fifteen design variables and nine constraints. A comparison with conventional design methods demonstrates the potential of this optimization approach to achieve substantial cost reductions, with savings of approximately 10% to 15% in financial costs and about 20% in environmental costs for the case study, while ensuring structural integrity. Replication of Results Results from the finite element analyses can be made available upon request by e-mailing the corresponding author. The methods used for the optimization and creation of surrogate models were performed as described in the article.
BibTeX
@inproceedings{rstum2025efficient,
  author = {Heine Havneraas Røstum and Joseph Morlier and Sébastien Gros and Ketil Aas-Jakobsen},
  title = {{Efficient multi-objective constrained Bayesian optimization of bridge girder}},
  booktitle = {Structural},
  year = {2025},
  doi = {10.1007/s00158-025-04157-x},
}
Figure