Abstract
This paper presents an efficient Bayesian optimization framework designed to minimize both the financial and environmental costs of bridge girders. By integrating Proper Orthogonal Decomposition (POD) for dimensionality reduction and Kriging Partial Least Squares (KPLS) for surrogate modeling, the proposed approach achieves a 20% reduction in environmental impact and a 10-15% reduction in financial costs compared to traditional engineering benchmarks. The framework demonstrates high sample efficiency, reaching stable designs in significantly fewer iterations than standard methods.
Introduction
The construction industry is a major contributor to global greenhouse gas emissions, with cement production alone accounting for approximately 7% of the global total. Despite the growing need for sustainable infrastructure, traditional bridge design often relies on experience-based "rules of thumb" that prioritize safety and cost over environmental impact.
Optimizing bridge designs is computationally demanding, requiring high-fidelity Finite Element (FE) simulations that are both time-consuming and memory-intensive. Furthermore, balancing competing objectives—such as minimizing financial cost while reducing carbon footprint—presents a complex multi-objective challenge. Standard optimization methods typically require hundreds of simulations, making them impractical for large-scale infrastructure projects.
Methodology
The authors propose a specialized "stack" of machine learning and optimization techniques to address these challenges efficiently.
Optimization Framework
The framework utilizes Bayesian Optimization (BO) to build a surrogate model that predicts design performance without exhaustive simulation. Key components include:
- Proper Orthogonal Decomposition (POD): Used to compress high-dimensional functional outputs (e.g., stress distributions) from FE simulations into a few essential modes, simplifying the surrogate model training.
- Kriging Partial Least Squares (KPLS): A specialized Gaussian Process Regression that handles high-dimensional design spaces by reducing the number of internal parameters, ensuring accurate predictions even with limited samples.
- Preferred Trade-off Multi-Objective Optimization (PTMOO): Instead of searching for the entire Pareto front, PTMOO targets a single, balanced solution based on predefined trade-offs, drastically reducing the required number of simulations.
Case Study and Data
The framework was validated using a two-lane, three-span post-tensioned concrete bridge girder (70m length). The study utilized 15 design variables (e.g., girder height, reinforcement areas) and 9 high-dimensional constraints based on Eurocode standards. Initial data was generated using 30 FE analyses via Latin Hypercube Sampling (LHS).
Results and Discussion
The proposed method demonstrated superior performance in both efficiency and design quality:
- Economic and Environmental Gains: The optimized design achieved a 10-15% reduction in financial costs and a 20% reduction in carbon footprint compared to traditional benchmarks.
- Convergence Efficiency: The KPLS-based approach reached a stable, high-quality design in approximately 20 iterations, significantly faster than standard methods like cEHVI.
- Structural Utilization: Optimized designs showed higher material utilization ratios, suggesting that traditional designs are often over-engineered, leading to material waste without proportional safety benefits.
Conclusion
This research highlights the potential of combining dimensionality reduction with specialized surrogate models to solve complex structural optimization problems. While future work may explore Deep Gaussian Processes or direct integration of all Eurocode reinforcement rules, the current framework provides a robust path toward more sustainable and cost-effective infrastructure design.
Summary updated