Eurocode-Compliant Optimal Design of Steel Structures

The built environment is responsible for a large share of global greenhouse gas emissions. About 25 % of these emissions is due to the manufacturing and processing of structural building materials. Nevertheless, in current building practice, large amounts of structural material are wasted, as load-bearing structures are often grossly overdimensioned. Numerical optimization tools have the potential to reduce both material consumption and engineering costs, yet structural engineers are reluctant to use structural optimization tools in everyday engineering practice.

Formulating a design problem in engineering practice as a structural optimization problem is challenging: such problems are characterized by a mixture of discrete and continuous design variables and involve large numbers of complex building code constraints. In the academic literature, structural design optimization problems are usually either simplified or solved by means of metaheuristic methods (often genetic algorithms). Metaheuristic methods have the advantage of being universally applicable and the disadvantage of being relatively inefficient, and the consensus is that gradient-based optimization methods are better suited in case the design variables are continuous.

This doctoral study aims to promote the adoption of numerical optimization tools by structural engineers, by (1) investigating the effectiveness of a genetic algorithm in solving real-world structural design optimization problems and by (2) developing new optimization strategies for situations where the genetic algorithm proves inadequate. The main developments are (1) a new method for performance assessment and parameter tuning of metaheuristic algorithms, (2) a new gradient-based/metaheuristic optimization strategy to solve problems with mixed continuous/discrete design variables, and (3) a new heuristic rounding technique designed to efficiently solve discrete size optimization problems. The thesis focuses on steel structures. In order to accurately simulate a real design situation, two realistic benchmark problems are formulated, with the emphasis on taking into account all relevant design rules from the Eurocodes. Together with a set of simpler test problems, these realistic benchmark problems are used to test existing and new optimization algorithms.

The results of the study show that genetic algorithms are adequate for solving real-world structural design optimization problems, provided that the number of design variables is limited. With optimal control parameter values, determined by the performance assessment method, the genetic algorithm shows significantly better performance on the simpler test problems than with default control parameter values, although the difference decreases the longer the genetic algorithm is allowed to run. The performance assessment method is too computationally expensive to be applied directly to the realistic benchmark problems, and it proves difficult to extrapolate trends in well-performing control parameter values found for the simpler test cases. The heuristic rounding technique generally outperforms the genetic algorithm, and is also more efficient in terms of the number of objective and constraint function evaluations. Finally, the hybrid gradient-based/metaheuristic optimization strategy performs significantly better than the conventional genetic algorithm for size optimization problems combined with shape and topology optimization, characterized by a combination of continuous and discrete design variables.