The construction industry is a significant driver of carbon-related emissions and climate change. In response, the field of structural engineering and design has been focusing on structural efficiency with regards to material choice, deployment, and design utilization. The use of high-strength and high-performance steel grades in steel structures has gained popularity due to their potential for significant savings in steel tonnages and overall emissions. However, traditional analysis and design methods have not fully adapted to these developments, leaving room for more accurate methods to fully exploit the material benefits of modern steels.
To address this need, recent research has focused on the optimization and inclusion of numerical analysis methods in the design of steel structures. The Hollosstab project, carried out by the steel and composite structures research group at ETH Zurich from 2016 to 2019, aimed to overcome code-related shortcomings in the design of hollow sections made of mild and high-strength steel.
Building upon the work of the Hollosstab project, a new thesis from ETH Zurich has developed a novel, computer-aided, data-driven approach for the analysis and design of large-scale steel structures. The approach, called DNN-DSM, utilizes Deep Neural Networks (DNN) to predict the non-linear deformation path of steel structures, including the pre- and post-buckling range. By using beam finite elements that mimic the behavior of advanced shell finite element models, the approach combines the accuracy and precision of shell element analysis with the computational efficiency of beam element analysis.
The thesis demonstrates the feasibility of the DNN-DSM method by conceptualizing and developing it to a mature degree. It focuses on individual load cases and the simpler scenario of hollow sections loaded in bending about a single axis. The methodology includes a review of traditional and advanced methods in structural steel design, as well as the use of machine learning in engineering sciences. Physically validated simulations of local buckling performance are used to train the DNNs for stiffness and strength prediction in a beam finite element formulation. To create the DNN-DSM method, a bespoke simulation tool is built using the programming language python. Finally, the DNN-DSM method is validated against benchmark shell finite element models and conventional code-based steel design.
The thesis concludes by outlining the future steps necessary to advance the method from feasibility demonstration to industrial implementation for a wide range of steel structure applications.
This research represents an important development in the field of structural engineering and design, offering a more accurate and efficient method for analyzing and designing high-strength steel structures. By leveraging data-driven, machine-learning techniques, the potential of modern steels as structural materials can be fully exploited, leading to further reductions in emissions and environmental impact in the construction industry. With ongoing advancements in numerical simulation and analysis, the industry is poised to continue improving the structural efficiency and sustainability of steel structures.
