Advanced Inelastic Analysis and Design of High Strength Steel Structures with Machine-Learning-Derived Predictive Methods

Abstract

The construction industry is a significant driver of carbon-related emissions and of cli-mate change. In the field of structural engineering and design, this fact has led to a stricter focus on structural efficiency with respect to material choice, deployment and design utilization. For steel structures in particular, recent years have seen a push to-wards the use of high-strength and other high-performance steel grades, which poten-tially allow for significant savings in steel tonnages and thus in overall emissions and grey energy caused during the fabrication of structures and its components. However, traditional structural analysis and design methods and codes have not fully kept pace with these developments and are known to leave important margins of unexploited uti-lization. There is thus a need for more accurate analysis and design methods in order to fully exploit the material benefits of modern steels as a structural material. As steel structures are typically relatively thin-walled and slender, instabilities in the elastic and plastic range play a particularly large role in the description of their their load-carrying capacity and, ultimately, of their non-linear load-deformation behaviour. Continuous advances in numerical simulation of metallic structures has opened the door to ever more precise, simulation-based design methods for steel structures. Recent research work thus focused on various aspects of the optimization and the inclusion of numerical analysis methods in the design tasks and methods for steel members and sec-tions prone to instabilities. A project carried out by current members of the steel and composite structures research group at ETH Zurich, the RFCS-funded project Hollosstab (2016-’19), also formed the basis and defined the scope for the work carried out in this thesis: overcoming code related shortcomings in the design of hollow sections and struc-tures composed thereof, made of mild and high-strength steel. However, significantly broadening the ambitions and methodological scope of Hollosstab, this thesis set out to exploit the potential of data-driven, machine-learning techniques in the analysis and de-sign of steel structures of this type. Accordingly, the present thesis is concerned with the development and establishment of a novel, computer-aided, data-driven approach for the analysis and design of large-scale steel structures capable of predicting the entire non-linear deformation path, i.e., the pre- and post-buckling range, using beam finite elements that mimic the behaviour of ad-vanced shell finite element models of closed RHS/SHS steel profiles. The approach pro-posed and developed in this thesis, which is referred to as DNN-DSM, uses Deep Neural Networks (DNN) to predict the non-linear stiffness matrix terms in a beam element for-mulation suitable for implementing within the Direct Stiffness Method (DSM). The ap-proach is intended to allow developers of structural analysis software to combine the accuracy and precision of shell element analysis with the computational efficiency of beam element analysis, while allowing local, slenderness-dependent instability phe-nomena to be considered directly in the analysis. Within the thesis, the methodology is conceptualized and developed to a degree of ma-turity that allowed to demonstrate its feasibility. It thereby applies a restricted scope, i.e. for individual load cases and the simpler case of hollow sections loaded in bending about a single axis. It pursues its objectives in a manner that is reflected in the thesis structure: i. Firstly, it reviews the traditional and advanced methods in structural steel de-sign, and the use of machine learning in the engineering sciences, identifying their potential and shortcomings. ii. Subsequently, it makes use of physically validated, non-linear (GMNIA) simula-tions of the local buckling performance of hollow sections as key component of the data development for DNN training. iii. It then carries out the necessary steps to “train” DNNs for the prediction of stiff-nesses and strength values in a beam finite element formulation. iv. It formulates the “deep neural network direct stiffness method” (DNN-DSM), by creating a bespoke simulation tool and using the programming language python. v. Finally, it validates the DNN-DSM method with more sophisticated benchmark shell finite element models that explicitly capture local buckling, as well as with conventional, code-based steel design and further experimental results from lit-erature. The thesis is wrapped up by an outlook on further steps and a widening of scope that will be needed to advance the method from the feasibility demonstration stage to indus-trial implementation for a wide range of steel structure applications.