Machine Learning Approaches for the Reliable Design of Steel Joints

Abstract This thesis presents a framework for developing machine-learned prediction models design of steel connections, with a focus on maintaining the reliability level required by Eurocode design procedures. The work responds to the increasing use of finite element analysis (FEA) in structural joint design, where traditional analytical or empirical design equations are often insufficient to describe nonlinear and interaction-dependent behavior. The study outlines the motivation to replace simplified analytical expressions with data-driven models trained on validated numerical simulations, while preserving reliability in line with the requirements of structural engineering design practice. A column web in transverse compression is selected as a representative example to illustrate the proposed workflow. The procedure begins with automated generation of a large dataset through parametric numerical design calculations using IDEA StatiCa Connection, where more than 9~000 FE models were computed. The dataset covers a broad range of geometric and material parameters, and each simulation was evaluated according to a newly developed equivalent plastic strain limit criterion. The definition of this strain limit is one of the central contributions of the thesis, as it directly influences the calculated resistance and the consistency of the resulting machine learning models. A detailed discussion is provided on the identification of a reliable equivalent plastic strain limit for steel structural components. The thesis compares recommendations given in different parts of Eurocode~3 and highlights the inconsistencies that currently exist. A new plastic strain limit proposal is derived based on numerous physical tests, numerical simulations and numerical design calculations for the net section failure of plates with notches and bolt holes, which is deemed to be the governing case. The proposed limit is clear and simple; it enables reproducible resistance determination in numerical design calculations and serves as a solid basis for subsequent data-driven modeling. The thesis further introduces a complete process for the verification, validation, and reliability assessment of machine learning models in structural engineering applications. The process is based on the principles of EN~1990 and the draft standard FprEN~1993-1-14, ensuring that predictive models meet equivalent reliability levels to traditional design methods. Neural networks were selected as the primary modeling technique due to their ability to approximate complex nonlinear relationships. Using Python-based automation and systematic hyperparameter tuning, models were trained to predict resistances with a coefficient of variation below 2~\% compared with FEA results. The study also investigates the influence of dataset coverage, feature scaling, and model regularization on prediction accuracy and reliability. The results demonstrate that when trained on sufficiently rich and consistent data, machine learning models can replicate the results of detailed numerical simulations with no bias and negligible coefficient of variation. However, the reliability of such models remains limited to the range of input parameters represented in the training dataset. The thesis emphasizes that careful dataset generation, transparent preprocessing, and continuous verification are prerequisites for trustworthy machine learning applications in structural engineering. A discussion is provided on extending the range of validity through geometric similarity scaling. The thesis concludes with recommendations for how machine learning can be integrated into structural design practice while keeping the same level of safety and transparency required by engineering standards. Several directions for future research are suggested: (i) improving the quality of finite element models used for training, including possible expansion through geometric similarity scaling or the use of simplified models with coarser meshes trained to produce high-quality results; (ii) enhancing reliability assessment by considering objectively safer FEA modifications, such as reducing the plastic strain limit or neglecting strain hardening; and (iii) creating well-documented datasets of experimental studies and numerical simulations, and encouraging researchers to share their results -- a practice increasingly important in today’s data-driven research environment. The developed framework helps to connect advanced numerical simulations with practical design procedures and opens the path toward the next generation of efficient and reliable design tools.