Traditional and Machine-Learning Numerical Models for Partial-Strength Extended Endplate Connections

Steel joints with partial strength extended endplate are widely used in conventional construction and in regions with low to moderate seismicity. It was shown recently that many of the existing analytical, mechanical and empirical models fail in accurately predicting the connection behavior, yielding inconsistent predictions with large errors exceeding ±100%. The accurate characterization of the moment-rotation response of steel joints is necessary for conducting accurate structural analyses and to achieve efficient and robust designs. Accordingly, a recently collated large experimental database is used to develop more accurate numerical models. Traditional approaches, such as multivariate regression analysis, as well as machine-learning approaches, such as neural networks and decision trees, are employed to reach this objective and demonstrate the differences between the two approaches. This study outlines the procedure used to identify the significant features controlling the connections key response parameters (elastic stiffness, post-yield stiffness and plastic strength) and to regress or train the mathematical models. The models’ performance is then evaluated considering the observed error metrics and the advantages and disadvantages of each model. The new models demonstrated an improved accuracy, compared to currently available alternatives.