COMPARISON AND TUNING OF MACHINE LEARNING MODELS FOR COLUMN WEB PANEL IN TRANSVERSE COMPRESSION

Abstract

<p>The component method enables efficient design of steel joints by decomposing them into individual components characterized by their resistance and stiffness. This paper presents a machine learning-based approach for resistance prediction of the column web panel in transverse compression, based on a numerical dataset generated using IDEA StatiCa Connection. Three machine learning models are investigated and compared: a deep neural network, Random Forest, and XGBoost. Particular attention is given to the influence of model tuning on predictive performance, where the deep neural network is tuned using a manual sensitivity-based procedure, while Random Forest and XGBoost are tuned using grid search. The results highlight differences in model behavior, sensitivity to tuning, and overall prediction accuracy. The study emphasizes the importance of appropriate model selection and tuning when applying machine learning models to resistance prediction problems in steel joint design.</p>