Response-specific machine-learning surrogates for preliminary assessment of a parametric steel I-beam family with four circular web openings

Steel I-beams with circular web openings facilitate the passage of building services but exhibit complex deformation and localized stress-concentration behaviour. This study develops machine-learning surrogate models for rapid preliminary assessment of such beams using numerical responses generated in ANSYS. A total of 1126 parametric simulations were performed, of which 981 configurations were retained after geometric, meshing, solution, equilibrium, and response-extraction checks. Three independent single-output regression models were developed for maximum deflection, maximum web-stress demand, and maximum opening-edge stress demand. Ridge Regression, Support Vector Regression, Random Forest, XGBoost, and Gaussian Process Regression were evaluated through nested repeated cross-validation. Among the evaluated models, the dimensional Gaussian Process Regression configuration yielded the lowest mean NRMSE for maximum deflection, with a coefficient of determination of 0.784 and an NRMSE of 0.045. The dimensional XGBoost configuration yielded the lowest mean NRMSE for maximum web-stress demand, with a coefficient of determination of 0.527 and an NRMSE of 0.065. The dimensionless-augmented Support Vector Regression configuration yielded the lowest mean NRMSE for maximum opening-edge-stress demand, with a coefficient of determination of 0.509 and an NRMSE of 0.054. Regression Error Characteristic analysis and locally adaptive conformal prediction were used to examine error tolerance and predictive uncertainty.