Interpretable soft computing predictions of elastic shear buckling in tapered steel plate girders

The complexity of the shear buckling in tapered plate girders has motivated researchers to conduct experimental and numerical investigations to understand the underlying mechanisms controlling such phenomenon, and subsequently develop related design-oriented expressions. However, existing predictive models have been developed and validated using limited datasets and/or traditional regression techniques—restricting both the model utility, when considering a wider range of design parameters, and the model generalizability, due to associated uncertainties. To address these issues, the present study employed a powerful soft computing technique—multi-gene genetic programming (MGGP), to develop design expressions to predict the elastic shear buckling strength of tapered end plate girder web panels. A dataset of 427 experimental and experimentally validated numerical results was used in training, validating, and testing the developed MGGP models. Guided by mechanics and findings from previous studies, the key parameters controlling the strength were identified, and MGGP were employed to reveal the interdependence between such parameters and subsequently develop interpretable predictive models. The prediction accuracy of the developed models was evaluated against that of other existing models using various statistical measures. Several filter and embedded variable importance techniques were used to rank the model input parameters according to their significance in predicting the elastic shear buckling strength. These techniques include the variable importance random forest and the relative influence gradient boosting techniques. Moreover, partial dependence plots were employed to explore the effect of the input variables on the strength. The results obtained from this study demonstrated the robustness of the developed MGGP expression for predicting the elastic shear buckling strength of tapered plate girder end web panel. The developed model also exhibited a superior prediction accuracy and generalizability compared to currently existing ones. Furthermore, the developed partial dependence plots facilitated interpreting the influence of all input variables on the predicted elastic shear buckling strength.