Strength prediction of concrete-filled steel tubular columns using Categorical Gradient Boosting algorithm

Abstract Due to complexities from the interaction between steel tube and concrete filling of concrete-filled steel tubular (CFST) columns, their strengths are very complicated, which is a highly nonlinear relation with material strengths and geometry. Categorical gradient Boosting (CatBoost), which is advanced boosting machine, is presented to solve the problems. A total of 3103 tests, which is divided in four datasets, is trained and tested the learners to determine the ultimate axial strength as the output variable while the strength of materials (concrete and steel) and geometry (e.g., diameters/width/heights, thickness, effective length, eccentricities) are the input ones. The comparison of the present results from 10-fold cross validation and those from the code predictions (AISC 360-16, Eurocode 4 and AS/NZS 2327) and previous study shows very high prediction accuracy in terms of coefficient of determination (R2), which is the lowest value (R2 = 0.964) for Dataset 2 and the highest one (R2 = 0.996) for Dataset 1. While the predictions from three codes beyond material limit and slenderness are less conservative than those within it, CatBoost provides nearly similar experiment results with the mean values as unity without any limits. This algorithm can be used to predict an accurate strength of CFST columns.