Torsional Design Of CFST Columns Using A Data-Driven Optimization Approach

In practical engineering,Concrete-filled steel tube(CFST)columns under the action of self-weight together with the bending and torsion moment caused by wind or earthquake load will be subjected to several load combinations(i.e.,compression,bending,and torsion).However,it is difficult to establish a tranditional theoretical model for predicting the torsional strength of(CFST)column reflecting the combined load state.In order to obtain a prediction model for the torsional strength of CFST columns with higher computational accuracy,based on transfor learning algorithms,this paper successfully predicts the torsional strength of steel pipe concrete columns with high accuracy and ease the overfitting of the tranditional machine learn model.Moreover,an end-to-end optimization framework for the optimal design of the torsional performance of CFST columns is designed,which can quickly and automatically give the design parameters of the CFST columns: such as steel tube thickness,concrete strength,section size et al.,according to the torsion design objectives.The main work and related results are as follows:The dataset used in this study compiled from existing literature,and the most influencing parameters and the ultimate torsion strength of CFST columns are identified as inputs and outputs of the ML model,respectively.Preliminary assessment of traditional ML models: Using the 199 experimental data,a preliminary analysis was conducted based on the selected 16 types of learning algorithms.A new sensitivity analysis method based on the SHAP approach is used to interpret the XGBoost model,which studies the sensitivity of each parameter to the ultimate torsion strength.These results need to keep by parametric studies found in the literature.Transfer learning algorithm is adopted to reduce the over-fitting problem of machine learning.The results show that the proposed approach can transfer useful information from the source domain(simulation data set)to the target domain(physical experiment data set),effectively reducing the small sample bias of the target domain(physical experiment data set).The generated two-stage Tr Ada Boost model incorporated with nondominated sorting genetic algorithm II(NSGA-II)is used to optimize structure design.The Pareto fronts of the two objectives(ultimate torsion strength and cost)are successfully obtained.