Strength prediction of circular CFST columns through advanced machine learning methods

Concrete-filled steel tube (CFST), well recognized for its excellent mechanical behaviour and economic efficiency, is widely used as a main load-carrying component in various kinds of structures. Machine learning (ML) is one of the promising artificial intelligence methods which just starts to be utilized for the advanced prediction of structural performances. This paper attempts to evaluate the feasibility of combining mechanism analysis to optimize ML models in predicting the axial compression strength of circular CFST columns. A comprehensive database containing 2,045 circular CFSTs under axial loading was established through extensive literature survey. Based on correlation analysis and mechanism analysis, input parameters for ML models were rationally selected. Then back-propagation neural network (BPNN), genetic algorithm (GA)BPNN, radial basis function neural network (RBFNN), Gaussian process regression (GPR) and multiple linear regression (MLR) models were established. It was revealed that the established ML models, especially GPR, could reliably predict the strengths of CFST with higher accuracies and wider applicable ranges than existing methods in current design standards. By subdividing the database according to column slenderness, ML models achieved improved accuracy for strength prediction, whilst little effect on the model accuracy was generated by random subdivisions. This indicates that when adopting ML methods in structural engineering sector, optimization of the models can be expected on the basis of rational understanding towards the corresponding structural mechanism.

Automated summary

This study evaluated the feasibility of combining mechanism analysis with machine learning models for predicting the axial compression strength of circular CFST columns, establishing several ML models for strength prediction and finding that the GPR model had higher accuracy and wider applicability range. The performance of ML models varied slightly under different column slenderness conditions, with random subdivisions having little effect on the model accuracy.