Time-Based Fire Resistance Performance of Axially Loaded, Circular, Long CFST Columns: Developing Analytical Design Models Using ANN and GEP Techniques

Abstract

Concrete-filled steel tube (CFST) columns are composite structural elements preferred in various engineering structures due to their superior properties compared to those of traditional structural elements. However, fire resistance analyses are complex due to CFST columns consisting of two components with different thermal and mechanical properties. Significant challenges arise because current design codes and guidelines do not provide clear guidance for determining the time-dependent fire performance of these composite elements. This study aimed to address the existing design gap by investigating the fire behavior of circular long CFST columns under axial compressive load and developing robust, accurate, and reliable design models to predict their fire performance. To this end, an up-to-date database consisting of 62 data-points obtained from experimental studies involving variable material properties, dimensions, and load ratios was created. Analytical design models were meticulously developed using two advanced soft computing techniques: artificial neural networks (ANNs) and genetic expression programming (GEP). The model inputs were determined as six main independent parameters: steel tube diameter (D), wall thickness (ts), concrete compressive strength (fc), steel yield strength (fsy), the slenderness ratio (L/D), and the load ratio (μ). The performance of the developed models was comprehensively compared with experimental data and existing design models. While existing design formulas could not predict time-based fire performance, the developed models demonstrated superior prediction accuracy. The GEP-based model performed well with an R-squared value of 0.937, while the ANN-based model achieved the highest prediction performance with an R-squared value of 0.972. Furthermore, the ANN model demonstrated its excellent prediction capability with a minimal mean absolute percentage error (MAPE = 4.41). Based on the nRMSE classification, the GEP-based model proved to be in the good performance category with an nRMSE value of 0.15, whereas the ANN model was in the excellent performance category with a value of 0.10. Fitness function (f) and performance index (PI) values were used to assess the models’ accuracy; the ANN (f = 1.13; PI = 0.05) and GEP (f = 1.19; PI = 0.08) models demonstrated statistical reliability by offering values appropriate for the expected targets (f ≈ 1; PI ≈ 0). Consequently, it was concluded that these statistically convincing and reliable design models can be used to consistently and accurately predict the time-dependent fire resistance of axially loaded, circular, long CFST columns when adequate design formulas are not available in existing codes.