Prediction of axial compressive capacity of

CFRP-confined concrete-filled steel tubular

short columns based on XGBoost algorithm
Author links open overlay panel
,,,
https://doi.org/10.1016/j.engstruct.2022.114239Get rights and content

Abstract
Carbon fiber reinforced polymer (CFRP) is widely used in the field of strengthening concrete-
filled steel tubular (CFST) columns due to its excellent performance. However, the interactions
between steel tube and concrete, and between steel tube and CFRP are very complex, which
make it difficult to predict the strength. To solve this problem, a new method using an advanced
machine learning algorithm, XGBoost, to predict the axial compressive capacity of CFRP-
confined CFST short columns is proposed. The data set for training and testing contains 379
pieces of data out of two parts: 271 from the literature and 108 from our experiments presented
in this paper. Firstly, the typical failure modes and stress mechanisms of the specimens and the
effects of the number of CFRP layers, the strength of core concrete and section forms on axial
compressive capacity are analyzed through experiments. Then, the machine learning is carried
out with the material strengths (concrete, steel, CFRP), the cross-sectional areas of the materials
(concrete, steel tube, CFRP) and the cross-sectional shapes of the specimen as the features, and
the axial compression bearing capacity of the specimen as the labels. Eight algorithms: Linear
Regression (LR), K-Nearest Neighbor (KNN), Support Vector Machine (SVM), and typical
ensemble learning models such as Random Forest (RF), Adaptive Boosting (AdaBoost),
Gradient Boosting Decision Tree (GBDT), Extreme Gradient Boosting (XGBoost), LightGBM
Regressor (LGB) are selected for calculations. Based on RandomizedSearchCV and 5-fold cross
validation, the performance of the eight algorithms is compared. It is found that XGBoost has the
best prediction performance, with the coefficient of determination R2 = 0.9719. Finally, based on
the learning curves and grid search, nine hyperparameters of XGBoost model are adjusted, the
optimal combination of hyperparameters of XGBoost model is obtained. The results show that
the performance of XGBoost model is improved, with the coefficient of determination
R2 = 0.9850.

Introduction
CFST can fully combine the advantages of steel and concrete. It has the characteristics of high
bearing capacity, excellent seismic and fire resistance performance, convenient construction and
good economic benefit. It is widely used in modern building structures. The outer steel tube
provides confinement for the core concrete and limits the lateral deformation of the concrete,
while the concrete can prevent the inner buckling of the outer steel tube. However, after the
concrete core of CFST is destructed, the local bending of the pipe wall usually takes place, which
may lead to the sharp decline of mechanical properties of members. FRP has high tensile
strength. When it is used to strengthen CFST, it will increase the strength of concrete core, so
that the local bending of the pipe will occur at higher loads. In addition, the aging of steel tube
will also lead to the deterioration of the performance of CFST members. FRP has good corrosion
resistance and can inhibit the aging of steel tube. Because of the above characteristics, FRP is
more and more widely used in the field of building structure reinforcement. However, the
interactions between steel tube, concrete and FRP are very complex, which makes it not easy to
predict the strength of FRP-confined CFST columns.

Many scholars have carried out extensive experimental and theoretical research on the axial
compressive static performance of FRP-confined concrete-filled circular steel tubular (CFCST)
columns [36], [37], [31], [45], [15], [22], [2], [1], [17], [34], [18], [6], [16], [24], [28], [40], [41].
Choi and Xiao proposed a theoretical model to predict the axial compressive properties of
CFCST [4], and then, they proposed a theoretical model to predict the axial compressive
properties of CFRP-confined CFCST columns [5]. Ding et al. proposed a full elasto-plastic
model and a simplified model for CFCST stub columns under concentrically loaded conditions
[8]. Based on the limit equilibrium method and elastic–plastic analysis method, Ding et al.
proposed a simplified formula for calculating the ultimate bearing capacity of CFRP-confined
CFCST columns [7]. Hu et al. conducted axial compression tests on 12 glass fiber reinforced
plastic (GFRP) confined CFCST columns with thin-walled steel tubes, and the diameter
thickness ratio of steel tube was more than 100 [11]. Tao et al. carried out the axial compression
tests of six CFRP-confined CFCST columns and three CFRP-confined concrete-filled
rectangular steel tubular (CFRST) columns, and proposed a simple model of ultimate strength of
CFRP-confined CFST short columns [29]. Yu et al. conducted the performance test of GFRP-
confined CFCST under cyclic axial compression, and proposed the cyclic stress–strain model of
core concrete [39].

Compared with circular section columns, due to the lower shear strength of FRP, CFST columns
with rectangular section are prone to failure due to the stress concentration at the corner.
Therefore, the failure mechanism of FRP-confined CFRST columns is more complex. However,
the research on FRP-confined CFRST columns is not sufficient. Park et al. carried out the axial
compression test and beam column test of concrete-filled square steel tubular (CFSST) columns
and CFRP-confined CFSST columns [20]. Prabhu and Sundarraja studied the axial compression
mechanical properties of CFSST columns wrapped with CFRP strips, and proposed a prediction
model for the bearing capacity of CFSST columns confined with CFRP strips [27], [23]. Soman
and Chandrakumar studied the bearing capacity of GFRP-confined reinforced concrete square
columns under axial compression, and proposed an analytical model for the ultimate bearing
capacity [25]. Tao et al. tested CFRST columns with different thickness of CFRP [29].Wang et
al. conducted the axial compression static test of CFRP-confined CFSST columns, and discussed
the failure mode, load stress–strain curve, and the cooperative work of steel tube and CFRP [32],
[33]. Wei et al. studied the influence of corner radius of cross-section on the axial mechanical
properties of CFRP-confined CFRST columns, and proposed a unified prediction model of
ultimate bearing capacity of CFST columns under axial compression [35]. Yan et al. investigated
the axial bearing capacity of ultra-high performance concrete filled square high-strength steel
tube short columns through experiments [38]. Zhang et al. established the whole process stress–
strain relationship model of a FRP-confined CFST columns and the calculation methods of
ultimate strain, ultimate stress, peak strain and peak stress [42]. Zhang and Xiao carried out axial
compression tests on CFST columns confined with different FRP materials (carbon fiber, glass
fiber and basalt fiber) [43].

However, the above calculation models are not universal. In recent years, the application of
machine learning in engineering has attracted more and more interest. In the aspect of CFSTs,
scholars have developed different prediction models of bearing performance based on different
algorithms. Du et al. [9] and Le et al. [12] used (artificial neural network) ANN to predict the
axial bearing capacity of CFRST columns. Tran et al. [50] used ANN to predict the axial bearing
capacity of CFCST columns with UHSC. Lyu et al. [48] applied SVR optimized by SCA to
predict the axial bearing capacities of CFCST columns. Naser et al. [49] employed (genetic
algorithms) GA evaluate structural response of CFCST and CFRST columns under axial
compression and eccentric compression. Lee et al. [13] used Catboost model to predict the
bearing capacity of CFCST and CFRST columns under axial compression and eccentric
compression.

In the aspect of FRP-confined CFSTs, however, there are few bearing capacity prediction models
based on machine learning. Güneyisi and Nour [10] used (gene expression programing) GEP
technique to predict the axial compressive capacity of CFRP/GFRP-confined CFCST columns.
The interaction between FRP-confined CFST materials is much more complex than that between
CFST materials. It is necessary to establish bearing capacity models for FRP-confined CFSTs
with different section forms based on the powerful and robust ensemble learning algorithms.

In this paper, the mechanical properties of 66 CFRP-confined CFST short columns and 42 non-
CFRP-confined CFST short columns under axial compression are studied. The specimens consist
of 57 circular section columns and 51 square section columns. Through experiments, the typical
failure modes and stress mechanisms of the specimens are analyzed. The effects of the number of
CFRP layers, core concrete strength and section forms are discussed. 271 pieces of experimental
data of CFRP-confined CFST short columns are collected from the existing literature. Combined
with the 108 pieces of experimental data of this experiment, a new mathematical formula for
calculating the bearing capacity of CFRP-confined CFST is proposed. Then, based on machine
learning, eight algorithms: LR, KNN, SVM, RF, AdaBoost, GBDT, XGBoost, LGB were
selected to predict the axial compressive capacity of CFRP-confined CFST short columns. After
comparison, it is found that XGBoost has the best prediction performance. Finally, based on the
learning curves and grid search, the parameters of XGBoost model are adjusted, the best
combination of hyperparameters is obtained, which improves the performance of XGBoost.

Section snippets
Specimen design
Considering the influence of different confinement levels, different concrete strength, different
section forms and other factors, a total of 108 specimens (51 specimens with square section and
57 specimens with circular section) were made and tested in four series. The specific parameters
of the specimens are shown in Table 1. The length of all specimens is 400 mm, the section size
of square section specimens is 110 × 110 mm, and the section diameter of circular section
specimens is 114 mm.

Axial load–displacement curves

The load–displacement curves of CFST specimens without CFRP confinement and CFRP-
confined CFST specimens are shown in Figs. 6 and 7. The load–displacement curves of all
specimens show three stages: elastic, elastic–plastic and descending stage. After the peak load,
the specimen enters the descending stage. For the specimen without CFRP confinement, the
curve begins to decline rapidly. The bearing capacity of the specimen with square section
decreases faster than that of the specimen with

Some representative models

Table 7 presents some representative formulas for calculating the ultimate bearing capacity of
CFSTs under axial compression in existing codes and literature.

Based on the existing bearing capacity model, the predicted values Ncu,pred is obtained and
compared with the real values Ncu,exp of experiments to observe their calculation accuracy. Fig. 12
shows different degrees of calculation accuracy of different existing models. As can be seen
from Fig. 12, for CFSSTs, the calculated results of all

Machine learning
The calculation formulas obtained in Section 4 are effective within the scope of this experiment.
In order to obtain a more universal method for predicting bearing capacity, machine learning is
utilized to predict the bearing capacity of CFRP-confined CFST short columns. The method
based on machine learning can be divided into the following five main steps:

Step 1: Collection of experimental data;

Step 2: Feature Engineering;

Step 3: Model construction and validation;

Step 4: Comparison of models;

Conclusions
In this paper, XGBoost is used to predict the axial compressive capacity of CFRP-confined
CFST short columns. Firstly, the mechanical properties experiments of 108 specimens under
axial compression are carried out, including 57 specimens with circular section and 51 specimens
with square section. The typical failure modes and load–displacement curves of the specimens
are obtained, and the stress mechanisms of the specimens are analyzed. The effects of the layer
number of CFRP, the strength of