Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice Materials (Basel) . 2026 Mar 30;19(7):1360. doi: 10.3390/ma19071360 Search in PMC Search in PubMed View in NLM Catalog Add to search Research on the Prediction Method for Ultimate Bearing Capacity of Circular Concrete-Filled Steel Tubular Columns Based on Random Search-Optimized CatBoost Algorithm Zhenyu Wang Zhenyu Wang 1 Institute of Defense Engineering, AMS, PLA, Beijing 100850, China; [email protected] (Z.W.); Conceptualization, Software, Data curation, Writing – original draft Find articles by Zhenyu Wang 1 , Yunqiang Wang Yunqiang Wang 2 School of Water Conservancy and Transportation, Zhengzhou University, Zhengzhou 450001, China Conceptualization, Methodology, Writing – review & editing Find articles by Yunqiang Wang 2, * , Xiangyu Xu Xiangyu Xu 1 Institute of Defense Engineering, AMS, PLA, Beijing 100850, China; [email protected] (Z.W.); Methodology Find articles by Xiangyu Xu 1 , Zihan Zhang Zihan Zhang 1 Institute of Defense Engineering, AMS, PLA, Beijing 100850, China; [email protected] (Z.W.); Software Find articles by Zihan Zhang 1 , Yaxing Wei Yaxing Wei 1 Institute of Defense Engineering, AMS, PLA, Beijing 100850, China; [email protected] (Z.W.); Supervision, Data curation Find articles by Yaxing Wei 1 , Dan Luo Dan Luo 1 Institute of Defense Engineering, AMS, PLA, Beijing 100850, China; [email protected] (Z.W.); Formal analysis, Investigation Find articles by Dan Luo 1 Editor: Oldrich Sucharda Author information Article notes Copyright and License information 1 Institute of Defense Engineering, AMS, PLA, Beijing 100850, China; [email protected] (Z.W.); 2 School of Water Conservancy and Transportation, Zhengzhou University, Zhengzhou 450001, China * Correspondence: [email protected] Roles Zhenyu Wang : Conceptualization, Software, Data curation, Writing – original draft Yunqiang Wang : Conceptualization, Methodology, Writing – review & editing Xiangyu Xu : Methodology Zihan Zhang : Software Yaxing Wei : Supervision, Data curation Dan Luo : Formal analysis, Investigation Oldrich Sucharda : Academic Editor Received 2026 Mar 4; Revised 2026 Mar 21; Accepted 2026 Mar 23; Collection date 2026 Apr. © 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license . PMC Copyright notice PMCID: PMC13074762 PMID: 41976648 Abstract With the development of various emerging structures, concrete-filled steel tubular (CFST) columns have become critical load-bearing components in key infrastructures such as subways and underground utility tunnels. Accurately predicting their ultimate bearing capacity ( N u ) is essential for guaranteeing structural safety. To address the limitations of traditional empirical formulas and code-based calculation approaches, this paper proposes a prediction model for ultimate bearing capacity based on the CatBoost algorithm optimized by Random Search. Furthermore, the marginal contribution of each key feature to the prediction results is measured through interpretability analysis. First, a database containing 438 CFST column ultimate bearing capacity test cases was established, with key parameters such as geometric dimensions and material properties as input variables. Second, the predictive performance of six machine learning algorithms—CatBoost, LightGBM, Random Forest (RF), Gradient Boosting (GB), K-Nearest Neighbors (KNN), and XGBoost—was compared. A five-fold cross-validation integrated with a Random Search strategy was employed for joint hyperparameter optimization. The results show that the optimized CatBoost model significantly outperforms other algorithms and conventional design codes, achieving a coefficient of determination ( R 2 ) as high as 0.99 and a root mean square error (RMSE) of 174.29 kN. Furthermore, the SHAP (Shapley Additive exPlanations) method was used to perform global and local interpretability analyses of the prediction model. This not only quantified the individual contribution and interaction effects of each feature parameter on the bearing capacity but also revealed that geometric parameters are the primary influencing factor. This finding confirms a high degree of consistency between the prediction mechanism of the data-driven model and classical mechanical theories, effectively validating the model’s reliability. This study provides an efficient and reliable tool for the optimal design and rapid evaluation of CFST columns and establishes a new data-driven paradigm for the design and reinforcement of key components in underground structures. Keywords: concrete-filled steel tube (CFST), ultimate bearing capacity, machine learning, hyperparameter optimization, interpretability 1. Introduction Concrete-filled steel tube (CFST) structures have become a preferred structural form in high-rise buildings and underground space development owing to their exceptional load-bearing capacity, ductility, and cost-effectiveness [ 1 , 2 , 3 ]. Particularly in underground infrastructure, such as deep excavation support systems and subway station columns, CFSTs significantly enhance the structural capability to withstand high in situ stress environments by leveraging their compact cross-sections and the effective confinement provided by the steel tube to the concrete [ 4 ]. Nevertheless, the complex environments and demanding load scenarios typical of underground engineering place stringent requirements on the accuracy and reliability of CFST capacity predictions. Conventional approaches primarily rely on empirical design formulas and finite element modeling (FEM). Empirical formulas are usually calibrated on limited datasets and may be overly conservative or unreliable under complex loading conditions. Conversely, while FEM can achieve high fidelity, it comes at the cost of substantial computational effort, restricting its utility for rapid design iteration. Accordingly, there is a clear need for a prediction method that delivers both high accuracy and high efficiency, which is of significant value for engineering design. In recent years, the accumulation of experimental data has substantially enriched research on the mechanical behavior of CFST structures. Scholars such as Le et al. [ 5 ], Li et al. [ 6 , 7 ], and Wang et al. [ 8 ] have systematically investigated the effects of aspect ratio, complex cross-sectional geometries, long-term loading, and high-temperature environments on CFST bearing capacity through extensive experimentation and numerical simulation. These investigations have not only clarified fundamental resistance mechanisms but also established high-quality datasets that underpin subsequent model development. Based on these datasets, data-driven approaches have been increasingly adopted in this field [ 9 , 10 ]. Machine learning (ML), leveraging its robust nonlinear mapping capabilities, has been widely adopted for structural performance prediction [ 11 , 12 , 13 , 14 , 15 ]. For instance, Mohammad et al. [ 16 ] compared six algorithms, including SVM and ANN, demonstrating that ML models outperform existing design codes in accuracy. Lou et al. [ 17 ] proposed a knowledge-enhanced learning framework to improve prediction precision for eccentrically loaded columns, while Wei et al. [ 18 ] derived simplified design formulas based on ANNs. Kazemi et al. [ 19 ] developed an ensemble learning framework that combines synthetic data generation with heuristic optimization, achieving precise predictions of the full mechanical curves and axial compressive capacity of concrete-filled timber-steel tube (CTFST) composite members. This work represents the latest advancements in utilizing ensemble learning to handle increasingly complex composite components and to predict the entire mechanical response process. However, significant gaps remain. Existing studies typically rely on comparisons of single or limited algorithms, lacking systematic evaluations that encompass advanced ensemble methods such as CatBoost and LightGBM. Moreover, standardized protocols for hyperparameter optimization are often overlooked. Furthermore, while current research focuses heavily on enhancing prediction accuracy, it frequently neglects model interpretability [ 20 , 21 , 22 , 23 ]. Although interpretability tools such as SHAP have been introduced in related fields [ 24 , 25 ], quantitative insights into how critical variables govern CFST capacity through nonlinear and interactive mechanisms remain insufficient, particularly under conditions typical of underground engineering. This inherent “black-box” opacity erodes engineering confidence in model outputs, thereby constraining the practical adoption of ML-based approaches in structural design. This paper presents a novel framework for predicting the axial compression capacity of CFST that balances computational accuracy with engineering transparency. Utilizing a validated database of 438 specimens, a rigorous benchmarking of six ensemble learning algorithms was performed. The CatBoost model, optimized via random-search hyperparameter tuning and validated through a five-fold cross-validation scheme, demonstrated superior predictive performance. Furthermore, to mitigate the opacity of data-driven approaches, SHapley Additive exPlanations (SHAP) were utilized to interpret the model, explicitly quantifying the nonlinear impacts and synergistic effects of geometric and material variables on axial capacity. By bridging the gap between stochastic modeling and structural mechanics, this framework serves as a vital instrument for the reliability-based design of CFST structures in underground engineering contexts. 2. Dataset and Feature Analysis 2.1. Database Compilation A comprehensive database comprising 438 CFST specimens under axial compression was established by compiling 38 independent experimental datasets from Ref. [ 26 ] and the original data sources are provided in the Appendix A . Each data point in the repository is characterized by eight parameters. The input features include the external diameter ( D /mm), wall thickness ( t /mm), column length ( L /mm), concrete size effect coefficient ( γ U ), yield strength of the steel tube ( f y /MPa), compressive strength of the concrete prism ( f pr /MPa, based on 150 mm × 150 mm × 450 mm dimensions), and the confinement factor ( Φ ). The ultimate load-carrying capacity ( N u ) serves as the target output variable. The dataset was randomly split into a training set and a test set using a 70:30 ratio, yielding 307 and 131 samples, respectively. The training set was used for model fitting and hyperparameter optimization, while the test set was reserved for evaluating the generalization performance of the developed models. 2.2. Statistical Description of the Database Table 1 summarizes the descriptive statistics of the dataset parameters, including the mean, standard deviation, extrema (minimum and maximum), median, and quartiles. The statistical results indicate a wide distribution of feature values, effectively covering the typical parameter ranges encountered in practical CFST engineering applications. Notably, the target variable Nu exhibits a mean of 1917.44 kN, with a variation ranging from 273 kN to 9835 kN. This significant span demonstrates the high representativeness of the dataset, which encompasses diverse samples ranging from small-scale laboratory specimens to large-scale structural members. The frequency distributions of these parameters are illustrated in Figure 1 . Table 1. Summary statistics of the database parameters. Features D /mm t /mm L /mm γ U f y /MPa f pr /MPa Φ N u /kN Max 361 11.88 1113 1.14 853 65.60 10.21 9835 Min 48 1.00 153 0.92 223 15.28 0.23 273 Mean 155.77 4.18 464.81 1.01 352.95 37.41 1.30 1917.44 Std. Dev 59.03 1.98 194 0.04 100.48 12.02 0.96 1552.45 Median 140 3.56 419 1.0 347.90 34.72 1.09 1580 25% 109.66 2.97 312 0.99 303.50 27.02 0.71 870.50 75% 172.80 5.01 572.10 1.04 369.50 44.46 1.67 2270.25 Open in a new tab Figure 1. Open in a new tab Statistical distributions of input and output features. 2.3. Feature Correlation Analysis To elucidate the linear dependencies between the input features and the ultimate axial capacity ( N u ), a Pearson correlation analysis was conducted, with the resulting matrix presented in Figure 2 . The analysis reveals that the D exhibits the strongest positive correlation with N u ( r = 0.90), followed closely by the L ( r = 0.85). Notably, a significant collinearity exists between D and L ( r = 0.93). This observation reflects engineering practice, where larger diameters are typically associated with longer members to satisfy code-mandated slenderness limits. Furthermore, moderate correlations are observed for t ( r = 0.66) and f y ( r = 0.46). Overall, N u demonstrates a predominant linear sensitivity to the geometric and material properties of the steel tube, underscoring their critical role in determining axial compressive capacity. Figure 2. Open in a new tab Pearson correlation coefficient matrix. 3. Theoretical Background of Machine Learning Algorithms In this study, six representative machine learning algorithms were employed to predict the ultimate axial bearing capacity ( N u ) of CFST: LightGBM, Random Forest (RF), Gradient Boosting Decision Tree (GBDT), K-Nearest Neighbors (KNN), CatBoost, and XGBoost. This selection encompasses both tree-based ensemble learning techniques and distance-based instance learning methods, thereby ensuring the capability to accommodate diverse data distributions and capture complex feature dependencies. 3.1. LightGBM Model LightGBM (Light Gradient Boosting Machine) is an efficient machine-learning framework built upon gradient-boosted decision trees. Distinguished by its computational speed and low memory usage, LightGBM implements a histogram-based algorithm that discretizes continuous floating-point feature values into discrete bins. Furthermore, it employs Exclusive Feature Bundling (EFB) to reduce feature dimensionality by bundling mutually exclusive features. Unlike the conventional level-wise growth strategy, LightGBM adopts a leaf-wise tree growth strategy with depth limitations, which typically achieves better accuracy and faster convergence by minimizing the training loss more aggressively. Additionally, Gradient-based One-Side Sampling (GOSS) is incorporated to prioritize training instances with larger gradients, thereby accelerating the learning process while maintaining accuracy. The objective function of the LightGBM model can be formulated as follows: y ^ i = ∑ k = 1 K f k x i , f k ∈ F (1) Here, y ^ i denotes the predicted value, f k represents the k -th regression tree, and F is the space of all possible trees. The training process involves minimizing the following objective function: L = ∑ i = 1 n l y i , y ^ i + ∑ k = 1 K Ω f k (2) where l is the loss function, and Ω serves as the regularization term to prevent overfitting. 3.2. Random Forest (RF) Model Random Forest (RF) is a robust ensemble learning algorithm that enhances generalization by integrating a multitude of independent decision trees through majority voting or averaging. Central to this framework is Bootstrap Aggregation (Bagging), which generates diverse training subsets via sampling with replacement. This is coupled with a random feature selection strategy at each split node to maximize the decorrelation among constituent base estimators. Notably, RF is insensitive to variations in feature scales, rendering data normalization or standardization unnecessary. Additionally, the algorithm offers an intrinsic capability to quantify feature importance by assessing the performance degradation resulting from feature permutation. The mathematical formulation of the RF model is defined as: y ^ = 1 B ∑ b = 1 B f b x (3) where y ^ is the model prediction, f b denotes the predictive function of the b -th base learner, and B represents the total number of trees in the random forest. 3.3. Gradient Boosting (GB) Model Gradient Boosting (GB) is an ensemble learning method based on the forward stagewise additive modeling framework. The algorithm iteratively improves predictive performance by fitting new base learners to the negative gradients of the loss function with respect to the current ensemble predictions, thereby correcting the errors made in previous iterations. This training process can be interpreted as performing gradient descent optimization in the function space to minimize a specified loss function. Unlike parallel ensemble strategies like Random Forest, GB adopts a serial iterative approach, forcing subsequent base learners to focus on correcting samples that the current model has not fully fitted, which significantly improves the overall prediction accuracy. Due to its compatibility with arbitrary differentiable loss functions, the GB model demonstrates excellent generalization performance and robustness in both regression and classification tasks. The mathematical expression of the GB is expressed as: F m x = F m − 1 x + γ m h m x (4) Here, F m x denotes the ensemble model after the m -th iteration, h m x represents the newly added base learner, and γ m is the learning rate controlling the step size of the update. The base learner h m x is trained to approximate the negative gradient of the loss function by minimizing the following objective: h m = arg min h ∑ i = 1 n − ∂ L y i , F m − 1 x i ∂ F m − 1 x i − h x i 2 (5) 3.4. K-Nearest Neighbors (KNN) Model The k-Nearest Neighbors (k-NN) algorithm stands as a quintessential instance-based, non-parametric learning paradigm. Fundamentally, the algorithm operates by identifying the top-k training samples closest to a specific query instance based on a predefined distance metric, subsequently deriving predictions from the target values of these neighbors. In the context of regression tasks, the output is typically determined by computing the average of the target values associated with the k nearest neighbors. Characterized as a “lazy learning” strategy, k-NN eschews an explicit training phase; instead, it retains the entire training dataset and defers computation until the inference stage, utilizing local neighborhood information. This mechanism empowers k-NN to effectively model complex nonlinear mapping relationships. However, the model’s performance exhibits high sensitivity to the hyperparameter k. Furthermore, since the inference phase necessitates exhaustive distance calculations between the query sample and the entire training corpus, the computational overhead can become prohibitive when processing large-scale datasets. The mathematical formulation for k-NN regression is expressed as: y ^ = 1 K ∑ i ∈ N K x y i (6) where N K ( x ) denotes the set of indices of the k training samples nearest to the query sample x , and y i represents the target value of the i -th nearest neighbor. 3.5. CatBoost Model CatBoost is a gradient boosting decision tree algorithm specifically engineered to handle categorical features with high efficiency. While sharing a similar mathematical foundation with standard gradient boosting frameworks, CatBoost distinguishes itself through several innovative mechanisms. It employs symmetric decision trees to mitigate overfitting and utilizes Ordered Boosting combined with Target Statistics to process categorical variables directly. This approach effectively replaces traditional One-Hot Encoding with a more efficient representation. Furthermore, by introducing unbiased gradient estimation, the algorithm addresses the issue of prediction shift, ensuring robust performance even on datasets characterized by complex categorical dependencies. For a categorical feature, CatBoost computes the statistic using the following formula: x k , j = ∑ i = 1 n y i ⋅ x i , j = x k , j + a ⋅ b ∑ i = 1 n x i , j = x k , j + a (7) where x i , j denotes the encoded value of the j -th feature of the k -th sample, and a and b are smoothing parameters. 3.6. XGBoost Model XGBoost (eXtreme Gradient Boosting) is an efficient and scalable distributed gradient boosting system designed for optimized speed and performance. By incorporating L1 and L2 regularization terms into the objective function, the algorithm enhances traditional boosting frameworks, effectively controlling model complexity and suppressing overfitting. XGBoost leverages the second-order Taylor expansion of the loss function to accelerate convergence and adopts a parallel learning scheme based on block structures, combined with pre-sorting and block compression techniques to significantly improve computational efficiency. Furthermore, its built-in sparsity-aware mechanism automatically learns the optimal default branching direction for missing values. The objective function of XGBoost is formally expressed as: L = ∑ i = 1 n l y i , y ^ i + ∑ k = 1 K γ T k + 1 2 λ ω k 2 (8) where l is the loss function, T k represents the number of leaf nodes in the k -th tree, and γ are regularization parameters, and λ denotes the leaf weights. 4. Model Optimization Strategies and Evaluation 4.1. Input Feature Standardization Performing Z-Score standardization on the dataset prior to machine learning model training is a critical step for enhancing model robustness and training efficiency. This method effectively eliminates dimensional discrepancies between features by transforming feature values into a standard normal distribution with a mean of 0 and a standard deviation of 1. Its primary functions are as follows: (1) Mitigating Numerical Instability. Z-Score standardization effectively averts the imbalance in model parameter weight distribution resulting from significant disparities in feature scales. By preventing features with large magnitudes from dominating weight updates during gradient descent, it preserves numerical stability in computations. (2) Enhancing Convergence Efficiency. By applying an approximately uniform scaling to the feature space, this method ensures that parameter dimensions exhibit comparable scales of variation during optimization, facilitating coordinated gradient updates across all directions. Consequently, the contours within the parameter search space approximate a spherical shape, which significantly improves the convergence efficiency of gradient-based optimization algorithms and accelerates model training. (3) Strengthening Feature Representation Capabilities. For learning algorithms reliant on distance metrics or inner-product structures, Z-Score standardization bolsters the model’s capacity for comprehensive multi-dimensional feature representation. By eliminating the “pseudo-importance” arising from variations in units and value ranges, it ensures that all dimensions contribute balanced weights to distance calculations, thereby more accurately characterizing the intrinsic similarity structure among samples. (4) Correcting Regularization Bias. Standardization eliminates model performance biases caused by feature scale heterogeneity. In linear models incorporating regularization terms, the absence of standardization often leads to features with larger scales being disproportionately penalized. Z-Score standardization prevents this systematic bias in parameter estimation, ensuring that regularization constraints are applied equitably across all dimensions. Extensive empirical studies indicate that, while preserving the relative distributional structure of the original data, Z-Score standardization effectively enhances model generalization performance and training stability on complex, high-dimensional datasets. Its mathematical formulation is expressed as follows: x ′ = x − μ σ (9) where μ denotes the mean of the dataset x , and σ represents the standard deviation of the dataset elements. The term x ′ corresponds to the standardized result, which approximates a standard normal distribution, specifically the x ′ ∼ N ( 0 , 1 ) distribution. 4.2. Hyperparameter Optimization Based on Random Search The performance of machine learning models relies heavily on the selection of hyperparameters, as appropriate hyperparameters can significantly enhance a model’s predictive accuracy and generalization capability. In this study, hyperparameter optimization is conducted using a five-fold cross-validation ( K = 5) approach combined with a Random Search algorithm. 4.2.1. Five-Fold Cross-Validation As illustrated in Figure 3 , five-fold cross-validation divides the training dataset into five subsets of approximately equal size. In each iteration, four subsets serve as the training data, while the remaining single subset functions as the validation data. By repeating this process with five different partition combinations, limited data resources are fully utilized, yielding a reliable estimate of model performance. Five-fold cross-validation effectively maximizes the utility of limited training data, prevents overfitting, and significantly reduces random sampling bias, thereby ensuring the stability of the evaluation results. Furthermore, through multiple rounds of validation, it provides a more robust estimate of model performance, offering more accurate evaluation metrics for hyperparameter optimization. Figure 3. Open in a new tab Schematic diagram of five-fold cross-validation. 4.2.2. Random Search Algorithm In this paper, we adopt Random Search over the traditional Grid Search method for hyperparameter optimization. Random Search operates by randomly sampling multiple sets of parameter combinations from a predefined hyperparameter space, subsequently selecting the combination that yields the best performance as the final model configuration. Compared to Grid Search, Random Search demonstrates higher computational efficiency, particularly in high-dimensional parameter spaces. Furthermore, it allows for the exploration of a broader range of parameter combinations, often identifying solutions that are comparable to or better than those found via Grid Search. In this study, the number of iterations for Random Search was set to n_iter = 100 which means that 100 sets of parameter combinations were randomly sampled and evaluated within each model’s hyperparameter space. The calculation formula for Random Search is as follows: θ * = arg × min θ ~ P ( Θ ) 1 K ∑ k = 1 K L D k v a l , A D k t r a i n , θ (10) where θ denotes the hyperparameter combination, P ( Θ ) represents the distribution over the hyperparameter space, D k t r a i n and D k v a l are the training data and validation data for the k -th fold, respectively, A is the learning algorithm, and L denotes the loss function. Distinct hyperparameter search spaces were established for the six machine learning algorithms [ 27 , 28 ], as detailed in Table 2 . These spaces are designed to cover the key tunable parameters of each model, thereby ensuring a comprehensive exploration scope for the Random Search process. Table 2. Hyperparameter spaces for the machine learning algorithms. Algorithm Hyperparameter Search Space Light GBM colsample_bytree [0.6–1.0] learning_rate [0.01–0.3] max_depth [3–15] min_child_samples [1–50] n_estimators [50–500] num_leaves [20–1500] reg_alpha [0–2] reg_lambda [0–2] subsample [0.5–1.0] Random Forest bootstrap [True, False] max_depth [3–15] max_features [0.1–1.0] min_samples_leaf [1–10] min_samples_split [2–20] n_estimators [50–500] GradientBoosting learning_rate [0.01–0.3] max_depth [3–10] max_features [0.1–1.0] min_samples_leaf [1–20] min_samples_split [2–20] n_estimators [50–500] c [0.5–1.0] KNN leaf_size [5–100] n_neighbors [1–20] p [1, 2] CatBoost n_estimators [100–2000] learning_rate [0.01–0.3] l2_leaf_reg [1–10] depth [3–10] border_count [32–255] bagging_temperature [0–10] XGBoost colsample_bytree [0.3–1.0] gamma [0–1] learning_rate [0.01–0.3] max_depth [3–10] n_estimators [50–1000] reg_alpha [0–2] reg_lambda [0–2] subsample [0.5–1.0] Open in a new tab 4.3. Evaluation Metrics for Machine Learning Models To comprehensively evaluate model performance, this study utilizes four evaluation metrics: Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), Mean Absolute Error (MAE), and the Coefficient of Determination (R 2 ). These metrics reflect the accuracy and reliability of model predictions from different perspectives. Specifically, RMSE measures the average deviation between predicted and actual values and is more sensitive to large errors. MAPE captures the ratio of prediction error to the actual value, making it suitable for evaluating performance across data of different scales. MAE measures the average absolute difference between predicted and actual values and is less sensitive to outliers. Finally, R 2 measures the proportion of variance in the target variable explained by the model. The specific calculation formulas are as follows: R M S E ( MPa ) = 1 n × ∑ i = 1 n ( y i p − y i ) 2 (11) M A P E ( % ) = 1 n × ∑ i = 1 n y i p − y i y i × 100 (12) M A E ( MPa ) = 1 n × ∑ i = 1 n y i p − y i (13) R 2 = 1 − R S S T S S = 1 − ∑ i = 1 n ( y i − y i p ) 2 ∑ i = 1 n ( y i − y ¯ ) 2 (14) 5. Results and Analysis 5.1. Basic Experimental Results To verify the performance of the proposed machine learning models in predicting the ultimate bearing capacity of CFST, basic experiments were conducted on six classic algorithms: LightGBM, RandomForest, GradientBoosting, KNN, CatBoost, and XGBoost. The experimental results are presented in Table 3 . Table 3. Simulation results on the training and validation sets. Model Evaluation Indicators Dataset RMSE MAPE MAE R 2 LightGBM Training 293.78 0.06 126.92 0.96 Validation 446.35 0.10 228.22 0.92 RandomForest Training 140.21 0.03 68.40 0.99 Validation 290.27 0.09 173.81 0.97 GradientBoosting Training 66.49 0.04 51.34 1.00 Validation 216.23 0.09 136.10 0.98 KNN Training 329.94 0.08 165.13 0.95 Validation 486.71 0.12 262.47 0.91 CatBoost Training 34.45 0.02 25.49 1.00 Validation 223.21 0.06 112.43 0.98 XGBoost Training 20.68 0.01 8.45 1.00 Validation 258.66 0.07 132.62 0.97 Open in a new tab As shown in Table 3 , the XGBoost model achieved the best performance on the training set (RMSE = 20.68 kN, MAPE = 0.01, MAE = 8.45 kN and an R 2 = 1.00). Similarly, the CatBoost and Gradient Boosting models reached an R 2 of 1.00 on the training set, indicating that these three tree-based ensemble models possess an exceptional fitting capacity for the training data. However, a comparison with the test set reveals a performance gap; for instance, the RMSE of XGBoost increased from 20.68 kN on the training set to 258.66 kN on the test set, while CatBoost increased from 34.45 kN to 223.21 kN. This suggests that under default hyperparameters, the models exhibit a degree of overfitting. This phenomenon is common in tree-based models and stems from high model complexity and insufficient regularization under default settings, leading the models to ‘memorize’ noise in the training data rather than learning the underlying mechanical laws. Additionally, the KNN model performed poorly on both sets, with an RMSE of 486.71 kN on the test set, indicating its weak adaptability to the predictive task in this study. Figure 4 displays the scatter plots of the predictions from the basic experiments. It is evident from the plots that the GB, CatBoost, and XGBoost models fit the experimental data well, as the prediction points are concentrated near the diagonal line. This demonstrates the robust ability of these three models to predict the ultimate bearing capacity of CFST members. Figure 4. Open in a new tab Prediction scatter plots of the basic experiments for six models (( a ) LightGBM, ( b ) RF, ( c ) GB, ( d ) KNN, ( e ) CatBoost, ( f ) XGBoost). 5.2. Results of 5-Fold Cross-Validation and Hyperparameter Optimization To further improve model performance, this study combined 5-fold cross-validation with random search optimization to identify the optimal hyperparameters. For each machine learning model, a specific hyperparameter space was defined, and the optimal parameter combination within that space was sought using the random search method. After random search optimization, the optimal parameter combination for the six ML models is shown in Table 4 . Table 4. The Optimal Parameter Combination for the six ML models. Algorithm Hyperparameter Optimal Parameters Light GBM colsample_bytree 0.87 max_depth 13 n_estimators 238 reg_alpha 1.56 subsample 0.56 learning_rate 0.29 min_child_samples 12 num_leaves 1075 reg_lambda 1.56 Random Forest bootstrap FALSE max_features 0.79 min_samples_split 2 max_depth 11 min_samples_leaf 1 n_estimators 288 GradientBoosting learning_rate 0.17 max_features 0.68 min_samples_split 16 subsample 0.96 max_depth 5 min_samples_leaf 11 n_estimators 298 KNN leaf_size 54 n_neighbors 1 p 1 CatBoost n_estimators 2000 l2_leaf_reg 3 border_count 128 learning_rate 0.3 depth 4 bagging_temperature 1.5 XGBoost subsample 0.6 reg_alpha 0.1 max_depth 3 gamma 0.5 reg_lambda 0.1 n_estimators 500 learning_rate 0.1 colsample_bytree 1 Open in a new tab 5.3. Comparison Between Optimized and Original Models Table 5 presents the performance comparison between the optimized model and the original model on the test set. It illustrates that the performance of all models on the test set improved after hyperparameter optimization, and the overfitting observed in the baseline experiments was significantly mitigated. The R-CatBoost model demonstrated the superior performance, with the RMSE decreasing to 174.29 kN and the R 2 increasing to 0.99, representing the highest predictive accuracy and generalization capability. The most significant improvement was observed in the R-LightGBM model, where the RMSE dropped from 446.35 kN to 272.16 kN, a reduction of 39.0%. This demonstrates that the optimization strategy—utilizing Random Search and 5-fold cross-validation—effectively controlled model complexity and achieved an optimal balance between fitting capacity and generalization performance. This confirms the critical role of hyperparameter optimization in mitigating overfitting. While the R-KNN model showed improvement, it remained the weakest performer, suggesting that the KNN algorithm may not be well-suited for the specific data features of this study. Table 5. Performance Comparison Between the Original and Improved Models on the Test Set. Model Evaluation Indicators RMSE MAPE MAE R 2 LightGBM 446.35 0.10 228.22 0.92 R-LightGBM 272.16 0.09 157.08 0.97 RandomFores 290.27 0.09 173.81 0.97 R-RandomFores 280.57 0.07 140.32 0.97 GradientBoosting 216.23 0.09 136.10 0.98 R-GradientBoosting 223.63 0.07 127.61 0.98 KNN 486.71 0.12 262.47 0.91 R-KNN 419.83 0.08 184.50 0.93 CatBoost 223.21 0.06 112.43 0.98 R- CatBoost 174.29 0.06 107.30 0.99 XGBoost 258.66 0.07 132.62 0.97 R- XGBoost 218.35 0.08 130.54 0.98 Open in a new tab Figure 5 illustrates the relationship between the actual and predicted values for all improved models. As observed from the figure, the prediction points of the R-CatBoost model are closest to the diagonal line and exhibit the most concentrated distribution, which further demonstrates its superior predictive performance. Figure 5. Open in a new tab Actual vs. Predicted Values for the Improved Models. 5.4. Comparison of This Study with Existing Literature Based on the comparative analysis of the research findings in Table 6 , the following conclusions can be drawn across three key dimensions: Table 6. Comparison of the Results of This Study with Relevant Recent Research Findings. Reference Machine Learning Methods Dataset Source & Scale Validation & Evaluation Methods This Study R-CatBoost, LightGBM, RF, GB, KNN, XGBoost 38 independent test programs; 438 data points Train/Test = 70/30; 5-fold CV; SHAP interpretability analysis Kazemi [ 19 ] Ensemble learning framework (BR, XGBoost, GBM, RF, etc.) 88 numerical models/12 tests; 2000 data points Train/Test = 80/20; Multi-dimensional metrics; GUI interface showcase Li [ 29 ] PSO-GPR, BPNN, SVR, GPR 15 independent test programs; 162 data points Experimental data of large-scale members Lyu [ 30 ] SCA-SVR, ANN, RF, MLR Experimental data; 478 data points Train/Test = 70/30; 100 random trials for stability; Inverse design parameter prediction Xie [ 31 ] XGBoost, RF, LightGBM, AdaBoost, CatBoost, LSTM 66 tests/134 numerical simulations; 200 data points Train/Test = 70/30; Taylor diagram comparison; SHAP interpretability analysis Zhao [ 32 ] (Multilevel Extension) + AHP 25 independent tests; 449 data points 3-specimen uniaxial compression test Open in a new tab (1) Data-Driven Methodology: This study relies exclusively on 438 authentic experimental cases, prioritizing the physical reliability and empirical groundedness of the data. In contrast, Kazemi [ 19 ] and Xie [ 31 ] extensively utilized Finite Element Analysis (FEA) simulation data to compensate for the scarcity of experimental records. Kazemi [ 19 ] further advanced this by incorporating synthetic data generation techniques (GANs/VAEs) to address the challenges of small-sample learning. (2) Algorithmic Evolution: There is a clear transition in the complexity of the predictive models used. Early research, such as Lyu [ 30 ], focused on the heuristic optimization of individual base models (e.g., SVR). However, more recent studies—including this work, Xie [ 31 ], and Kazemi [ 19 ]—have shifted toward Ensemble Learning frameworks. Algorithms like CatBoost and XGBoost leverage multi-model synergy to significantly enhance generalization accuracy and predictive robustness. (3) Validation Depth: This study distinguishes itself by utilizing SHAP interpretability analysis to demonstrate the scientific alignment between machine learning results and classical structural mechanics, moving beyond simple curve fitting. For comparison, Li [ 29 ] emphasized the applicability of models to large-scale external specimens, while Zhao [ 32 ] focused on assessing model accuracy through in situ experimental validation. 6. SHAP-Based Interpretability Analysis SHAP (SHapley Additive exPlanations) is a unified interpretation framework based on cooperative game theory, which is capable of measuring the marginal contribution of each feature in a single prediction. Given that the R-CatBoost model demonstrated the optimal performance among all models, this study employs SHAP values to interpret the R-CatBoost model. Specifically, the influence of each input feature on the prediction of the ultimate bearing capacity of CFST is systematically analyzed from both global and local perspectives. 6.1. SHAP Summary Plot The SHAP summary plot is illustrated in Figure 6 . The D and t exhibit the widest range of SHAP value distribution. Specifically, scatter points corresponding to high feature values are predominantly located in the positive SHAP region, whereas those for low values are concentrated in the negative region. This indicates that larger diameters and wall thicknesses significantly enhance the predicted ultimate bearing capacity, while smaller dimensions tend to reduce the predicted values. These findings suggest that geometric dimensions are the dominant factors controlling the ultimate bearing capacity of CFST members. Overall, these observations are consistent with empirical engineering knowledge, identifying section size and material strength as decisive factors, whereas member length and size effects play a secondary role. Figure 6. Open in a new tab SHAP Summary Plot. 6.2. SHAP Feature Importance Plot Consistent with the results from the summary plot, Figure 7 indicates that D and t are the two most important features influencing model predictions, with the sum of their absolute SHAP values being significantly higher than that of other features. This reconfirms the decisive impact of steel pipe diameter and wall thickness on the ultimate bearing capacity. Figure 7. Open in a new tab SHAP Feature Importance Plot. 6.3. SHAP Dependence Plot To intuitively analyze the impact of feature value variations on prediction results and feature interactions, SHAP dependence plots for key features were generated, as shown in Figure 8 . The results indicate the following: (1) D and t : Their SHAP values generally show an upward trend as the feature values increase, indicating that larger diameters and thicker walls typically enhance the ultimate bearing capacity. (2) f y : This feature exhibits a non-linear relationship with SHAP values. The marginal influence is more sensitive in the medium strength range, while it plateaus at lower or higher ranges. This suggests that capacity gains in these extreme ranges may be limited by other failure modes or concrete/construction constraints. (3) f pr : There is an overall positive correlation, yet with fluctuations, reflecting that its effect is influenced by the confinement status and parameter combinations. (4) L : The influence of length is more complex. For small LL, SHAP values fluctuate near zero, showing minimal impact. However, once LL exceeds a certain threshold, SHAP values drop sharply into the negative range. This accurately captures the Euler Buckling phenomenon: as the slenderness ratio increases, the failure mode shifts from material failure to instability failure, and the P−δ effect significantly reduces the ultimate bearing capacity. Figure 8. Open in a new tab SHAP Dependence Plot. While traditional empirical formulas typically describe this behavior using a stability reduction factor ( φ ), the CatBoost model has automatically learned this physical law from the data. Overall, these dependence plots quantitatively reveal the non-linear mechanisms of geometric dimensions, material strength, and slenderness ratio on bearing capacity from a data-driven perspective, providing a reference for the rational selection of section and material parameters. These dependence plots quantitatively reveal the nonlinear mapping relationships between the feature parameters and the predicted bearing capacity, with the predicted trends showing a high degree of correlation with classical mechanical theories. In particular, the model successfully captures the Euler buckling and P-δ effects resulting from an increase in the slenderness ratio. This demonstrates that the data-driven model possesses robust physical consistency and engineering reliability. 6.4. SHAP Value Heatmap As shown in Figure 9 , features D and t exert the strongest influence on the prediction results, which is consistent with previous analyses. Furthermore, the heatmap reveals the interaction patterns among different features, further illustrating the combined effect of various feature combinations on the prediction outcome. Figure 9. Open in a new tab SHAP Value Heatmap. 7. Conclusions This study addresses the prediction of the ultimate bearing capacity of Concrete-Filled Steel Tubular (CFST) members. Various machine learning models were developed and compared, followed by a systematic interpretability analysis based on the optimal model. The main conclusions are drawn as follows: (1) Through the training and evaluation of multiple machine learning models, and utilizing random search combined with five-fold cross-validation for hyperparameter optimization, the R-CatBoost model achieved optimal performance on the test set. With an RMSE of 174.29, MAPE of 0.06, MAE of 107.30, and a coefficient of determination ( R 2 ) as high as 0.99, the model demonstrates the ability to predict the ultimate bearing capacity of CFST members with high precision while maintaining strong generalization capabilities on unseen data. Compared with traditional empirical formulas and other mainstream machine learning methods, R-CatBoost exhibits significant advantages in terms of both accuracy and robustness, serving as a reliable numerical tool for engineering design and assessment. (2) Global interpretation results based on the SHAP framework indicate that the D and t are the primary factors influencing the ultimate bearing capacity, followed by the f y and the f pr . The impacts of L and the γ U are relatively minor. This ranking of feature importance is highly consistent with existing theories and engineering experience: larger cross-sectional dimensions and thicker steel tube walls significantly enhance the bearing and confinement capacities of the member, while increasing steel and concrete strengths effectively raises the ultimate bearing capacity within a certain range. This consistency demonstrates that the R-CatBoost model not only possesses excellent numerical fitting performance but its internal decision logic also aligns with structural mechanical mechanisms, thereby enhancing the credibility of the model in engineering applications. (3) Systematic hyperparameter optimization plays a crucial role in enhancing model performance. By combining random search with five-fold cross-validation, this study effectively improved the generalization ability of multiple models. The improvement was particularly notable for the LightGBM model, which saw a 39.0% reduction in RMSE. This indicates that rational hyperparameter settings can fully unleash model potential while preventing overfitting, making it an indispensable step when applying machine learning models to practical engineering problems. Acknowledgments The authors are grateful for the constructive comments by the anonymous reviewers. Appendix A As detailed in Table A1 , the original sources for the 438 datasets (previously cited collectively in Ref. [ 26 ]) have been explicitly identified. The compiled database encompasses a wide range of geometric and material properties: the outer diameter ( D ) ranges from 48 mm to 361 mm, the wall thickness ( t ) from 1 mm to 11.88 mm, and the member length ( L ) from 153 mm to 1113 mm. Furthermore, the concrete grades span from C20 to C60. These parameters effectively cover the majority of standard CFST configurations commonly employed in structural engineering practice. Table A1. Source of the Original Dataset and Characteristics of Data Distribution. No. References D/mm t/mm L/mm Sample Size Percentage/% 1 Wang [ 33 ] 133.1~167.8 3.28~5.44 396~504 38 8.68 2 Lam D [ 34 ] 114~115 3.75~5.02 300 5 1.14 3 Ho J C M [ 35 ] 114~69.2 2.86~9.96 248~420 12 2.74 4 Schneider S P [ 36 ] 140 3~6.68 467~472 3 0.68 5 Miao [ 37 ] 100~273 1.45~8.75 309~1113 45 10.27 6 Tang [ 38 ] 92.0~210.0 1.5~4.0 276.0~630.0 16 3.65 7 Cai [ 39 ] 300/166 3/5 1000/660 4 0.91 8 Ding [ 40 ] 165~219 2.72~4.78 500~650 13 2.97 9 Gu [ 41 ] 152~166 4.5~10 480~500 6 1.37 10 Jiang [ 42 ] 166~202 3.0~4.5 500~600 12 2.74 11 Yao [ 43 ] 100~200 3 300~600 4 0.91 12 Zhao [ 44 ] 90 1.2~1.5 270 9 2.05 13 Tan [ 45 ] 133 4.7 465 4 0.91 14 Motoi [ 46 ] 101.6~139.8 2.37~2.99 305~419 18 4.11 15 He [ 47 ] 150~273 3.0~4.5 465~704 21 4.79 16 Chen [ 48 ] 100 1.6~2.5 300 8 1.83 17 Goode [ 49 ] 48~165 2.38~6.2 192~660 24 5.48 18 Sakino [ 50 ] 174~179 3.0~9.0 360 8 1.83 19 Luksha [ 51 ] 159~1020 5.07~13.25 477~3060 18 4.11 20 Sakino [ 52 ] 101.8 2.94~5.7 200 32 7.31 21 Huang [ 53 ] 159~164 3.8~6.3 520 12 2.74 22 Shea O [ 54 ] 165/190 1.94/2.82 580.5/663.5 6 1.37 24 Gardner N J [ 55 ] 76.5~152.6 1.7~4.09 153~305 8 1.83 25 Kato [ 56 ] 297.0~301.5 4.5~11.88 891~905 9 2.05 26 Yamamoto [ 57 ] 101.3~318.5 3.02~10.38 304~956 13 2.97 27 Weng [ 58 ] 200/280 5/4 600/840 4 0.91 28 Johansson [ 59 ] 159 5~10 650 3 0.68 29 Gu [ 60 ] 127~133 1.5~4.5 400 4 0.91 30 Yu [ 61 ] 100/200 3 300/600 16 3.65 31 Gupta [ 62 ] 89.3/112.6 2.89/2.74 340 16 3.65 32 Abed [ 63 ] 114~167 3.1~5.6 228~334 6 1.37 33 Liao [ 64 ] 180 3.8 720 2 0.46 34 Zhao [ 44 ] 140.0~216.3 4.5~8.2 420.0~648.9 24 5.48 35 Ye [ 65 ] 165 2.37 615 2 0.46 36 Ekmekyapar [ 66 ] 114.3 2.74/5.9 300 4 0.91 37 Chang [ 67 ] 111.64/113.64 1.9/3.64 400 6 1.37 38 Han [ 68 ] 120 2.65 360 2 0.46 39 Uenaka [ 69 ] 157.7 2.14 450 1 0.23 Total 438 100 Open in a new tab Author Contributions Conceptualization, Z.W. and Y.W. (Yunqiang Wang); methodology, X.X. and Y.W. (Yunqiang Wang); formal analysis, D.L.; software, Z.W. and Z.Z. investigation, D.L.; data curation, Z.W. and Y.W. (Yaxing Wei); writing—original draft preparation, Z.W.; writing—review and editing, Y.W. (Yunqiang Wang); supervision, Y.W. (Yaxing Wei). All authors have read and agreed to the published version of the manuscript. Data Availability Statement The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author. Conflicts of Interest The authors declare no conflicts of interest. Funding Statement This research received no external funding. Footnotes Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. References 1. Alatshan F., Osman S.A., Hamid R., Mashiri F. Stiffened Concrete-Filled Steel Tubes: A Systematic Review. Thin-Walled Struct. 2020;148:106590. doi: 10.1016/j.tws.2019.106590. [ DOI ] [ Google Scholar ] 2. Shaker F.M.F., Daif M.S., Deifalla A.F., Ayash N.M. Parametric Study on the Behavior of Steel Tube Columns with Infilled Concrete—An Analytical Study. Sustainability. 2022;14:14024. doi: 10.3390/su142114024. [ DOI ] [ Google Scholar ] 3. Feng Y., Wan X. Numerical simulation of dynamic response of RPC-FST columns under blast load. Prot. Eng. 2019;41:1–7. (In Chinese) [ Google Scholar ] 4. Na L., Yiyan L., Shan L., Lan L. Slenderness Effects on Concrete-Filled Steel Tube Columns Confined with CFRP. J. Constr. Steel Res. 2018;143:110–118. doi: 10.1016/j.jcsr.2017.12.014. [ DOI ] [ Google Scholar ] 5. Le K.B., Cao V.V., Cao H.X. Circular Concrete Filled Thin-Walled Steel Tubes under Pure Torsion: Experiments. Thin-Walled Struct. 2021;164:107874. doi: 10.1016/j.tws.2021.107874. [ DOI ] [ Google Scholar ] 6. Li H., Fang Y., Yang H., Wang Y. Behaviour of Concrete-Filled Double-Skin Steel Tubular Columns with Outer Galvanized Corrugated Steel Tubes under Monotonic and Cyclic Axial Compression. Structures. 2026;84:111083. doi: 10.1016/j.istruc.2026.111083. [ DOI ] [ Google Scholar ] 7. Li W., Zhu C.-Y., Chen H. Performance of Square High-Strength Steel Concrete-Filled Steel Tubes Subjected to Long-Term Sustained Loading: Axial Compression. J. Build. Eng. 2025;111:113362. doi: 10.1016/j.jobe.2025.113362. [ DOI ] [ Google Scholar ] 8. Wang T., Yu M., Zhang X., Xu L., Huang L. Experimental Study and Proposal of a Design Model of Ultra-High Performance Concrete Filled Steel Tube Columns Subjected to Fire. Eng. Struct. 2023;280:115697. doi: 10.1016/j.engstruct.2023.115697. [ DOI ] [ Google Scholar ] 9. Yu S., Huang S., Li Y., Liang Z. Insights into the Frost Cracking Mechanisms of Concrete by Using the Coupled Thermo-Hydro-Mechanical-Damage Meshless Method. Theor. Appl. Fract. Mech. 2025;136:104814. doi: 10.1016/j.tafmec.2024.104814. [ DOI ] [ Google Scholar ] 10. Yu S., Ren X., Zhang J. Modeling the Rock Frost Cracking Processes Using an Improved Ice-Stress-Damage Coupling Method. Theor. Appl. Fract. Mech. 2024;131:104421. doi: 10.1016/j.tafmec.2024.104421. [ DOI ] [ Google Scholar ] 11. Feng D.-C., Liu Z.-T., Wang X.-D., Chen Y., Chang J.-Q., Wei D.-F., Jiang Z.-M. Machine Learning-Based Compressive Strength Prediction for Concrete: An Adaptive Boosting Approach. Constr. Build. Mater. 2020;230:117000. doi: 10.1016/j.conbuildmat.2019.117000. [ DOI ] [ Google Scholar ] 12. Chen D., Fan Y., Zha X. Machine Learning-Based Strength Prediction of Round-Ended Concrete-Filled Steel Tube. Buildings. 2024;14:3244. doi: 10.3390/buildings14103244. [ DOI ] [ Google Scholar ] 13. Duong T.H., Le T.-T., Le M.V. Practical Machine Learning Application for Predicting Axial Capacity of Composite Concrete-Filled Steel Tube Columns Considering Effect of Cross-Sectional Shapes. Int. J. Steel Struct. 2023;23:263–278. doi: 10.1007/s13296-022-00693-0. [ DOI ] [ Google Scholar ] 14. Deng C., Xue X., Tao L. Prediction of Ultimate Bearing Capacity of Concrete Filled Steel Tube Stub Columns via Machine Learning. Soft Comput. 2024;28:5953–5967. doi: 10.1007/s00500-023-09343-x. [ DOI ] [ Google Scholar ] 15. Hai X., Hao Q., Hui Z.H. Comparative study on the application of machine learning algorithms in lithology identification. Prot. Eng. 2024;46:54–61. (In Chinese) [ Google Scholar ] 16. Tamimi M.F., Soliman M., Shojaeikhah S., Abu Qamar M.I., Alshannaq A.A. Reliability-Based Design of Concrete-Filled Stainless-Steel Tubular Columns Using Machine Learning. Structures. 2026;85:111193. doi: 10.1016/j.istruc.2026.111193. [ DOI ] [ Google Scholar ] 17. Lou J., Li Y., Zheng S., Feng Q., Wang G., Xu R., Qian X. Predicting the Ultimate Strength of Rectangular Concrete-Filled Steel Tube Columns under Eccentric Loading Using a Knowledge-Enhanced Machine Learning Framework. Eng. Appl. Artif. Intell. 2025;160:111935. doi: 10.1016/j.engappai.2025.111935. [ DOI ] [ Google Scholar ] 18. Wei B., Wei Y., Yi J., Chen J., Lin Y., Ding Y. Machine Learning-Driven Prediction of Axial Capacity of Concrete-Filled Steel Tubes Columns with Built-in Bamboo or Timber Cores. Eng. Struct. 2025;341:120817. doi: 10.1016/j.engstruct.2025.120817. [ DOI ] [ Google Scholar ] 19. Kazemi F., Asgarkhani N., Ghanbari-Ghazijahani T., Jankowski R. Ensemble Machine Learning Models for Estimating Mechanical Curves of Concrete-Timber-Filled Steel Tubes. Eng. Appl. Artif. Intell. 2025;156:111234. doi: 10.1016/j.engappai.2025.111234. [ DOI ] [ Google Scholar ] 20. Hafiz H., Li D., Luo P., Rasa A.R., Yang B. Interpretable K-Nearest Neighbors for Rockburst Risk Prediction: Robust Performance and Mechanistic Insights for Deep Rock Engineering. Eng. Fail. Anal. 2026;186:110443. doi: 10.1016/j.engfailanal.2025.110443. [ DOI ] [ Google Scholar ] 21. Liang D., Xue F. Integrating Automated Machine Learning and Interpretability Analysis in Architecture, Engineering and Construction Industry: A Case of Identifying Failure Modes of Reinforced Concrete Shear Walls. Comput. Ind. 2023;147:103883. doi: 10.1016/j.compind.2023.103883. [ DOI ] [ Google Scholar ] 22. Yang D., Yang J., Shi J. Interpretable Prediction and Simplified Calculation of Blast Load on Structure Surface Based on Machine Learning and Theoretical Model. Eng. Appl. Artif. Intell. 2026;167:113764. doi: 10.1016/j.engappai.2026.113764. [ DOI ] [ Google Scholar ] 23. Xinyu L., Fei L., Jian Y., Yong M., Hongtao L. Deep Learning-Based Intelligent Safety Protection Identification Method at Construction Sites. Prot. Eng. 2025;47:44–50. (In Chinese) [ Google Scholar ] 24. Kim D.K., Sung S.H., Song S.W., Kim S.J., Prabowo A.R., Kim S., Seo J.H., Ringsberg J.W. A SHAP Value Method for Ultimate Strength Prediction of Stiffened Panel: A Data-Driven Tool in Engineering. Ocean Eng. 2026;343:123159. doi: 10.1016/j.oceaneng.2025.123159. [ DOI ] [ Google Scholar ] 25. Alotaibi K.S., Alkhalaf A.M. Interpretable Machine Learning Framework for Predicting Cement Adhesive Bond Strength in NSM FRP Systems Using Differential Evolution and SHAP Analysis. Case Stud. Constr. Mater. 2026;24:e05761. doi: 10.1016/j.cscm.2026.e05761. [ DOI ] [ Google Scholar ] 26. Gen T. Steel Structures and Steel–Concrete Composite Structure Design Method. China Architecture & Building Press; Beijing, China: 2022. [ Google Scholar ] 27. Wang H., Qin B., Su Y., Li F., Hong S., Ding T. Coordinated Planning of Mobile Electric-hydrogen Energy Storage for Remote Power System Resilience Enhancement. J. Energy Storage. 2026;147:120160. doi: 10.1016/j.est.2025.120160. [ DOI ] [ Google Scholar ] 28. Qin B., Wang H., Liao Y., Li H., Ding T., Wang Z., Li F., Liu D. Challenges and Opportunities for Long-Distance Renewable Energy Transmission in China. Sustain. Energy Technol. Assess. 2024;69:103925. doi: 10.1016/j.seta.2024.103925. [ DOI ] [ Google Scholar ] 29. Li S.-Z., Wang J.-J., Jiang L., Deng R., Wang Y.-H. Machine Learning-Based Strength Prediction for Circular Concrete-Filled Double-Skin Steel Tubular Columns under Axial Compression. Eng. Struct. 2025;325:119460. doi: 10.1016/j.engstruct.2024.119460. [ DOI ] [ Google Scholar ] 30. Lyu F., Fan X., Ding F., Chen Z. Prediction of the Axial Compressive Strength of Circular Concrete-Filled Steel Tube Columns Using Sine Cosine Algorithm-Support Vector Regression. Compos. Struct. 2021;273:114282. doi: 10.1016/j.compstruct.2021.114282. [ DOI ] [ Google Scholar ] 31. Xie S., Isleem H.F., Almoghayer W.J.K., Ibrahim A.A. Axial Compressive Behavior of Reinforced Concrete-Filled Circular Steel Tubular Columns: Finite Element and Machine Learning Modelling. J. Big Data. 2025;12:233. doi: 10.1186/s40537-025-01282-8. [ DOI ] [ Google Scholar ] 32. Zhao Z., Wei Y., Wang G., Lin Y., Ding M. Evaluation of the Load-Carrying Effect of Rectangular Concrete-Filled Steel Tubular Columns under Axial Compression Based on the Multilevel Extension Method. Case Stud. Constr. Mater. 2023;18:e02041. doi: 10.1016/j.cscm.2023.e02041. [ DOI ] [ Google Scholar ] 33. Wang Y.Y. Ph.D. Thesis. Harbin Institute of Technology; Harbin, China: 2004. Study on the Basic Performance of Axially Compressed Short Columns of High-Strength Concrete-Filled Circular Steel Tubes. (In Chinese) [ Google Scholar ] 34. Giakoumelis G., Lam D. Axial capacity of circular concrete-filled tube columns. J. Constr. Steel Res. 2004;60:1049–1068. doi: 10.1016/j.jcsr.2003.10.001. [ DOI ] [ Google Scholar ] 35. Lai M.H., Ho J.C.M. A theoretical axial stress-strain model for circular concrete-filled-steel-tube columns. Eng. Struct. 2016;125:124–143. doi: 10.1016/j.engstruct.2016.06.048. [ DOI ] [ Google Scholar ] 36. Schneider S.P. Axially loaded concrete-filled steel tubes. J. Struct. Eng. 1998;124:1125–1138. doi: 10.1061/(ASCE)0733-9445(1998)124:10(1125). [ DOI ] [ Google Scholar ] 37. Miao R.Y. Application of plasticity theory to determine the bearing capacity of concrete-filled steel tube short columns under axial compression. J. Harbin Inst. Archit. Eng. 1982;2:36–39. [ Google Scholar ] 38. Tang G.Z., Zhao B.Q., Zhu H.X., Shen X.M. Study on the basic mechanical properties of concrete-filled steel tubes. J. Build. Struct. 1982;3:13–31. (In Chinese) [ Google Scholar ] 39. Cai S.H., Jiao Z.S. Basic Behavior and Strength Calculation of Concrete-Filled Steel Tubular Short Columns. J. Build. Struct. 1984;6:13–29. (In Chinese) [ Google Scholar ] 40. Ding F.X. Ph.D. Thesis. Central South University; Changsha, China: 2006. Research on the Structural Performance and Design Method of Circular Concrete-Filled Steel Tubes. (In Chinese) [ Google Scholar ] 41. Gu W.P., Cai S.H. Research on the behavior and bearing capacity of eccentrically loaded high-strength concrete-filled steel tubular columns. Build. Sci. 1993;9:8–12. (In Chinese) [ Google Scholar ] 42. Jiang J.W. Ph.D. Thesis. Tsinghua University; Beijing, China: 1997. Experimental Study on the Seismic Performance of High-Strength Concrete-Filled Steel Tubular Beam-Columns Under Cyclic Loading. (In Chinese) [ Google Scholar ] 43. Yao G.H., Han L.H. Mechanical behavior of concrete-filled steel tubular beam-columns with self-compacting high-performance concrete. J. Build. Struct. 2004;25:34–42. (In Chinese) [ Google Scholar ] 44. Miao S.F., Zhao J.H., Gu Q. Study on the bearing capacity of axially loaded concrete-filled steel tubes based on the twin-shear unified strength theory. Eng. Mech. 2002;19:32–35. (In Chinese) [ Google Scholar ] 45. Tan K.F. Ph.D. Thesis. Chongqing Jianzhu University; Chongqing, China: 1999. Study on the Mechanical Properties and Bearing Capacity of Steel Tube and Ultra-High-Strength Concrete Composites. (In Chinese) [ Google Scholar ] 46. Motoi M., Abe T., Nakaya H. Study on the ultimate bearing capacity of ultra-high-strength concrete-filled steel tubular columns. J. Struct. Constr. Eng. AIJ. 1999;523:133–140. (In Japanese) [ Google Scholar ] 47. He F., Zhou X.H. Experimental study on the bearing behavior of axially compressed high-strength concrete-filled steel tubular stub columns. Eng. Mech. 2000;17:61–66. (In Chinese) [ Google Scholar ] 48. Chen Z.Y. Performance of Reinforced Concrete Structures under Impact Loads, Scientific Report Collection No. 4. Tsinghua University Press; Beijing, China: 1986. Performance of Concrete-Filled Steel Tubular Stub Columns as Protective Structural Members; pp. 45–52. (In Chinese) [ Google Scholar ] 49. Goode C.D. Composite columns—1819 tests on concrete-filled steel tube columns compared with Eurocode 4. Struct. Eng. 2008;86:33–38. [ Google Scholar ] 50. Sakino K., Hayashi H. Behavior of concrete filled steel tubular stub columns under concentric loading; Proceedings of the 3rd International Conference on Steel-Concrete Composite Structures; Fukuoka, Japan. 26–29 September 1991; pp. 25–30. [ Google Scholar ] 51. Luksha L.K., Nesterovich A.P. Strength test of large-diameter concrete on steel-concrete composite structures; Proceedings of the 3rd International Conference on Steel-Concrete Composite Structures; Fukuoka, Japan. 26–29 September 1991; pp. 67–70. [ Google Scholar ] 52. Sakino K., Nakahara H., Morino S., Nishiyama I. Behavior of centrally loaded concrete-filled steel tube short columns. J. Struct. Eng. 2004;130:180–188. doi: 10.1061/(ASCE)0733-9445(2004)130:2(180). [ DOI ] [ Google Scholar ] 53. Huang M.K., Li B., Wen Y. Analysis of the influence of confinement effect coefficient on the mechanical properties of concrete-filled steel tubular members. J. Chongqing Jianzhu Univ. 2008;30:90–93. (In Chinese) [ Google Scholar ] 54. Martin D., Shea O., Russell Q. Bridge, design of circular thin-walled concrete filled steel tubes. J. Struct. Eng. 2000;126:1295–1303. [ Google Scholar ] 55. Gardner N.J., Jacobson E.R. Structural behavior of concrete filled steel tubes. ACI J. 1967;64:404–412. [ Google Scholar ] 56. Kato B. Compressive strength and deformation capacity of concrete-filled tubular stub columns. J. Struct. Constr. Eng. 1995;468:183–191. doi: 10.3130/aijs.60.183. [ DOI ] [ Google Scholar ] 57. Yamamoto K., Kawaguchi J., Morino S. Experimental study of the size effect on the behaviour of concrete filled circular steel tube columns under axial compression. J. Struct. Constr. Eng. 2002;561:237–244. doi: 10.3130/aijs.67.237_2. [ DOI ] [ Google Scholar ] 58. Huang C.S., Yeh Y.-K., Liu G.Y., Hu H.-T., Tsai K.C., Weng Y.T., Wang S.H., Wu M.-H. Axial load behavior of stiffened concrete-filled steel columns. J. Struct. Eng. 2002;128:1222–1230. doi: 10.1061/(ASCE)0733-9445(2002)128:9(1222). [ DOI ] [ Google Scholar ] 59. Johansson M. The efficiency of passive confinement in CFT columns. Steel Compos. Struct. 2002;2:379–396. doi: 10.12989/scs.2002.2.5.379. [ DOI ] [ Google Scholar ] 60. Gu W., Guan C.W., Zhao Y.H., Cao H. Experimental study on axially compressed circular CFRP-steel composite tubed concrete stub columns. J. Shenyang Jianzhu Univ. 2004;20:118–120. (In Chinese) [ Google Scholar ] 61. Yu Q.G., Han L.H. Experimental behaviour of thin-walled hollow structural steel (HSS) columns filled with self-consolidating concrete (SCC) Thin-Walled Struct. 2004;42:1357–1377. [ Google Scholar ] 62. Gupta P.K., Sarda S.M., Kumar M.S. Experimental and computational study of concrete filled steel tubular columns under axial loads. J. Constr. Steel Res. 2007;63:182–193. doi: 10.1016/j.jcsr.2006.04.004. [ DOI ] [ Google Scholar ] 63. Abed F., AlHamaydeh M., Abdalla S. Experimental and numerical investigations of the compressive behavior of concrete filled steel tubes (CFSTs) J. Constr. Steel Res. 2013;80:429–439. doi: 10.1016/j.jcsr.2012.10.005. [ DOI ] [ Google Scholar ] 64. Liao F.Y., Han L.H., He S. Behavior of CFST short column and beam with initial concrete imperfection: Experiments. J. Constr. Steel Res. 2011;67:1922–1935. doi: 10.1016/j.jcsr.2011.06.009. [ DOI ] [ Google Scholar ] 65. Ye Y., Han L.H., Sheehan T., Guo Z.-X. Concrete-filled bimetallic tubes under axial compression: Experimental investigation. Thin-Walled Struct. 2016;108:321–332. doi: 10.1016/j.tws.2016.09.004. [ DOI ] [ Google Scholar ] 66. Ekmekyapar T., Al-Eliwi B.J.M. Experimental behavior of circular concrete filled steel tube columns and design specifications. Thin-Walled Struct. 2016;105:220–230. doi: 10.1016/j.tws.2016.04.004. [ DOI ] [ Google Scholar ] 67. Chang X., Fu L., Zhao H.B., Zhang Y.-B. Behaviors of axially loaded circular concrete-filled steel tube (CFT) stub columns with notch in steel tubes. Thin-Walled Struct. 2013;73:273–280. doi: 10.1016/j.tws.2013.08.018. [ DOI ] [ Google Scholar ] 68. Han L.H., Yao G.H. Influence of initial stress in steel tube on the bearing capacity of concrete-filled steel tubular beam-columns. China Civ. Eng. J. 2003;36:9–18. (In Chinese) [ Google Scholar ] 69. Uenaka K., Kitoh H., Sonoda K. Concrete filled double skin circular stub columns under compression. Thin-Walled Struct. 2010;48:19–24. doi: 10.1016/j.tws.2009.08.001. [ DOI ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The original contributions presented in this study are included in the article. Further inquiries can be directed to the corresponding author. Articles from Materials are provided here courtesy of Multidisciplinary Digital Publishing Institute (MDPI) ACTIONS View on publisher site PDF (2.4 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top