硅酸盐学报

掺入钢纤维虽能显著提升素混凝土的强度与韧性，但也导致其力学响应更为复杂，随机性和不可预测性增强。为深入探究钢纤维混凝土的抗压与劈拉强度，本工作构建了一个包含636组试验样本的数据库，特别涵盖了粗骨料尺寸、纤维形状、几何特征及含量等关键变量。基于此数据库，随即开发并对比了5种人工智能预测模型：随机森林(RF)、梯度提升回归树(GBRT)、极端梯度提升(XGB)、轻量级梯度提升机(LGBM)和贝叶斯神经网络(BNN)。在有效预防过拟合与欠拟合的基础上，详细评估了各模型的预测精度与鲁棒性，并进一步结合沙普利加性解释(SHAP)值分析、部分依赖图(PDP)和个体条件期望(ICE)等解释技术揭示模型的内部机制。结果显示：模型在测试集上的决定系数(R²)介于0.84～0.90之间，训练集R²介于0.80～0.90，差距较小，均高于0.75，展现出良好的预测能力与泛化性能，其中以LGBM模型表现最佳。变量贡献分析表明，影响抗压强度的关键因素依次为水灰比、砂率、纤维增强系数、骨灰比与粗骨料最大粒径；而劈拉强度主要受纤维增强系数、砂率和水灰比控制。

Introduction Conventional concrete suffers from its inherently low tensile strength. To address this drawback, researchers propose steel fiber reinforced concrete(SFRC) as a novel material that significantly improves tensile properties and exhibits outstanding performance in cracking resistance, toughness, and durability. However, the incorporation of steel fibers makes the mechanical behavior of SFRC more complex. To ensure its safe application in structural engineering, an accurate prediction of its mechanical properties is essential. Artificial intelligence technologies are widely applied to accurately evaluate the mechanical properties of concrete. However, the existing machine learning models for SFRC are typically trained on a limited number of samples(i.e., fewer than 300), which restricts their generalization capability. Moreover, these models often lack interpretability, undermining their credibility and limiting their practical application in engineering projects. This study was to establish a database for the compressive and splitting tensile strengths of SFRC, develop corresponding machine learning prediction models, and evaluate their predictive performance. Furthermore, the interpretability of the models was analyzed. The findings of this study could offer some insights into the development of high-precision machine learning models for SFRC and their interpretability analysis, thereby promoting their application in engineering practices. Methods In this study, a total of 636 experimental data points on SFRC were collected from 24 independent studies to establish a comprehensive database, including 419 data points for compressive strength and 217 for splitting tensile strength. Each data entry contains a mix-related information such as the water-to-binder ratio, sand ratio, aggregate-to-binder ratio, maximum size of coarse aggregate, fiber shape factor, fiber aspect ratio, fiber volume fraction, and the corresponding compressive or splitting tensile strength. Based on this database, five machine learning models(i.e., Random Forest(RF), Gradient Boosting Regression Tree(GBRT), Extreme Gradient Boosting(XGB), Light Gradient Boosting Machine(LGBM) and Bayesian Neural Network(BNN)) were proposed. The predictive performance of these models was evaluated using the coefficient of determination(R²), mean absolute percentage error(MAPE), and root mean square error(RMSE), in order to identify the most effective prediction model. For the selected model, the SHAP(SHapley Additive exPlanations) method was employed to analyze the trends and contribution of each input parameter on the compressive and splitting tensile strengths of SFRC. In addition, the Partial Dependence Plot(PDP) and Individual Conditional Expectation(ICE) methods were also used to quantitatively investigate the variation patterns of predicted values with respect to individual input parameters. Results and discussion All the models proposed demonstrate a great predictive performance. The R² for the test set ranges from 0.84 to 0.90, while for the training set it ranges from 0.80 to 0.90. A small difference between them(i.e., both being > 0.75) indicates a good predictive accuracy and a generalization ability.For the overall performance across R², MAPE and RMSE metrics, the LGBM model shows the optimum prediction performance. The SHAP analysis reveals that the compressive and splitting tensile strengths of SFRC decrease with increasing water-to-binder ratio and aggregate-to-binder ratio, while the strengths both increase at a higher sand ratio and a greater fiber reinforcement factor. In addition, as the maximum size of coarse aggregate increases, the compressive strength decreases, whereas the splitting tensile strength increaseds. These results are consistent with the practical observations. Based on the mean absolute SHAP values, the input parameters influencing the compressive strength in a descending order are water-to-cement ratio, sand ratio, fiber reinforcement factor, aggregate-to-cement ratio, and maximum coarse aggregate size. For splitting tensile strength, the order is fiber reinforcement factor, sand ratio, water-to-binder ratio, aggregate-to-binder ratio, and maximum coarse aggregate size. The influence and parameter importance rankings obtained through the PDP and ICE analysis are in a reasonable agreement with those derived from SHAP, further validating the interpretability and reliability of the model. Conclusion A large-scale database containing 636 data entries was established for the compressive and splitting tensile strengths of SFRC. Based on this database, five machine learning prediction models(i.e., RF, GBRT, XGB, LGBM and BNN) were proposed. All the models exhibited good predictive performance and generalization ability, significantly outperforming the existing prediction methods. Among them, the LGBM model showed the optimum overall performance. The SHAP, PDP and ICE techniques were employed to analyze the influence and importance of each input parameter on the compressive and splitting tensile strengths. Based on the variable contribution analysis, the main factors affecting the compressive strength were the water-to-cement ratio, sand ratio, fiber reinforcement factor, aggregate-to-cement ratio, and maximum coarse aggregate size. In contrast, the key determinants of splitting tensile strength were the fiber reinforcement coefficient, sand ratio and water-to-cement ratio.