工程科学与技术2025，Vol.57Issue(5)：1-13,13.DOI:10.12454/j.jsuese.202401032
基于BO‒XGBoost算法的气凝胶混凝土力学和保温性能预测方法
Prediction Method for the Mechanical and Thermal Insulation Properties of Aerogel Concrete Based on the BO‒XGBoost Algorithm
摘要
Abstract
Objective Accurately predicting the compressive strength and thermal conductivity of aerogel concrete with different mix ratios is essential.Con-ventional performance prediction methods involve repeatedly designing mix ratios and measuring performance in the laboratory,which requires substantial manpower and time.Alternatively,formula-based or statistical methods are employed to obtain optimal performance;however,the em-pirical formulas derived from these methods demonstrate low accuracy,remain highly dependent on specific experiments,and cannot be general-ized or applied to other cases.This study proposes an aerogel concrete performance prediction model based on the Extreme Gradient Boosting(XGBoost)algorithm,which enables a clearer understanding of the nonlinear relationships between the components and the performance of aero-gel concrete. Methods First,relevant data were collected from research literature to construct a database containing 183 sets of compressive strength data and 152 sets of thermal conductivity data.The input features primarily consisted of mix ratios of materials such as cement,while the output features represented the performance of aerogel concrete,specifically compressive strength and thermal conductivity.Second,the XGBoost algorithm was employed for model training and performance prediction,and evaluation metrics such as the coefficient of determination(R2)and root mean square error(RMSE)were utilized to assess the model's accuracy.The Bayesian optimization algorithm was applied to determine the optimal hy-perparameters of the XGBoost model to enhance prediction reliability and reduce overfitting.Third,since the model contained two types of input methods,one approach used the amounts of materials such as aerogel,cement,silica fume,sand,and water as input variables,while the other used interpretable features such as the water-to-binder ratio,aggregate ratio,aerogel content,silica fume content,and curing age as input vari-ables.The model also generated two output results,compressive strength and thermal conductivity.Three comparison schemes were designed,and their prediction accuracy and performance were assessed using the same database to compare the two input methods and to evaluate the strate-gies of"building one model to predict both performances"versus"building two models to predict the performances separately".Fourth,to dem-onstrate the advantages of the BO‒XGBoost algorithm in predicting aerogel concrete performance,comparisons were conducted with multiple classical machine learning algorithms,including Random Forest(RF)and Artificial Neural Networks(ANN).For objectivity,all algorithms used the same unified database and underwent Bayesian optimization for hyperparameter tuning.Fifth,to investigate the generalization capability of the BO‒XGBoost model,it was applied to predict the performance of 12 new sets of aerogel concrete data,and the prediction accuracy of the model was observed.Finally,since machine learning models were inherently"black boxes,"making it challenging to characterize complex non-linear relationships between input and output variables,the Shapley Additive Explanations(SHAP)model was employed for feature interpretabil-ity analysis,calculating the contribution of each feature to the model's predictions and explaining the relationships between input features and out-put results. Results and Discussions 1)The analysis of the prediction accuracy of models established under the three schemes revealed that all three schemes achieved high accuracy in predicting the performance of aerogel concrete,with R2 values greater than 0.92 for the test set.The model under Scheme 1 yielded the best results,with R2 values of 0.977 and 0.978 for compressive strength and thermal conductivity,respectively,and RMSE values of 2.366 MPa and 0.104 0.128 W/(m·K).2)When compared to four other traditional machine learning algorithms,the XGBoost model was shown to be more suitable for predicting the performance of aerogel concrete.The compressive strength and thermal conductivity models based on the XGBoost algorithm achieved R2 values of 1.000 and 0.992 for the training set,and R2 values of 0.977 and 0.978 for the test set.In contrast,the R2 values of the test set for RF,ANN,and other classical algorithms were all below 0.963,indicating poor fitting accuracy and signifi-cant overfitting,which rendered them unsuitable for predicting aerogel concrete performance.3)The generalization ability of the model was veri-fied using 12 new sets of aerogel concrete mix ratio data.The R2 values for compressive strength and thermal conductivity prediction were 0.986 and 0.895,respectively,with RMSE values of 1.539 MPa and 0.128 W/(m·K),demonstrating the model's strong generalization capability.4)The SHAP model analysis revealed that the primary factors influencing the compressive strength of aerogel concrete are aerogel content and water-to-binder ratio.As the aerogel content and water-to-binder ratio increase,the compressive strength decreases.The primary factors influencing the thermal conductivity of aerogel concrete are also aerogel content and water-to-binder ratio,with higher values leading to reduced thermal conduc-tivity.The results obtained from SHAP analysis are consistent with conclusions drawn from laboratory experiments. Conclusions This study proposes an XGBoost-based performance prediction model for aerogel concrete,designed to predict both the compres-sive strength and the thermal conductivity of ordinary aerogel concrete.Compared to conventional empirical fitting formulas,the prediction model exhibited higher accuracy and stronger generalization capacity,presenting a novel approach for predicting performance and designing mix ratios of aerogel concrete.关键词
气凝胶混凝土/性能预测/贝叶斯优化/XGBoost算法/SHAP可解释分析Key words
aerogel concrete/performance prediction/bayesian optimization/XGBoost algorithm/SHAP interpretability analysis分类
建筑与水利引用本文复制引用
熊峰,陈腾盛,邓楚兵,李云飞,曾一..基于BO‒XGBoost算法的气凝胶混凝土力学和保温性能预测方法[J].工程科学与技术,2025,57(5):1-13,13.基金项目
国家自然科学基金项目(52278197) （52278197）
四川省博士后创新人才支持项目(BX202314) （BX202314）
