机器学习预测混凝土抗压强度梯度提升决定系数达0.940：《科学报告》以1030条观测比较四种模型
《科学报告》11月13日刊出一项混凝土抗压强度预测研究。研究使用UCI混凝土抗压强度数据集，共1030条观测、428种配合比，比较了线性回归、支持向量回归、随机森林与梯度提升四种模型。
研究采用按强度分层的八二划分，得到824条训练观测与206条留出观测，超参数在训练部分内以五折交叉验证选取，留出集只使用一次。
结果显示，梯度提升表现最好，决定系数R²为0.940，均方根误差为4.07兆帕，平均绝对误差为2.70兆帕；随机森林R²为0.926，支持向量回归R²为0.887。
研究同时指出，多数配合比在多个龄期受测，随机划分衡量的是对已知配合比新观测的泛化；改用按配合比分组的交叉验证后，梯度提升得分由0.933降至0.862，但模型排名不变。
SHAP归因将龄期、水泥、水与矿渣列为领先预测变量。作者表示，该框架支持配合比的初步筛选并配合实验室验证，外部验证仍属必要。该研究未获外部资助，作者来自科威特中东美国大学工程学院等机构。
（来源：《科学报告》）
Gradient Boosting Reaches R² of 0.940 Predicting Concrete Strength in Scientific Reports Study of 1,030 Observations
Scientific Reports published a study on concrete compressive strength prediction on November 13, 2026. The study used the UCI Concrete Compressive Strength dataset, comprising 1,030 observations and 428 mixtures, and compared Linear Regression, Support Vector Regression, Random Forest and Gradient Boosting. A strength-stratified 80/20 split gave 824 training and 206 held-out observations; hyperparameters were selected by five-fold cross-validation within the training partition and the held-out set was used once. Gradient Boosting performed best, with R² of 0.940, RMSE of 4.07 MPa and MAE of 2.70 MPa, followed by Random Forest at R² of 0.926 and Support Vector Regression at R² of 0.887. Because most mixtures are tested at several ages, the authors also ran mixture-grouped cross-validation, which lowered the Gradient Boosting score from 0.933 to 0.862 with the ranking preserved. SHAP attributions ranked age, cement, water and slag as the leading predictors. The authors said the framework supports preliminary mixture screening with laboratory verification, and that external validation remains necessary. The research received no external funding.
(Source: Scientific)
来源：《科学报告》
发布：GSM·美国（发布链接）
图片来源：图：一个受压试验后开裂的混凝土圆柱试件。Xb-70 at English Wikipedia，公共领域，经 Wikimedia Commons。 / Photo: Xb-70 at English Wikipedia, public domain, via Wikimedia Commons.
GSM Global · 本篇英文版 English edition of this story
