Comparison of Predictive Modeling Concrete Compressive Strength with Machine Learning Approaches

Abstract

Accurately predicting concrete compressive strength is fundamental for optimizing mix designs, ensuring structural integrity, and advancing sustainable construction practices. Increased demands for safer, more durable infrastructure necessitate effective predictive concrete compressive strength models. This research aims to compare the effectiveness of six machine learning models such as Linear Regression, Random Forest, Support Vector Regression (SVR), K-Nearest Neighbors (KNN), Gradient Boosting, and XGBoost to predict concrete compressive strength. Used a dataset of 1030 instances with varying mixture compositions, conducted extensive exploratory data analysis, and applied feature engineering and data scaling to enhance model performance. Assessments were performed with a 5-fold cross-validation approach with the R-squared (R²) metric. In addition, the SHAP value is used to understand the influence of each feature on the compressive strength results. The results revealed that XGBoost significantly outperformed other models, achieving an average R² value of 0.9178 with a standard deviation of 0.0296. Notably, Random Forest and Gradient Boosting also demonstrated robust capabilities. Based on our experiment, these models effectively predicted compressive strengths close to actual measured values, confirming their practical applicability in civil engineering. SHAP values provided insights into the significant impact of age and cement quantity on model outputs. These results highlight the advanced ensemble methods' effectiveness in concrete compressive strength prediction and underscore the importance of feature engineering in enhancing model accuracy.

Keywords

• concrete compressive strength
• machine learning
• prediction
• shap values