What question did this study set out to answer?

The aim is to create a machine learning framework to better predict the compressive strength of CO₂-cured concrete.

February 22, 2026Open Access

Machine Learning-Based Prediction of the Compressive Strength of CO₂-Cured Concrete: Benchmarking and SHAP Interpretability

FHFiras HilalogluKing Abdullah University of Science and Technology QKQudsia KanwalKing Abdullah University of Science and Technology MAMohammed A. AlmarshoudUniversity of Jeddah

Key Points

The aim is to create a machine learning framework to better predict the compressive strength of CO₂-cured concrete.
Developed a framework using eight experimentally traceable inputs from 280 concrete mixes.
Tested six supervised machine learning models including random forest, gradient boosting trees, and neural networks.
Conducted Shapley additive explanation analysis for interpretability of model results.
Ensemble models achieved a test coefficient of determination (R²) ranging from 0.950 to 0.963.
Root mean square error between 3.0 and 3.7 MPa indicates strong predictive performance.
Identified cement content and water-to-binder ratio as primary factors affecting compressive strength.

Abstract

The construction sector is a major source of global CO₂ emissions, with cement production a primary contributor. Accelerated carbonation curing (ACC) represents a viable low-carbon strategy to address this; however, predicting the compressive strength of CO₂-cured concrete remains challenging because this strength depends on the combination of the mix design and carbonation parameters, which tend to interact nonlinearly. Existing empirical and machine learning (ML)-based studies have largely focused on conventional curing systems or single-model predictions and lack mechanistic interpretability for carbonation systems. This study thus develops a comprehensive and interpretable ML-based framework specifically tailored to CO₂-cured concrete using eight experimentally traceable inputs from 280 laboratory mixes and testing six supervised models: multiple linear regression, support vector regression, random forest (RF), gradient boosting trees (GBT), extreme gradient boosting (XGBoost), and an artificial neural network (ANN). The ensemble models RF, GBT, and XGBoost achieved the highest accuracy, with a test coefficient of determination (R²) of 0.950–0.963 and a root mean square error of 3.0–3.7 MPa, indicating strong generalization. Importantly, the R² values achieved for these ensemble models exceeded those reported in previous studies on CO₂-cured concrete and related systems by up to 0.065, highlighting the enhanced reliability and novelty of the proposed framework. Multi-level Shapley additive explanation analysis was also conducted for the mechanism-consistent interpretation of the modeling results, revealing that the cement content, water-to-binder ratio, curing duration, and CO₂ pressure had the strongest positive effect on the compressive strength, with CO₂ concentration playing a secondary role. The proposed framework provides an accurate, transparent, and scalable tool for optimizing ACC parameters, reducing experimental costs, and promoting the data-driven deployment of low-carbon concrete technologies.

اسأل الذكاء الاصطناعي

Bookmark

View Full Paper