Data-Driven Prediction of Concrete Behaviour and Performance Enhancement Through CO₂ Inclusion Techniques

Abstract

Concrete remains the most widely used construction material globally, valued for its versatility, strength, and durability. However, the conventional approach to concrete mix design, governed by empirical codal provisions such as IS 10262:2019, is inherently iterative, labor-intensive, and prone to inefficiencies. It relies heavily on trial-and-error experimentation to achieve desired performance metrics such as slump and compressive strength. Moreover, the process lacks a feedback mechanism to incorporate historical data, resulting in the loss of experiential knowledge, increased material and labor costs. Simultaneously, the environmental impact of concrete production, particularly the carbon emissions associated with cement manufacturing, poses a significant challenge to sustainable development. This thesis addresses these dual concerns by presenting a novel machine learning-based modeling approach to predict compressive strength, slump, and carbonation depth of concrete subjected to accelerated and natural carbonation, with experimental CO₂ sequestration techniques to enhance both the technical and environmental performance of concrete. This thesis presents an innovative machine learning based framework for predicting key concrete properties—slump, compressive strength at 7 and 28 days, and carbonation depth—using a dataset of 963 mix designs with 31 input features. Seven ML algorithms were evaluated: K-Nearest Neighbors (KNN), Support Vector Machine (SVM), Decision Tree (DT), Random Forest Model (RFM), Gradient Boosting (GB), Linear Regression (LR), and Back-Propagation Neural Network (BPNN). Random Forest Model delivered the highest accuracy, with R² values of 0.77 (slump), 0.84 (7-day strength), and 0.83 (28-day strength). Further optimization of RFM using Bayesian Optimization, Grid Search, and Random Search significantly improved performance, with the Bayesian-optimized RFM achieving an R² of 0.916 for 28-day strength. SHAP (SHapley Additive exPlanations) analysis enabled feature importance ranking and the development of reduced-input models, maintaining high accuracy with fewer parameters and improved computational efficiency. In a pioneering step toward performance-based concrete mix design, the Bayesian optimized Random Forest model has been employed to predict carbonation depth across both naturally and artificially carbonated environments—a capability not previously demonstrated. Leveraging three diverse datasets spanning CO₂ concentrations from ambient levels to 100%, the model achieved a high generalization performance (R² = 0.92), marking a significant advancement in durability prediction. This cross-environmental applicability suggests a paradigm shift in how carbonation behavior can be modelled. SHAP analysis further revealed key drivers of carbonation, laying the groundwork for simplified yet robust lifecycle assessment tools. In a parallel line of investigation, the thesis explores the potential of integrating CO₂ directly into cement paste during mixing as a means of enhancing mechanical performance and promoting carbon sequestration. Two experimental techniques were studied: the inclusion of dry ice (solid CO₂) and the bubbling of gaseous CO₂ into the mix. Their effects on initial setting time, compressive strength, and microstructural development were analyzed using TGA, XRD, FTIR, and SEM. Dry ice reduced setting time by up to 100 minutes, acting as an accelerator, while gaseous CO₂ increased it by 20 minutes, functioning as a retarder. Both treatments improved compressive strength: 4% dry ice inclusion yielded 57.75 MPa (10.5% increase), and gaseous CO₂ treatment achieved 58.86 MPa (12.6% increase). TGA and XRD confirmed increased CaCO₃ formation, with the D12 specimen showing a 40.54% rise in CaCO₃ and 3% CO₂ capture; gaseous CO₂ treatment resulted in a 32.43% CaCO₃ increase and 2.4% CO₂ capture. The findings demonstrate that ML models can effectively verify codal mix proportions and reduce experimental trials, while CO₂ integration during mixing offers a promising route for performance enhancement and decarbonization. The research also opens avenues for reverse mix design, deriving optimal mix proportions from desired properties, and for extending data-driven modelling to novel materials like geopolymer concrete and recycled aggregates. In conclusion, this thesis presents a comprehensive framework that combines advanced computational intelligence and experimental innovation to modernize concrete mix design. It contributes to sustainable construction practices by improving efficiency and reducing environmental impact, aligning with the United Nations Sustainable Development Goals.