<![CDATA[Minimising embodied CO2 in concrete is one of the most important tasks that needs to be performed in order to attain a sustain‑able construction and prevent the effects of the changes in climatic conditions. This paper is a comparative analysis of three machine learning models: Linear Regression (LR), AdaBoost (ADB), and K‑Nearest Neighbours (KNN) to predict embodied_CO2 based on a dataset of 1,000 ob‑servations in the form of mixture composition, material properties, and environmental indicators. The descriptive statistical analysis assured the balanced distribution of most variables with little skewness, whereas the correlation analysis revealed cement and resource consumption as the leading factors contributing to embodied_CO2. Training, testing, split, and k‑fold cross‑validation based on the R, MAE, RMSE, RAE, and RRSE metrics were used to measure the model performance. Findings reveal that KNN was a better method in comparison with LR and ADB in all assessment systems. KNN with k‑fold validation had a correlation coefficient of 0.9996, MAE of 1.8668, and RMSE of 2.5041 versus LR (R = 0.9874, MAE = 11.3218, RMSE = 13.0931) and ADB (R = 0.9764, MAE = 14.5647, RMSE = 18.0974). The same tendencies were noted in the testing stage, with KNN having R = 0.9996, MAE = 1.9273, and RMSE = 2.7044, which are considerably lower than LR (MAE = 11.0947; RMSE = 12.8293) and ADB (MAE = 13.9921; RMSE = 16.8487). The residual analysis also indicated that KNN has better stability, with tightly clustered and symmetric error distributions and a small generalisation gap. The results show that instance‑based learning is effective to learn complex nonlinear associations in embodied carbon prediction. This paper emphasizes the significance of strong cross‑validation and residual diagnos‑tics in model selection and shows the feasibility of machine learning in aiding the design of low‑carbon concrete with regard to design strategies.]]>
