With increasing global emphasis on sustainable development, the CO₂ emissions from concrete as the most widely used building material have attracted much attention. Statistics show that cement production accounts for approximately 8% of global CO₂ emissions. Low-carbon concrete minimizes lifecycle CO₂ emissions through optimized cementitious materials, use of solid waste, and low-carbon technique (e.g., carbon capture), while still meeting engineering performance standards. However, incorporating recycled aggregates or mineral admixtures, although reducing CO₂ emissions, often leads to a deterioration in mechanical properties and durability of concrete. Therefore, how to effectively balance reduced CO₂ emissions while maintaining the engineering performance of concrete remains a challenge in low-carbon concrete mix designs.

The complex interactions among multiple components of concrete materials make mix proportion design highly complex. Conventional trial-and-error methods show obvious deficiencies in efficiency, cost and precision. Previous research established some performance prediction models based on machine learning methods, which often focused on single performance indicators. There is little systematic research on multi-objective collaborative optimization of mechanical properties. This failure to adequately balance environmental benefits with performance hinders the effective use of AI in designing low-carbon concrete. The result demonstrates that optimizing the proportions of recycled aggregates and supplementary cementitious materials can effectively limit carbon emissions to the range of 240-260 kg CO₂/m³ when the mechanical property requirements (i.e., compressive strength >50 MPa) and chloride ion permeability resistance (i.e., electric flux <500 C) are satisfied. This study was to offer an approach for achieving a balanced design between material performance and CO₂ emissions in low-carbon concrete, and to promote the application of Artificial Intelligence (AI) technology in its mix proportion design.

This study introduced an AI-based intelligent design optimization method for low-carbon concrete. The process started with data preprocessing, including normalization and outlier treatment, to enhance dataset quality. Subsequently, various models, optimized via different hyperparameter tuning methods, were benchmarked for high-precision prediction of key performance indicators. The SHapley Additive Explanations (SHAP) analysis was employed to elucidate the impact of mix proportion parameters on concrete performance. The The Non-dominated Sorting Genetic Algorithm III (NSGA-III) algorithm was then utilized for multi-objective optimization of CO₂ emissions, cost, strength, and chloride ion penetration resistance. Two indices, C_i (Carbon intensity index) and Q_i (Chloride ion penetration carbon efficiency index), were introduced to quantify the balance between environmental impact and concrete performance. These indices were subsequently employed to evaluate the solutions from the multi-objective optimization, guiding the decision-making process towards a comprehensive and synergistic optimization of material properties and low-carbon characteristics.

During data preprocessing, normalization addresses dimensional inconsistencies among feature variables, enhancing their comparability. Furthermore, the Density-Based Spatial Clustering of Applications with Noise (DBSCAN) algorithm effectively identifies specific outliers (i.e., 43 related to compressive strength and 4 concerning electric charge passed), leading to a marked improvement in overall data quality. Comparison of machine learning models shows that eXtreme Gradient Boosting (XGB) model exhibits a superior performance in all prediction tasks, with Particle Swarm Optimization (PSO) achieving R² of 0.91 for compressive strength and 0.89 for electric charge passed on test sets. Interpretability through SHAP analysis reveals that cement content and water content are the main factors affecting compressive strength, while superplasticizer dosage is a key factor affecting electric charge passed. For carbon efficiency indices, fly ash replacement of cement is an important approach to reduce C_i and Q_i, while the negative impact of recycled aggregates on performance often exceeds their carbon reduction benefits. Multi-objective optimization results indicate significant trade-offs among compressive strength, durability, CO₂ emissions, and cost. Decision schemes based on carbon efficiency indices further confirm that CO₂ emissions can be effectively reduced, while ensuring concrete performance via scientifically proportioning supplementary cementitious materials, and optimizing water-binder ratio and superplasticizer content.

This study highlighted a significant potential of an AI-driven approach for designing low-carbon concrete, effectively balancing engineering performance with environmental impact. This study could provide valuable insights into achieving tailored concrete properties via utilizing predictive modeling, parameter influence analysis with SHAP, and NSGA-III for multi-objective optimization. The proposed intelligent design method incorporating carbon efficiency indices could offer a practical strategy for developing low-carbon concrete, promoting sustainable practices in the construction industry.