Ultra-high performance concrete (UHPC) has garnered widespread attention in the field of civil engineering due to its exceptional mechanical properties, durability, and workability. These superior properties are primarily attributed to its dense microstructure and unique material composition, which typically includes cement, silica fume, quartz powder, and superplasticizers, along with reinforcing materials like steel fibers. The remarkable performance of UHPC not only meets the demands for high-performance materials in modern engineering but also contributes to reducing the life-cycle costs of structures by minimizing section sizes and extending service life.

However, the development and application of UHPC face several challenges. The complexity of the preparation process, stringent requirements for raw materials, and the high sensitivity of UHPC properties to mix design variations increase both production costs and technical difficulties. Traditional experimental design methods struggle to efficiently optimize UHPC mixtures, and issues such as shrinkage and autogenous cracking require precise design solutions. These challenges make the composition design and performance prediction of UHPC a focal point in the field. With advancements in computer technology and data science, machine learning (ML) offers new tools and approaches for UHPC research and development. Therefore, this study aims to promote the application of ML technology in the UHPC field, thereby advancing the intelligent development of advanced construction materials.

The study introduced a comprehensive ML-based framework for UHPC performance prediction and precise design. The framework began with data cleaning, where the multiple imputation by chained equation (MICE) method with a predictive mean matching (PMM) kernel was employed to impute missing data in the UHPC database. Subsequently, the isolation forest algorithm was applied to identify and eliminate outlier data, thereby improving the quality of the dataset. The cleaned and optimized dataset was then used to train an XGBoost model, optimized via Bayesian hyperparameter tuning, to accurately predict various UHPC properties. Two AI-driven approaches for UHPC mix design were also proposed: one that combined the MAA model with ML predictions for multi-performance optimization, and another that integrated the ML model with genetic algorithm (GA) for multi-objective optimization.

The application of MICE combined with PMM resulted in a substantial improvement in the accuracy of the imputed data, as evidenced by the reduced average deviation between imputed and original data values. This enhanced imputation process allowed for a more reliable dataset, which directly contributed to the performance of the XGBoost prediction models. Outlier detection via the isolation forest algorithm effectively removed data points that exhibited significant deviation from the norm, particularly in compressive strength and density measurements. The refined dataset led to more accurate and consistent predictions across all performance metrics. After hyperparameter optimization, the XGBoost model demonstrated exceptional predictive capabilities, with notable improvements in performance metrics. The predictive accuracy of the model was further validated against experimental data, confirming its effectiveness in real-world applications. These results highlight the potential of ML techniques in advancing the field of UHPC research. The integration of MICE and PMM into the data preprocessing pipeline ensured more accurate and reliable datasets, which were crucial for developing robust predictive models. The success of the XGBoost model, particularly after Bayesian optimization, underscores the importance of hyperparameter tuning in enhancing model performance.

Furthermore, the study proposes two ML-assisted design strategies for UHPC: 1) a combined physical packing theory and ML prediction model for initial mixture design, and 2) a multi-objective optimization approach using a metaheuristic algorithm for fine-tuning UHPC compositions. These strategies provide a framework for the intelligent and efficient design of UHPC materials, aligning with the growing demand for high-performance, sustainable construction materials.

This research demonstrates the significant potential of ML in predicting and optimizing the performance of UHPC. By addressing data challenges and enhancing prediction models, the study provides valuable insights into the application of ML in UHPC design. The proposed ML-assisted design strategies offer a practical approach to developing UHPC with tailored properties, paving the way for more sustainable and efficient construction practices.