Abstract:

Ultra-high performance concrete （UHPC） demonstrates immense potential for enhancing structural load-bearing capacity and service life. However， UHPC mix proportion design involves highly nonlinear couplings among multiple components and parameters. Conventional trial-and-error methods are constrained by long experimental cycles， high costs， and a single objective， making it difficult to achieve a systematic balance among mechanical properties， economic cost， and carbon emissions. In recent years， machine learning （ML） methods have been widely applied to performance prediction and mix proportion design of UHPC. Nevertheless， existing studies have yet to establish a systematic design framework that integrates high-quality data management， accurate interpretable prediction， and multi-objective collaborative decision-making.A UHPC mechanical property database was constructed based on an extensive review of published literature， comprising 17 input features and 2 output features. Optimal input dimensionality was determined through an embedded feature engineering method. A dynamic-weight stacking ensemble model （DW-Stacking） was developed using Random Forest （RF）， XGBoost （XGB）， CatBoost （CB）， and LightGBM （LGB） as base learners， incorporating a prediction error-based dynamic weighting and cross-fold smoothing mechanism. Three hyperparameter optimization strategies， namely Adaptive Grid Search （AGS）， Bayesian Optimization （BO）， and Enhanced Grey Wolf Optimization （EGWO）， were tested. The BO was selected for the final base learner optimization. SHAP analysis was further employed to quantify the contribution and directional influence of each input feature on UHPC mechanical properties. For multi-objective optimization， flexural strength， compressive strength， unit cost， and CO₂ emissions were defined as four parallel optimization objectives. An improved NSGA-Ⅱ algorithm was constructed， incorporating Latin hypercube sampling， dynamic parameter regulation， and parallel search operators. An entropy-weight TOPSIS decision method was then applied to select optimal mix proportions from the resulting Pareto front. The optimization results were verified through actual testing.The DW-Stacking model achieves test-set R² values of 0.927 and 0.958 for flexural and compressive strength prediction， respectively. Compared with the BO-RF baseline， RMSE is reduced by 45.3% and 25.6%； compared with standard Stacking， further RMSE reductions of 36.2% and 17.3% are attained. These results demonstrate that the dynamic weighting and smoothing mechanism effectively suppresses noise introduced by lower-quality sub-models and substantially enhances prediction robustness. SHAP analysis identifies steel fiber content， curing age， and silica fume content as the important factors governing UHPC mechanical properties， all of which exhibit positive correlations with both flexural and compressive strengths. In the multi-objective optimization stage， the improved NSGA-Ⅱ achieves an 18.4% increase in the hypervolume indicator and approximately 281% faster convergence relative to the original algorithm， reflecting superior solution quality and search efficiency. Experimental validation confirms that within the mix proportion design range， where cost and carbon emissions reductions do no exceed 25%， prediction errors for both flexural and compressive strengths of UHPC are consistently controlled within ±9%. When the cost and carbon-emission reduction targets are extended to 50%， the content of key materials in the optimized mix proportions deviate substantially from the training data distribution. This result in prediction errors exceeding ±25% for both output features and revealing the boundary conditions of the model’s reliable predictive range.This research formulates an interpretable， AI-driven collaborative design framework for UHPC. The proposed DW-Stacking model achieves superior accuracy in predicting the mechanical properties of UHPC. SHAP analysis effectively elucidates the positive effects of steel fiber， curing age， and silica fume content. Furthermore， the improved NSGA-Ⅱ algorithm efficiently realizes the synergistic optimization of different conflicting objectives. Experimental results confirm that the model can reduce the unit cost and carbon footprint of UHPC by about 25% while meeting the minimum mechanical-property requirements specified for structural UHPC， providing a reliable scientific tool for engineering practice.

Key words: ultra-high performance concrete, machine learning, mix proportion, multi-objective optimization, performance prediction