Sustainable Optimization of Ultra-High-Performance Concrete Using Recycled Fines and Co₂ Mitigation: A Machine Learning and Swarm Intelligence Framework

The increasing environmental repercussions associated with cement-based construction materials, particularly ultra-high-performance concrete (UHPC), necessitate the adoption of innovative methodologies that enhance mechanical properties while mitigating carbon emissions. This investigation examines the ecological implications of UHPC and demonstrates how machine learning methodologies and advanced optimization techniques can enhance its mechanical properties while reducing carbon emissions by analyzing UHPC formulations incorporating recycled concrete powder (RCP). The Random Forest (RF) and Long Short-Term Memory (LSTM) models, augmented by Particle Swarm Optimization (PSO) and Chicken Swarm Optimization (CSO), were employed to forecast and optimize performance metrics. The dataset comprised 380 instances amalgamating real and AI-generated methodologies utilizing a Variational Autoencoder, encompassing the principal components of UHPC such as cement, fly ash, silica fume, and the water-to-binder ratio, with CO2 emissions serving as the dependent variable. Findings indicated that RF models, particularly RF-CSO and RF-PSO, attained high accuracy Root Mean Square Error (RMSE: 7.04), Mean Absolute Error (MAE: 4.16), whereas LSTM-PSO demonstrated superior performance in recognizing temporal patterns (RMSE: 46.22, MAE: 4.53), underscoring its capability to manage non-linear relationships. The standalone LSTM exhibited subpar performance, thereby accentuating the necessity for optimization. The optimized UHPC formulations realized a CO₂ reduction of up to 20.83%, congruent with the guidelines established by the Global Cement and Concrete Association (GCCA), thereby representing a significant advancement toward the development of low-carbon concrete solutions. The research accentuates the potential of machine learning and optimization techniques in facilitating sustainable, data-driven design of UHPC, with RF models requiring minimal hyperparameter calibration, while LSTM models necessitate more extensive refinement to ensure reliability and minimize errors.