AI-Driven computational methodologies of optimized mix design and performance prediction of ultra-high-performance concrete (UHPC): a systematic review

Ultra-high-performance concrete (UHPC) has progressed from a laboratory innovation to a field-deployed structural material; however, its mixture design is still largely guided by empirical trial-and-error because targeted performance emerges from tightly coupled, nonlinear interactions among particle packing, binder chemistry, water-to-binder ratio, admixture dosage, fiber characteristics, curing history, and microstructural evolution. This systematic review synthesizes how artificial intelligence (AI), including machine learning (ML), deep learning (DL), fuzzy/expert systems, metaheuristic optimization, and emerging large language model (LLM) workflows, is being used to accelerate UHPC proportioning, predict fresh, hardened, durability properties, and enable sustainability-aware design. Across the reviewed studies, tree ensembles and nonlinear learners (e.g., XGBoost, Random Forest, ANN, SVR) consistently outperform conventional regression for capturing UHPC nonlinearity, with many works reporting high predictive accuracy (commonly R 2 approaching ∼0.90–0.99) for strength- and workability-related outputs, while hybrid optimization (e.g., PSO/GA-assisted pipelines) improves multi-objective search for mixtures balancing strength, flowability, cost, and CO 2 . AI-enabled sustainability pathways prominently include binder minimization, SCM/recycled-constituent integration, and CO 2 -oriented optimization, yet practical deployment remains constrained by heterogeneous and underspecified datasets, limited external validation, underrepresentation of time-dependent durability and microstructure-informed features, and low interpretability in black-box models. Unlike prior UHPC reviews that focus on isolated prediction tasks, this paper integrates prediction, optimization, sustainability objectives, microstructure-level learning, and LLM/NLP-assisted knowledge workflows, and proposes a hybrid AI–physics roadmap for reproducible, field-validated, and interpretable digital mix-design pipelines. Overall, AI is reshaping UHPC development toward autonomous, sustainability-aware material design, provided that future research prioritizes standardized reporting, leakage-safe validation, uncertainty/interpretability, and closed-loop lab-to-field verification.