Integrating uncertainty and explainable machine learning in performance-adjusted life cycle assessment of polyamide fiber-reinforced concrete

Accurate estimation of the environmental footprints of construction materials is critical for advancing sustainable design. However, life cycle assessment (LCA) studies of fiber-reinforced concrete often face methodological limitations, including inconsistent functional units, insufficient handling of performance variability, and inadequate treatment of data uncertainty. This study introduces a feature-driven framework that combines machine learning (ML) with cradle-to-gate LCA to evaluate 144 experimentally derived concrete mixes containing virgin and recycled polyamide (PA) fibers. To account for uncertainty arising from empirical limitations, experimental data were preprocessed through imputation, normalization, and outlier filtering. A tensile strength (TS)-based functional unit was adopted to enable fair comparisons among mixes with different structural capacities. Among five tested models, Support Vector Regression (SVR) achieved the highest predictive accuracy (R-squared = 0.98), and Shapley Additive Explanations (SHAP) enabled transparent attribution of feature contributions. SHAP analysis identified water and cement content as the most influential drivers of TS, while cement, superplasticizer, and virgin fiber content dominated environmental impacts. Although fiber type showed a modest influence on mechanical performance, it significantly affected environmental outcomes. Replacing virgin with recycled PA fibers reduced global warming potential, fossil fuel depletion, and human toxicity by up to 93 %, 97 %, and 41 %, respectively. Overall, this framework demonstrates how integrating ML with performance-adjusted LCA can explicitly capture data uncertainty and experimental variability, advancing methodological rigor and guiding the design of low-impact, high-performance concrete aligned with circular-economy goals.