Durability-informed life cycle assessment of concrete through machine learning for service life prediction

Sustainable concrete infrastructure cannot be achieved through prescriptive mix design or carbon accounting frameworks that neglect material deterioration. Despite advances in durability science, machine learning (ML), service-life modeling, and life-cycle assessment (LCA), these domains remain misaligned because they quantify performance using incompatible metrics. Durability research characterizes performance through transport-controlled degradation processes; ML infers performance statistically across heterogeneous datasets; service-life analysis defines performance by the timing of corrosion limit states; and LCA evaluates performance using functional units (FUs) often decoupled from degradation mechanisms. This fragmentation underpins many sustainability claims for low-clinker concretes. This review shows that transport parameters governing chloride ingress and carbonation vary by one to two orders of magnitude across binder systems due to chemistry, curing regime, moisture history, exposure conditions, and test methodology. Such variability destabilizes deterministic service-life predictions and renders conventional mass- or strength-based LCA comparisons physically inconsistent. While ML approaches can reduce statistical scatter, predictions remain unreliable when deterioration mechanisms, exposure descriptors, and depassivation criteria are not explicitly embedded. Within LCA, FU selection exerts a stronger influence on environmental rankings than mixture composition itself, with rankings frequently reversing when service-life- or transport-informed FUs replace volumetric metrics. To address these limitations, the review introduces a deterioration-informed sustainability framework (DISF) that unifies ML-based durability prediction, probabilistic service-life modeling, and LCA through a shared representation of deterioration trajectories. By embedding deterioration kinetics, exposure sensitivity, and temporal performance into functional units, the framework enables comparisons of binder systems and performance, shifting sustainability assessment from static carbon metrics toward performance-based decision-making. • Study reveals fundamental incompatibility between durability metrics, ML outputs, service life, and LCA frameworks. • Unifies ML, durability, and LCA through a novel Deterioration-Informed Sustainability Framework. • Critically establishes limits of ML durability predictions without embedded deterioration mechanisms. • Shows that sustainability rankings reverse when functional units scale with service life predictions. • Establishes logical admissibility gates to validate durability-based sustainability claims.