A data-driven multi-indicator quality assessment framework for concrete: conformal uncertainty quantification and copula-based tail-dependence analysis

Concrete acceptance continues to rely almost exclusively on compressive strength, yet deterioration-driven failures under aggressive exposure, such as reinforcement corrosion and freeze–thaw attack, remain a leading cause of premature decommissioning, even for strength-compliant members. This mismatch points to a fundamental gap: in-service performance is governed not by any single metric but by the joint risk of non-compliance across coupled strength, durability and denseness indicators, whose contradictions the current pass/fail paradigm systematically overlooks. Addressing this gap, the present study develops a data-driven multi-indicator quality assessment workflow that chains point prediction, distribution-free uncertainty quantification and tail-dependent joint non-compliance modelling into an integrated assessment pipeline. Eight machine-learning models are benchmarked for simultaneous prediction of compressive strength, electrical resistivity and ultrasonic pulse velocity on 4420 specimens with synchronous measurements; the reported accuracy (R 2 = 0.984–0.989 for compressive strength) strictly reflects in-domain interpolation, as leave-one-group-out cross-validation by mix design yields negative R 2 for key indicators. Split conformal prediction then furnishes finite-sample prediction intervals whose coverage guarantees hold without distributional assumptions, while grouped analyses expose pronounced coverage heterogeneity across curing ages and aggregate types. A trivariate t-copula fitted to the prediction residuals reveals that the dependence among indicators is concentrated in the extreme tails: when compressive-strength residuals fall in the worst 5% region, the conditional probability of concurrent resistivity deterioration is roughly sevenfold higher than the independence baseline (an order-of-magnitude estimate given the small tail sample), a finding that reframes “strength-compliant yet durability-deficient” cases as a robustness deficit of the acceptance system itself. Joint non-compliance probabilities estimated via Monte Carlo simulation provide a graded quality-screening map that can inform acceptance decisions and life-cycle quality management workflows. The framework is modular: base model, conformal strategy and copula family are each independently replaceable. It is also, in principle, readily transferable to broader multi-indicator quality assessment problems in construction materials and infrastructure systems.