Machine learning based conformal predictors for uncertainty-aware compressive strength estimation of concrete

Estimating concrete compressive strength is crucial for accurately predicting its performance, optimising material usage, and ensuring the durability and safety of the structure. Traditional machine learning (ML) models have primarily focused on deterministic predictions of compressive strength, often overlooking the uncertainty associated with these estimates. However, concrete is a non-homogeneous material with complex and variable behaviour, making it inherently difficult to predict compressive strength with precision. Therefore, incorporating uncertainty into predictive modelling is essential for producing more reliable and practical results in real-world engineering applications. This study addresses this gap by proposing a comprehensive framework for uncertainty quantification in concrete strength estimation using conformal prediction methods. In this comprehensive study, eight distinct machine learning models are systematically integrated with six conformal prediction variants to construct statistically rigorous prediction intervals. To evaluate the performance of the models holistically in engineering contexts, a novel Efficiency Score (ES) is proposed, combining empirical coverage, mean interval width, and point prediction accuracy. The findings reveal notable trade-offs between predicted interval width and empirical coverage across the model spectrum. Among the tested combinations, LightGBM coupled with Jackknife+ emerges as the most effective configuration, demonstrating the highest efficiency score. Additionally, conformal predictors exhibit satisfactory adaptation to heteroscedasticity, which arises in the predictions of higher-grade concrete ( > 40 MPa). Thus, the proposed framework empowers more informed decision-making in concrete design and quality control by providing robust uncertainty bounds advancing beyond traditional deterministic point predictions to support risk-aware infrastructure development. • Proposes ML-based conformal predictors for concrete strength estimation. • Provides distribution-free uncertainty estimates in concrete compressive strength. • Compares an array of machine learning models for uncertainty-aware concrete strength estimation. • Systematically analyses different conformal prediction methods. • Proposes a novel efficiency score for ML models.