Advancing non-destructive concrete compressive strength estimation: Large-Scale datasets and machine learning framework

Non-destructive test (NDT) methods provide an indirect assessment of the compressive strength of in-situ concrete structures. While traditional static models effectively capture the behaviour of small-scale localised datasets, their accuracy diminishes when applied to larger, aggregated datasets, where increased variability in NDT measurements introduces greater uncertainty in predicting concrete compressive strength. This paper presents three exhaustive, largest-to-date NDT databases on the ultrasonic pulse velocity (UPV), rebound hammer (RH), and SonReb methods, comprising 16,531 test results from 115 studies. First, existing empirical models are evaluated against global dataset trends. New relationships are fitted to reflect the global behaviour of each NDT method, highlighting their innate limitations in capturing large-scale variability. A comprehensive three-phase machine learning (ML) program is then introduced, studying the effects of incomplete features with varying levels of missing data on model performance. Seven diverse ML models are included in Phase 1, while Phase 2 assesses different imputation strategies. Phase 3 integrates the top-performers with a Tree-Structured Parzen estimator (TPE) optimisation algorithm to refine hyperparameters and maximise performance. Across all phases, CatBoost regression emerged as the most robust predictive model due to the high proportion of categorical variables included within the databases. The TPE-CatBoost models achieved final R² values of 0.928, 0.896, and 0.947 for UPV, RH, and SonReb, respectively. Finally, a Django-based web application was deployed on a cloud server (https://recreate-ndt.onrender.com/), allowing practitioners to generate real-time compressive strength predictions for new NDT results. These novel datasets and ML tools can power future innovation through more advanced data-driven modelling.