Harnessing machine learning for accurate estimation of compressive strength of high-performance self-compacting concrete from non-destructive tests: A comparative study

Abstract

This study investigates the compressive strength (CS-28d) of high-performance self-compacting concrete. To accurately estimate the CS-28d, the study employed regression techniques (linear regression (LR) and stepwise polynomial regression (SPR)) as well as machine learning techniques (M5Prime, Random Forest, Random Tree and REP Tree). The study was based on 600 experimental results were collected from literature, with the rebound number (R), and ultrasonic pulse velocity (Vp) used to predict the CS-28d. The results indicate that the M5Prime model performed the best, with the highest coefficient of determination (Pearson R of 95.06 %) and the lowest root mean square error (RMSE of 4.84516 %). Among the tested regression models, the SPR model demonstrated the best performance and an empirical equation was introduced to estimate the CS-28d of a high-performance self-compacting concrete with high accuracy. An in-depth exploration of the model's sensitivity to its input parameters was conducted to quantify their effect on the CS-28d outcome. The results showed that the rebound number (R) had the most significant impact, with a sensitivity analysis parameter of 95 %. This was followed by the ultrasonic pulse velocity (Vp) had a relatively insignificant effect of only 5 % on the CS-28d values.