Predicting frost resistance of rubberized concrete for sustainable design: a comparative analysis of regression trees and stepwise regression models

Abstract This study presents a machine learning-driven comparative analysis to predict frost resistance in rubberized concrete, focusing on the interpretability and performance of three models: Stepwise Linear Regression (SLR), Stepwise Polynomial Regression (SPR), and Classification and Regression Tree (CART). Leveraging historical experimental data, the methodology integrates rigorous preprocessing via Interquartile Range (IQR) outlier detection and tenfold cross-validation to ensure robustness. Key variables, including rubber content (0–20% fine aggregate replacement), water-cement ratio, and freeze–thaw cycles, were analyzed to balance durability and sustainability goals. The CART model offered interpretable decision rules (test R 2 = 97.01%) identifying critical thresholds like rubber content ≤ 15%, providing a clear guideline for maximizing durability. Comparatively, the study advances prior Artificial Neural Network (ANN)-based approaches by delivering mathematical transparency (SLR, SPR) and actionable insights (CART). These outputs directly guide mix design: the SPR equation enables precise prediction of frost resistance for specific combinations of mix parameters, while the CART rules establish safe application limits. This moves beyond prediction to offer practical tools for optimizing sustainable concrete mixes under frost exposure, prioritizing both accuracy and implementable insight.