Prediction of Fresh and Mechanical Properties of Self-Compacting Concrete Using Gaussian Process Regression

Abstract

Self-compacting concrete (SCC) incorporating ceramic waste powder (CWP) as a partial cement replacement offers a route to reduce both landfill burden and embodied carbon in construction, but characterising the combined effect of mix design variables on the full suite of fresh and hardened properties normally demands extensive laboratory testing. This study evaluates Gaussian Process Regression (GPR) as a data-efficient surrogate modelling approach for predicting twelve fresh and mechanical properties of CWP-based SCC from three mix design inputs: cement content, CWP content, and water-to-binder (w/b) ratio. A dataset of twenty-one experimentally characterised SCC mixes, spanning slump flow, T500 time, V-funnel time, L-box ratio, segregation resistance, compressive strength (7, 28, and 90 days), split tensile strength (7, 28, and 90 days), and 28-day flexural strength, was modelled using independent GPR models with a Matern 5/2 kernel and white-noise term, with performance assessed exclusively through leave-one-out cross-validation (LOOCV) given the small sample size. Eleven of the twelve properties were predicted with LOOCV coefficients of determination (R2 ) between 0.80 and 0.99, with the strongest performance observed for slump flow and split tensile strength (R2 ≥ 0.96). Twenty-eight-day flexural strength was the exception, with a negative LOOCV R2 indicating that the three mix design inputs alone do not explain its variability in this dataset. The results demonstrate that GPR, combined with rigorous LOOCV validation and explicit predictive uncertainty, is a viable and transparent tool for mix-design-stage property prediction in small experimental SCC datasets, while also illustrating the diagnostic value of LOOCV in identifying properties that require additional explanatory variables.