Jia Chen, Zhicheng Liao, Jianbo Huang
Accurate prediction of 3D printed concrete (3DPC) compressive strength (CS) with quantified uncertainty is central to reliability-based mix design verification, yet existing data-driven models return deterministic point estimates only, with no calibrated basis for structural safety assessment. A Gaussian process with an automatic relevance determination radial basis function (GP_ARD_RBF) kernel is trained on 254 CS records compiled from 24 experimental programmes (CS: 11.1–189.0 MPa, 13 mix-design features) and benchmarked against seven deterministic baselines over 30 independent 80/20 random splits. GP_ARD_RBF achieves and RMSE MPa, statistically equivalent to the four leading ensemble models (Wilcoxon signed-rank, – ), while providing a closed-form approximately calibrated predictive distribution: PICP , PICP , and a mean 90% prediction interval width of 32.32 MPa across 30 splits. ARD length-scales identify (OPC) as the dominant feature (62.9% of normalised importance), convergently validated by GP-SHAP attribution (Spearman , ); the four leading features ( (OPC), (OPC), (W/B), and (WRA)) account for 86.5% of total SHAP attribution. Uncertainty decomposition yields a learned aleatoric noise floor of MPa, fixed by model construction as a single scalar shared across all inputs, alongside a composition-dependent epistemic component: SCM-rich mixtures exhibit the highest epistemic uncertainty (6.58 MPa), identifying the compositional subspace where additional experimental data would most reduce prediction error. Reliability index maps over the (W/B)– (OPC) design space show that the fib Model Code target is achievable at MPa only at (W/B) and (OPC) ; at MPa the maximum across the full grid is 1.54, exposing the impact of GP epistemic uncertainty on the achievable reliability index. The dataset and source code are openly available at https://github.com/lucassivan/3DPC-GP .