Interpretable Machine Learning for Compressive and Flexural Strength Prediction of Fly Ash Blended 3D Printed Concrete with Uncertainty Quantification

Fly ash (FA) blended 3D printed concrete (3DPC) offers improved sustainability but requires strength prediction models validated at the mix-composition level rather than within familiar formulations. This study applies leave-one-mix-out (LOMO) cross-validation to benchmark eight machine learning algorithms on 126 experimental records spanning seven FA-blended 3DPC compositions (FA 0–15 wt.%, W/B 0.30–0.35, age 1–28 days). ExtraTrees and ElasticNet achieve the highest composition-level generalisation for compressive strength (CS, R2=0.786±0.253) and flexural strength (FS, R2=0.915±0.061, RMSE =0.301 MPa), respectively. SHAP analysis identifies FA replacement percentage as the dominant CS predictor (mean |SHAP|=3.92 MPa, negative effect) and curing age as the dominant FS predictor (1.06 MPa), a reversal consistent with distinct failure mechanisms under compression and flexure. FA substitution at 5%–7.5% with W/B =0.30 is identified as the optimal design range, with content limited to 5% when structural loading within three days is required. LOMO residual prediction intervals are well calibrated from 50% to 99% nominal coverage, with empirical deviations not exceeding 1.4 percentage points; at the 90% level, interval half-widths are ±4.60 MPa for CS and ±0.494 MPa for FS. Design-space mapping over the FA% × age prediction grid at W/B =0.30 identifies minimum curing ages of 3 days at FA ≤7.5 wt.% and 13.5 days at FA =10 wt.% for simultaneous CS and FS target compliance; Sobol variance decomposition confirms FA% (S1=0.553) and curing age (S1=0.526) as the dominant sensitivity drivers for CS and FS, respectively, in close agreement with the SHAP attribution. A graphical user interface integrating the two best models returns simultaneous CS and FS predictions with 90% calibrated prediction intervals from three mix design inputs, supporting uncertainty-aware strength estimation without programming expertise. Both models are local interpolators conditioned on the experimental design space (FA 0–15 wt.%, W/B 0.30–0.35, age 1–28 days); predictions outside these boundaries have not been validated and should be interpreted with caution. The dataset and source code are publicly available at https://github.com/lucassivan/ML-FA-3DPC.