Predicting the Workability of PCE-Modified Cement–Fly Ash Pastes with Explainable Machine Learning
Abstract
Reliable assessment of polycarboxylate ether (PCE)–binder combinations requires predictive models whose interpolation performance is distinguished from their performance when an entire formulation is absent from training. This study examined 616 cement–fly ash pastes prepared using twenty-two in-house PCE formulations, four fly ash replacement levels (0, 15, 30, and 45 wt %), seven PCE dosages (0.50–2.00 wt % of binder), and a fixed water-to-binder ratio of 0.35. Marsh funnel flow time was measured for all mixtures, whereas mini-slump measurements were available for 252 mixtures comprising nine PCE formulations. Ten regression algorithms were evaluated using 3 × 5-fold repeated cross-validation and nine non-algebraically redundant input variables. XGBoost achieved R2 = 0.946 ± 0.027 and RMSE = 8.40 s for flow time, and R2 = 0.778 ± 0.061 and RMSE = 0.471 cm for mini-slump. These random-resampling results describe interpolation among formulations represented in the training folds. When each PCE chemistry was withheld in turn, the mean R2 was 0.865 ± 0.198 for flow time and 0.034 ± 0.530 for mini-slump. Performance also decreased when the boundary fly ash levels were withheld. SHAP and permutation analyses identified PCE dose and fly ash replacement as the strongest predictors, while the formulation-level descriptors made smaller contributions. These findings represent associations learned from the present dataset and do not constitute direct evidence of adsorption or dispersion mechanisms. At a nominal 90% level, split-conformal intervals achieved empirical coverages of 85.5% for flow time—moderately below the nominal level, within finite-sample binomial fluctuation—and 90.2% for mini-slump on a random test partition; across fifty repeated train–calibration–test splits, the mean coverages were 89.4 ± 3.6% and 90.1 ± 5.8%. The flow-time model is suitable for preliminary screening within the investigated factor ranges, whereas the mini-slump model should be restricted to interpolation among the sampled formulations.