Data-Driven Compressive Strength Screening of Slag and Fly Ash Concrete Mixtures Before Trial Batching

Replacing part of the cement with blast-furnace slag or fly ash lowers the clinker content of concrete, but such mixtures often gain early strength more slowly depending on binder composition. On site, slower early strength can delay formwork removal, and in mix design the uncertain strength gain is managed with extra trial batches and higher cement contents. Stacked ensembles predict compressive strength accurately but usually apply one weight vector to all mixtures. This study tested whether regime-specific weights offer better predictions. K-means clustering divided 1030 results of the UCI dataset into four regimes of binder composition and water-to-cement ratio, and the weights of random forest, multilayer perceptron and long short-term memory learners tuned by the sparrow search algorithm were selected for each regime on out-of-fold predictions. This cluster-conditioned stacking (CCS) was compared with six other combination rules on 20 splits grouped by mixture. CCS reached a mean test root-mean-square error (RMSE) of 5.76 MPa, significantly lower than equal weighting (6.27 MPa) but not significantly different from global regression weights (5.78 MPa) or random forest alone (5.77 MPa). Screening with the model before trial batching can avoid some failed batches, each costing about a month of retesting.