Interpretable Machine Learning for Compressive Strength Prediction of GGBS-Fly Ash Blended Concrete: Generalization Across Training-Set Compositions and Feature-Attribution Sensitivity.

Supplementary cementitious materials (SCMs), particularly ground granulated blast-furnace slag (GGBS) and fly ash, are widely used in concrete, but the effect of training-set SCM composition on prediction of GGBS-fly ash blended mixtures remains insufficiently examined. Six regression algorithms were compared using formulation-grouped nested cross-validation, after which the retained algorithm was optimized on the complete development set and evaluated on a formulation-disjoint hold-out set. The retained model was then examined under six training-set compositions on the same GGBS-fly ash blended hold-out subset, with additional analyses of training-set size, blended-formulation exposure, hyperparameter sensitivity, and SHAP attribution under alternative feature representations. XGBoost achieved a mean outer-fold RMSE of 5.396 MPa and a hold-out RMSE of 5.095 MPa; on the blended subset, training on all mixture types and on blended mixtures only yielded RMSEs of 5.372 and 5.346 MPa, respectively, compared with 7.095 MPa when blended formulations were excluded, while SCM SHAP attributions varied across feature representations. The retained XGBoost model showed that blended-mixture prediction performance within this dataset varied with training-set SCM composition and blended-domain representation, while SCM feature attributions varied with input representation.