Matching metaheuristics to machine learning for concrete drying shrinkage prediction toward cost-effective mix design

Empirical models have limited accuracy for coupled mixture-environment-geometry-time effects on concrete drying shrinkage. An interpretable machine-learning framework was developed using 42,099 records from the NU Shrinkage-Creep Database, benchmarking nine algorithms under Bayesian optimization and four metaheuristics. GA-LightGBM achieved the best performance, with an RMSE 78.4% lower than GL2000. Bayesian optimization favored GBDT and CatBoost whereas PSO favored MLP, RF, and AdaBoost. SHAP analysis identified drying duration as the dominant variable, revealed a threshold effect of elastic modulus near 30,000 MPa, and showed step-like temperature effects between 20 and 25 °C. External validation on 116 data points from three independent studies yielded R 2 = 0.9627, RMSE = 37.7 με, and MAE = 31.4 με. In two shrinkage-constrained mix-design cases, the surrogate-assisted optimizer yielded mixtures with unit material costs 19.1% and 10.3% lower than feasible baselines under 365-day shrinkage constraints, supported by a graphical user interface.