Inaccurate prediction of shrinkage in self-compacting concrete (SCC) may result in underestimated cracking risk, increased permeability, serviceability deterioration, and reduced long-term durability of concrete structures. Although conventional empirical shrinkage models are widely used in engineering practice, their accuracy is often limited when applied to SCC mixtures with high paste volume, mineral admixtures, manufactured sand, and high-range water-reducing admixtures. Recent machine-learning-based models provide an alternative approach, but single learning algorithms may show limited robustness for small and heterogeneous datasets. In addition, random sample-level data splitting may introduce information leakage when shrinkage measurements obtained at different curing ages from the same mixture are assigned to both training and testing sets. To address these issues, this study develops a stacking-based ensemble learning framework for SCC shrinkage prediction using mixture proportions and curing age as input variables. A multi-source database containing 61 mixture designs and 448 data samples was established from published experimental studies. To obtain a more realistic assessment of model generalization, a mixture-level validation strategy was adopted, in which all age-dependent samples from the same mixture were assigned exclusively to either the training set or the testing set. Under this strategy, 358 data samples were used for model training and 90 data samples were used for independent testing. Four base learners, including multilayer perceptron (MLP), support vector regression (SVR), decision tree (DT), and gradient boosting decision tree (GBDT), were constructed and integrated through different ensemble configurations. The Stacking-SVR model achieved the best overall performance on the independent testing set, with a mean absolute error (MAE) of 13.6 με and a mean absolute percentage error (MAPE) of 7.5%. Compared with GBDT, Stacking-GBDT, and DT models, the proposed Stacking-SVR model reduced the MAPE by approximately 10.7%, 11.8%, and 35.3%, respectively. Stability and applicability analyses further indicate that the proposed framework can provide reliable shrinkage predictions within the investigated mixture and curing-age ranges. However, because the model was developed from a compiled database and does not explicitly include environmental variables such as relative humidity and temperature, its use should be limited to parameter ranges represented in the database. Overall, the results demonstrate that stacking ensemble learning combined with mixture-level validation offers a leakage-controlled and engineering-oriented approach for SCC shrinkage prediction.
Wang et al. (Tue,) studied this question.