Prediction and Control of Fabric-Form Concrete Flowability Using a BO-MLP-RF Ensemble Learning Model

To address the coupled effects of multiple factors on the flowability of fabric-form concrete during construction, as well as the reliance of conventional mixture-proportion design on empirical experience and its low trial-mixing efficiency, this study develops a Bayesian-optimization-based multilayer perceptron-random forest ensemble learning model (BO-MLP-RF) for slump-flow prediction and mixture-proportion optimization of fabric-form concrete. Based on 313 sets of self-compacting concrete test data, nine parameters are selected as model inputs: water-to-cement volume ratio (VW/VP), fly ash-to-cement volume ratio (VFA/VP), cement content (C), fly ash content (FA), fine aggregate content (Fagg), coarse aggregate content (Cagg), water content (W), polycarboxylate superplasticizer dosage (SP), and paste flow spread (SFpaste). Concrete slump flow (SF) is used as the output. Bayesian optimization is employed to tune the key hyperparameters of the MLP and RF models, and the ensemble weights are determined from the validation-set errors. The results show that the BO-MLP-RF model achieves high prediction accuracy and generalization performance, with a coefficient of determination R2 of 0.976 and a root mean square error (RMSE) of 14.80 mm for the training set, and an R2 of 0.941 and an RMSE of 18.95 mm for the test set, outperforming the individual MLP and RF models. SHAP-based interpretability analysis further indicates that SFpaste, W, and VW/VP are the three dominant features affecting slump-flow prediction, with relative importances of 24.6%, 20.5%, and 19.2%, respectively, and a cumulative contribution of 64.3%. This result highlights the important effects of paste flowability and water-related parameters on the flowability of fabric-form concrete. On this basis, with a target slump-flow range of 620–690 mm, the BO-MLP-RF prediction model is combined with a genetic algorithm to optimize the mixture proportions of fabric-form concrete. The optimized candidate solution yields an SFpaste of 339.35 mm and a predicted SF of 680.42 mm, satisfying the specified flowability requirement. The results demonstrate that the BO-MLP-RF ensemble learning model can effectively characterize the complex nonlinear relationships between mixture-proportion parameters and the flowability of fabric-form concrete and can provide quantitative support for mixture-proportion design and flowability control. However, the optimized result is a model-predicted candidate solution, and its engineering applicability should be further verified through independent mixture-proportion tests.