Concrete as a typical multiphase composite material exhibits properties closely related to raw material quality, mix proportion design, curing conditions, and types of admixtures. Especially in the field of alkali-activated concrete (AAC), a predominant use of industrial solid wastes as raw materials introduces inherent complexities, including intricate microscopic reaction mechanisms, highly unpredictable performance characteristics, leading to the absence of a widely accepted reliable design standards. However, conventional approaches such as experimental characterization and numerical simulations demand substantial financial and labor resources and yield outcomes constrained due to material variability and methodological assumptions. Machine learning (ML) as a data-driven approach has a potential to promote the implementation of the standard from another perspective (i.e. data aspect) with less cost waste via establishing a predictive model especially a multi-output predictive models with high accuracy and robustness. Therefore, this study was to use machine learning in property prediction of alkali-activated concrete (AAC) via establishing a robust dual-output hybrid BPNN framework, resolving challenges in multi-objective optimization and overfitting through metaheuristic algorithms, and providing mechanistic insights by a feature importance analysis.

This study established two independent databases for model training and generalization capability evaluation. Initially, the datasets were preprocessed and analyzed through data normalization, Pearson correlation coefficient calculation, and descriptive statistical analysis. Subsequently, four metaheuristic algorithms, i.e., Ant Colony Optimization (ACO), Genetic Algorithm (GA), Whale Optimization Algorithm (WOA), and Grey Wolf Optimizer (GWO), were employed to optimize the hyperparameters of the Backpropagation Neural Network (BPNN), with a predictive accuracy enhanced via ten-fold cross-validation. The predictive performance of five models (original BPNN and optimized variants) was systematically compared based on four evaluation metrics, i.e., coefficient of determination (R²), mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean square error (RMSE). The generalization capability was further validated using a novel independent database, ultimately identifying the optimal model capable of high-precision simultaneous prediction of compressive strength and slump. Finally, the principal component analysis (PCA) was utilized to rank the influence of input variables on the output properties, enabling the selection of optimal mix proportions from the database through comprehensive performance evaluation of dual-output variables.

The Pearson correlation coefficient analysis indicates the rationality of input variable selection via confirming the absence of significant multicollinearity, thereby mitigating its adverse effect on the model predictions. During model training, the metaheuristic optimization algorithms significantly enhances the predictive performance of the Backpropagation Neural Network (BPNN), demonstrating that a hyperparameter optimization effectively improves the model capability. However, the generalization capability evaluation using a novel independent database reveals that three optimized models exhibit an inferior performance, compared to the unoptimized BPNN, contradicting their training-phase results. This discrepancy indicates overfitting in these three models during training, potentially attributable to excessive feature extraction induced by ten-fold cross-validation. In contrast, the Ant Colony Optimization-enhanced BPNN (ACO-BPNN) consistently demonstrates superior predictive accuracy and stability across both training and generalization assessments.

Furthermore, the principal component analysis (PCA)-based feature importance ranking identifies ground granulated blast furnace slag (GGBFS) as the most influential factor for 28-d compressive strength, while the coarse aggregate/fine aggregate ratio (CA/FA) exerts the dominant control over slump variability.

A hybrid-BPNN model capable of predicting both compressive strength and slump was trained and modeled in an entirely new database with appreciable accuracy levels. Five predictive models before and after optimization were established and compared. The ACO-BPNN had the optimum predictive accuracy (based on four evaluation indicators), and the strongest stability and robustness (based on the error bars) in testing process for both two properties. For the assessment of compressive strength prediction in testing process, the evaluation indicators could involve R²=0.932, MAE=2.972, RMSE=4.043, and MAPE=0.083. For slump degree prediction, the evaluation indicators could involve R²=0.929, MAE=12.212, RMSE=19.570, and MAPE=0.223. The WOA-BPNN, GA-BPNN, and GWO-BPNN algorithms exhibited pronounced overfitting tendencies when applied to multi-output variable modeling. This phenomenon could stem from excessive feature space optimization caused by high-frequency data reuse inherent in the ten-fold cross-validation framework. The principal component analysis (PCA) indicated that the relative importance of eleven input variables for compressive strength could be arranged in a descending order as follows: GGBFS; curing time, Na₂SiO₃, total aggregates, CA/FA, superplasticizer, OPC, SF, curing temperature, FA, and NaOH; and those for slump could be arranged in a descending order as follows: CA/FA, curing time, superplasticizer, SF, Na₂SiO₃, GGBFS, OPC, total aggregates, curing temperature, NaOH, and FA.