Performance Analysis of Artificial Neural Network and Its Optimized Models on Compressive Strength Prediction of Recycled Cement Mortar

In the background of sustainable development in the construction industry, recycled cement mortar (RCM) has emerged as a research hotspot due to its eco-friendly features, where mechanical properties serve as critical indicators for evaluating its engineering applicability. This study proposes an artificial neural network (ANN) model optimized by intelligent algorithms, including the GWO (grey wolf optimizer), PSO (particle swarm optimization), and a GA (genetic algorithm), to predict the compressive strength of recycled mortar. By integrating experimental and prediction data, we establish a comprehensive database with eight input variables, including the water-cement ratio (W/C), cement-sand ratio (C/S), fly ash content (FA), aggregate replacement rate (ARR), and curing age. The predictive performance of neural network models with different database sizes (database 1: experimental data of RCM; database 2: experimental data of RCM and ordinary mortar; database 3: model prediction data of RCM, experimental data of RCM, and ordinary mortar) is analyzed. The results show that the intelligent optimization algorithms significantly enhance the predictive performance of the ANN model. Among them, the PSO-ANN model demonstrates optimal performance, with <i>R</i>² = 0.92, MSE = 0.007, and MAE = 0.0632, followed by the GA-ANN model and the GWO-ANN model. SHAP analysis reveals that the W/C, C/S, and curing age are the key variables influencing the compression strength. Furthermore, the size of the dataset does not significantly influence the computation time for the above models but is primarily governed by the complexity of the optimization algorithms. This study provides an efficient data-driven method for the mix design of RCM and a theoretical support for its engineering applications.