The evaluation of the mechanical performance of fly ash-recycled mortar (FARM) is a necessary condition to ensure the efficient utilization of recycled fine aggregates. This article describes the design of nine mix proportions of FARMs with a low water/cement ratio and screens six mix proportions with reasonable flowability. The compressive strengths of FARMs were tested, and the influence of the water/cement ratio (w/c) and age on the compressive strength was analyzed. Meanwhile, a backpropagation neural network (BPNN) model optimized by the grey wolf optimizer (GWO), namely the GWO-BPNN model, was established to predict the compressive strength of FARM. The input layer of the model consisted of w/c, a cement/sand ratio, water reducer, age, and fly ash content, while the output layer was the compressive strength. The data set consisted of 150 sets from this article and existing research in the literature, of which 70% is used for model training and 30% for model validation. The results show that compared with the traditional BPNN, the coefficient of determination (R2) of GWO-BPNN increases from 0.85 to 0.93, and the mean squared error (MSE) of model training decreases from 0.018 to 0.015. Meanwhile, the convergence iterations of model validation decrease from 108 to 65. This indicates that GWO improved the prediction accuracy and computational efficiency of BPNN. The model results of characteristic heat, kernel density estimation, scatter matrix, and the SHAP value all indicated that the w/c was strongly negatively correlated with compressive strength, while the sand/cement ratio and age were strongly positively correlated with compressive strength. However, the relationship between the contents of fly ash, the water reducer, and the compressive strength was not obvious.
Keywords: backpropagation neural network; compressive strength; fly ash-recycled mortar; grey wolf optimizer; low water/cement ratio.