A hybrid flower pollination algorithm–deep learning approach for strength prediction of sustainable recycled fine aggregate concrete with GUI-based implementation

Abstract

This study presents an integrated experimental–machine learning framework for evaluating and predicting the compressive strength of sustainable rigid pavement concrete incorporating 50% washed recycled fine aggregates (WRFA) as a replacement for natural fine aggregates, together with zirconia silica fume (ZSF) and steel slag (SS) as supplementary cementitious materials. The novelty lies in the development of a hybrid Flower Pollination Algorithm–optimized deep neural network (FPA–DNN) that enhances prediction accuracy, robustness to noise, and interpretability through SHAP-based analysis, alongside a GUI-enabled decision-support tool for real-time application. Experimental findings indicated that 50% WRFA reduced compressive strength by 29.6%, 25.2%, and 23.9% at 7, 28, and 90 days, respectively, relative to the control mix. Among the investigated formulations, 20% SS cement replacement (SS20) provided the most favorable performance, limiting flexural strength reduction to approximately 3% and split tensile strength loss to 3.8% at 28 days while maintaining pavement-grade requirements. SEM-based microstructural evaluation of the optimal WRFA50 + SS20 mixture confirmed a denser cementitious matrix, reduced porosity, and improved interfacial transition zone bonding, attributed to SS-induced secondary hydration. A dataset comprising 264 samples (103 laboratory-generated and 161 collected from screened literature) was used to train and compare five regression models: kNN, RF, ANN, DNN, and FPA-DNN. Using 10-fold cross-validation, the FPA-DNN achieved the highest predictive accuracy, yielding testing performance of R² = 0.96 ± 0.006, RMSE = 2.87 ± 0.09 MPa, and MAPE = 4.38 ± 0.13%, outperforming kNN (R² = 0.85), RF (0.86), ANN (0.89), and standalone DNN (0.91). A noise-robustness assessment, conducted by applying additive white Gaussian noise to the target variable (σ up to ~ 1.6% of mean compressive strength), further confirmed the stability of the proposed model, with FPA-DNN retaining R² > 0.90 at the highest noise level (p = 0.20), whereas conventional models exhibited greater degradation. SHapley Additive ExPlanations (SHAP)-based interpretability identified cement content and steel slag as the most influential positive predictors, while WRFA showed a replacement-dependent effect. The proposed framework offers an interpretable and noise-resilient approach for strength prediction of recycled aggregate concrete; however, the present study is limited to a fixed WRFA replacement level and compressive-strength-focused modelling. Future work should extend validation to broader WRFA sources, wider replacement ranges, and durability-based performance indicators for long-term pavement applications.