GA–PSO–optimised dual-path attention network for predicting strength of nano/micro-modified alkali-activated concrete

Abstract

We present a novel, chemically-aware framework for predicting the compressive strength of nano-/micro-modified alkali-activated concrete subjected to multi-ionic exposure. A comprehensive dataset of 324 unique mixes—varying binder precursor, nano- and micro-additives, aggregates, silicate–hydroxide ratio, superplasticizer dosage, curing temperature, and ionic exposure—is assembled. We engineer a Chemical Aggressivity Index (CAI) to quantify combined chemical effects and propose a Dual-Path Attention Network (DPAN) that processes material and exposure features in parallel. A hybrid Genetic Algorithm–Particle Swarm Optimisation (GA–PSO) simultaneously tunes network hyperparameters and feature weights, yielding an optimised DPAN with \(R^2=0.90\), MAE = 2.98 MPa, and RMSE = 4.21 MPa on the test set—surpassing linear regression, SVR-RBF, Random Forest, and XGBoost. Monte Carlo dropout provides reliable uncertainty bands, while SHAP analysis reveals that precursor content, acid concentrations, and CAI most strongly influence strength. The proposed methodology advances data-driven mix design by capturing complex chemical–mechanical interactions and offering actionable insights for resilient, sustainable concrete in aggressive environments.