<p>A conditions-aware deep-learning framework based on a Bayesian-optimized Hierarchical Mixture-of-Experts (BO-HME) is introduced to predict steel corrosion rates in cementitious mortars, addressing the one-size-fits-all limitation of single global models. The approach aligns with corrosion mechanisms by using a gating network that learns exposure and chemistry regimes and routes each sample to specialized experts. A consolidated dataset of 275 specimens with 14 inputs spanning mixture, environmental/material, and electrochemical descriptors is analyzed under a leakage-safe pipeline with nested cross-validation; Bayesian optimization (Tree-structured Parzen Estimator) tunes architecture and training controls. Complementary parametric analysis combines linear screening, principal component analysis, and a Gradient Boosting explainer with SHAP to quantify nonlinear and interaction effects. On the test set, BO-HME attains R² = 0.9572, RMSE = 2.1378, MAE = 1.2243, and MAPE = 3.36%, outperforming multiple benchmark models under an identical pipeline; relative to leading tree-boosting alternatives, RMSE is reduced by about 20 ~ 25%. Predicted-versus-measured plots indicate good calibration over most of the range, with mild under-prediction confined to the highest corrosion rates. The analysis identifies electrochemical state as dominant for prediction, with pH most influential, followed by electrical resistivity, corrosion potential, and chloride-to-hydroxide ratio. A Streamlit interface reproduces the trained pipeline and enables schema-checked single-sample and batch prediction with CSV or JSON export. The approach yields an accurate and deployable tool for durability assessment, threshold-based alerts, and maintenance planning within structural health monitoring workflows.</p>