Long-Term Compressive Behavior Evaluation for SSC and Prediction Using Machine Learning Models with SHAP Interpretation

This study investigates the time-dependent compressive behavior of seawater sea-sand concrete (SSC) and establishes intelligent prediction models for its evaluation. SSC samples with three strength grades (C30, C40, and C60) were fabricated and tested at curing ages of up to 720 d, accompanied by microstructural characterization. A database containing 344 compressive-strength records was compiled from the present experiments and published literature. Four models were developed and compared, including a back-propagation (BP) neural network, a genetic algorithm-optimized BP (GA-BP) network, a Bagging-BP ensemble model, and a gradient boosting regression tree (GBRT) model. Model interpretability was further examined using SHapley Additive exPlanations (SHAP). The results indicate that SSC exhibits rapid early-age strength gain, while moderate fluctuations may occur during middle and later curing periods. Strength–elastic modulus relationships were established for SSC and showed satisfactory fitting accuracy, with R2 values exceeding 0.80. Among the investigated models, GA-BP demonstrated favorable predictive performance under the adopted data split, achieving the lowest root mean square error with R = 0.9413. SHAP interpretation highlighted curing duration and water-to-cement ratio as the dominant factors governing compressive strength. Overall, the proposed GA-BP model can serve as a useful reference for SSC strength prediction and mix-design optimization within the current dataset.