Strength Prediction of Cement-Stabilized Steel Slag Using Deep Learning and SHAP Analysis

TL;DR

This study uses deep learning and SHAP analysis to predict and explain the strength of cement-stabilized steel slag, finding that optimal performance occurs at 60% steel slag content.

Contribution

A novel CNN-GRU-Attention model is introduced for strength prediction of CSSS with high accuracy and SHAP analysis for interpretability.

Findings

CSSS strength increases nonlinearly with curing age and peaks at 60% steel slag content.

Microstructural analysis shows AFt formation and gel network densification enhance strength.

The CNN-GRU-Attention model achieves R2 scores of 0.9875 for UCS and 0.9911 for STS with high accuracy.

Abstract

This study combined experimental analysis with deep learning to investigate the effects of curing age, steel slag content, and gradation composition on the mechanical properties of cement-stabilized steel slag (CSSS). The strength evolution patterns and underlying microscopic mechanisms were systematically elucidated. Experimental results showed that CSSS strength grows nonlinearly with curing age, with optimal mechanical performance achieved at a 60% steel slag content. The microstructural evolution characterized by SEM-EDS and XRD revealed that steel slag incorporation promotes the formation of AFt and densifies the gel network. In later curing stages, natural carbonation of Ca(OH)2 and secondary hydration of reactive steel slag components produce CaCO3 and additional C-S-H gel, which fill pores and significantly enhance long-term strength. A CNN-GRU-Attention model was developed to predict the unconfined compressive strength (UCS) and splitting tensile strength (STS) of CSSS. In a single data split, the model achieved R2 values of 0.9875 for UCS and 0.9911 for STS, with RMSEs of 0.2577 MPa and 0.0234 MPa, and MAEs of 0.2059 MPa and 0.0184 MPa, outperforming all benchmark models. Under rigorous 5 × 5 repeated cross-validation, it maintained the highest average R2 (UCS: 0.9417, STS: 0.9329) and the lowest error metrics, confirming its robustness and generalization capability. SHAP and Pearson correlation analyses identified cement content as the primary strength determinant, while steel slag content exhibited a threshold effect, highlighting the importance of prudent gradation control in practical engineering. This study provides both a theoretical foundation and a methodological framework for analyzing variable interactions and predicting the strength development of CSSS.

Figures