Data-augmented machine learning for predicting flexural strength of steel fiber-reinforced concrete in underground structures

In underground structures, steel fiber-reinforced concrete (SFRC) has been widely applied for its notable ability to enhance the strength, toughness, and durability of the cementitious matrix, while also providing considerable economic benefits. However, the addition of steel fibers introduces extra complexities that affect performance, increasing the variability and uncertainty of material behavior and thus making mix proportioning and structural design more challenging. To address these issues, this study proposes an advanced data-driven framework for predicting the flexural strength of SFRC, which integrates data collection, sample expansion, machine learning prediction, and interpretability analysis. Specifically, a rigorously curated database of 195 experimental samples was established, incorporating key variables like fiber geometry, dimension, dosage, and coarse aggregate particle size. The dataset was then expanded to 500 samples using principal component analysis and Gaussian Copula, while preserving both the statistical distributions and overall variable trends. Subsequently, five advanced machine learning algorithms were employed to construct predictive models of flexural strength, with hyperparameters carefully optimized to avoid overfitting and underfitting. Finally, the best-performing model was further interpreted using Shapley additive explanations, partial dependence plots, and individual conditional expectation. Overall, the proposed framework demonstrates solid theoretical foundations and methodological rigor, and the predictive strategy is expected to provide valuable insights for addressing performance prediction challenges in other materials and structural systems.