Predicting the compressive strength of Steel Fiber Reinforced Concrete (SFRC) is challenging due to the complex, coupled influences of numerous mix design parameters. Traditional experimental methods are often costly and time-consuming for exploring this high-dimensional factor space. While machine learning (ML) models offer a powerful alternative, their practical application is hindered by two major limitations: the scarcity of comprehensive, high-quality experimental data which restricts model performance and generalizability, and the inherent "black-box" nature of advanced ML algorithms which obscures the underlying mechanisms linking mix proportions to mechanical properties. This study aims to address these critical gaps by developing a hybrid framework that integrates ML-based prediction, data augmentation via Generative Adversarial Networks (GAN), and post-hoc model interpretability using SHapley Additive exPlanations (SHAP). The objective is not only to achieve high-precision strength prediction but also to enhance model robustness with limited data and, crucially, to uncover the influential mechanisms of key components on SFRC compressive strength.

A database of 141 mix designs with 11 input features—Cement (C), Fly Ash (FA), Silica Fume (SF), Water (W), Coarse Sand (CS), Water Reducer (WR), Recycled Aggregate (RA), Natural Aggregate (NA), Fiber Length (FL), Fiber Diameter (FD), and Fiber Content Ratio (FCR)—and the corresponding compressive strength was compiled from peer-reviewed literature. Four tree-based ML algorithms (Random Forest, XGBoost, Stacking, and AdaBoost) were employed for regression modeling. To mitigate the data scarcity issue, a GAN architecture was implemented to generate 282 synthetic but physically plausible samples, expanding the training set. The quality of the GAN-generated data was validated using t-Distributed Stochastic Neighbor Embedding (t-SNE) visualization and statistical moment analysis. The model trained on the augmented dataset was evaluated using the coefficient of determination (R²) and Mean Squared Error (MSE). Finally, the optimal model (Random Forest) was interpreted using the SHAP framework to quantify and visualize the contribution and interaction effects of each input feature on the model's predictions.

Using the original dataset, all ML models demonstrated competent predictive capability. The Random Forest model achieved the best performance with an R² of 0.93 and an MSE of 15.6 MPa², followed closely by XGBoost (R²=0.92, MSE=16.6 MPa²). The introduction of GAN-generated data led to a remarkable and consistent improvement in all models' performance. For Random Forest, the R² increased to 0.98 and the MSE decreased by approximately 56% to 6.9 MPa². The Stacking model achieved an R² of 0.99. This significant enhancement confirms that GAN-based data augmentation effectively enriches the feature space and improves the models' ability to generalize from limited experimental data.

The SHAP analysis provided profound insights into the governing factors of compressive strength. The summary plot identified cement, water, and silica fume as the three most influential features. Cement and silica fume exhibited strong positive SHAP values, highlighting their fundamental role in forming a dense cementitious matrix. The impact of water displayed a distinct non-linear relationship: optimal content promoted hydration and strength, while excess water reduced strength, likely by increasing porosity. The interaction SHAP dependence plots revealed nuanced coupling effects. For instance, the positive contribution of a high WR dosage was more pronounced and stable under conditions of lower water content, emphasizing the critical coupling between water and chemical admixtures in determining the water-to-cementitious materials ratio. In contrast, parameters related to fiber geometry (FL, FD) and content (FCR) showed relatively low and concentrated SHAP values near zero, indicating their secondary role in influencing compressive strength, which aligns with their primary function of improving toughness rather than pure compressive capacity. The SHAP force plots and Sankey diagram further illustrated how specific combinations of high cement/silica fume and optimal water content drive predictions toward high strength, whereas high coarse sand content or inappropriate water levels contribute to lower strength outcomes.

This work successfully developed an integrated framework combining ML, GAN, and SHAP for predicting and interpreting the compressive strength of SFRC. The main conclusions are as follows:1) Among the evaluated ML models trained on the original dataset, Random Forest provided the most accurate predictions for SFRC compressive strength.2) GAN-based data augmentation proved highly effective, substantially improving the predictive accuracy and generalization ability of all models, with performance metrics (R² and MSE) showing dramatic improvement.3) SHAP interpretability analysis quantitatively identified cement, silica fume, and water as the most critical factors affecting compressive strength, with water exhibiting a non-linear effect. The analysis also elucidated important feature interactions, such as that between water and water reducer. Fiber parameters were found to have a minor influence on compressive strength prediction.4) The proposed framework offers a powerful, data-driven tool that moves beyond mere "black-box" prediction. It enhances prediction robustness from small datasets and provides actionable, mechanism-based insights for optimizing SFRC mix design, thereby bridging the gap between empirical data mining and fundamental materials science understanding.