Interfacial Area Concentration (IAC) significantly impacts the performance of two-fluid models in thermal–hydraulic codes. However, predicting IAC in rod bundle geometry is challenging due to data sparsity, geometric complexities, and the limitations of existing two-group IAC models. This study proposes a two-stage ANN model for predicting IAC in vertical rod bundle geometry, adopting SHAP (SHapley Additive exPlanations) for feature selection. A two-layer ANN architecture is trained on 425 experimental samples using dimensionless input features, with the Laplace parameter transforming the target variable into a dimensionless form. Two stages of training are performed, with the second stage utilizing a reduced feature set filtered based on SHAP values from the first stage. This simplifies the model by retaining only the most influential features, improving interpretability and reducing overfitting risks. Results show that the second stage achieves performance comparable to the first stage, with only a slight margin of difference, while significantly reducing model complexity. The proposed model outperforms existing two-group IAC models across various flow regimes, as validated by the Mean Relative Error (MRE) metric. The trade-off between a minor reduction in efficiency and the benefits of simplification—such as improved interpretability, computational efficiency, and robustness—is deemed acceptable. This study highlights the potential of ANN models combined with SHAP-based feature selection to enhance IAC prediction in complex geometries, offering a robust and efficient alternative to traditional approaches.