This study aims to develop and validate advanced machine learning models for predicting the punching shear strength of steel fiber-reinforced concrete (SFRC) slab-column connections. Using a curated dataset of 377 experimental results, 36 outliers were identified and removed using Cook's distance approach, which resulted in a refined dataset of 341 samples. Predictive frameworks were constructed employing artificial neural networks (ANN), Categorical Boosting (CatBoost), and a hybrid Auto search optimization (ATOM)-ANN approach. The performance of these models was rigorously evaluated against established design formulas to assess their predictive accuracy and robustness. The CatBoost model outperformed its counterparts, achieving a mean absolute error (MAE) of 0.007 and a coefficient of determination of 0.921 on the validation set, demonstrating its superior ability to handle highly nonlinear relationships. SHAP analysis identified critical factors influencing punching shear strength, including concrete strength, fiber volume content, reinforcement ratio, slab thickness, and column dimensions. Results highlighted optimal reinforcement ratios between 1.5 % and 1.7 %, beyond which fiber congestion compromises strength. Additionally, four nonlinear design models were introduced, with Model-2 providing the best performance among the proposed formulations. Benchmarking against existing design methods revealed moderate predictive capabilities, but their limitations in addressing nonlinear behaviors underscore the need for advanced approaches.