<p>Hybrid fiber-reinforced self-compacting concrete (HFR-SCC) is increasingly employed in structural applications requiring enhanced ductility and durability. However, its performance under elevated temperatures remains difficult to predict due to the complex interactions between mixture constituents, fiber degradation, and thermal damage mechanisms. This study proposes a novel data-driven framework based on the TabNet deep learning architecture to forecast the residual compressive and flexural strengths of HFR-SCC exposed to high temperatures. A diverse experimental dataset comprising 114 samples was compiled from the literature, incorporating eight key input parameters including binder composition, aggregate content, fiber dosage, and thermal exposure conditions. The TabNet model, optimized via Bayesian hyperparameter tuning, demonstrated excellent predictive accuracy and generalization capability, achieving R² values exceeding 0.98 and low error metrics across both training and testing sets. Comparative evaluations against seven conventional machine learning models—including ensemble and kernel-based approaches—highlighted TabNet’s superior performance, particularly in balancing accuracy and robustness. Importantly, TabNet’s intrinsic interpretability revealed that exposure temperature, slag content, and fiber volume were the most influential factors governing residual mechanical behavior. These findings affirm the potential of attention-based deep learning models to support reliable, interpretable, and efficient evaluation of fire-exposed concrete structures, advancing the integration of machine learning in materials engineering practice.</p>