Irreversible electroporation (IRE) is a promising non-thermal ablation technique for the treatment of deep-seated tumors. Its clinical efficacy, however, critically depends on accurate patient-specific prediction of the electric field (EF) distribution. While physics-based numerical solvers provide high-fidelity simulations, their computational cost limits their use for intraoperative guidance and real-time treatment adaptation. This work addresses this limitation by combining methodological advances in fast numerical modeling with clinical outcome validation. We propose a hybrid learning–numerical framework that couples a convolutional neural network with a lightweight iterative correction scheme to rapidly compute physically consistent electric potential and EF maps from electrode configurations and tissue properties. The neural network provides an initial solution, which is subsequently refined to enforce physical constraints, enabling fast and reliable dosimetry compatible with clinical workflows and intraoperative imaging updates. The method is evaluated on synthetic datasets and 15 patient-specific cases using high-resolution domains (100×100×100 voxels, 1 mm³). Under homogeneous conductivity assumptions, the hybrid solver achieves a 15-fold speedup compared to conventional numerical solvers while preserving dosimetric accuracy. In heterogeneous tissue settings, it outperforms classical approaches in accuracy while maintaining more than a twofold computational acceleration. In parallel, retrospective numerical simulations were performed on 31 IRE procedures for hepatocellular carcinoma to investigate the relationship between EF coverage and clinical outcomes. Logistic regression analysis revealed that insufficient tumor coverage by EF isolines strongly correlates with local treatment failure, with the 400 V/cm threshold providing the highest predictive value (ROC-AUC = 0.904). Together, these results demonstrate that fast, patient-specific EF simulation is both computationally feasible and clinically meaningful. The proposed framework enables near–real-time IRE dosimetry --within a few tens of seconds-- while preserving predictive relevance for treatment outcome, supporting more adaptive and precise ablation strategies.