Radiofrequency ablation (RFA) has emerged as a promising minimally invasive technique for tumor treatment, offering reduced recovery time and postoperative pain. Accordingly, accurately simulating ablation dynamics—including electrode positioning and timing—is critical for minimizing recurrence and preventing damage to surrounding healthy tissue. While a numerical model provides physically reliable predictions, it is inherently slow and computationally intensive, rendering it unsuitable for real-time clinical applications. To address this issue, we propose a transformer-based U-Net (UNETR) model trained on numerically computed thermal effects to enable real-time prediction of the spatiotemporal dynamics of 3D ablation regions and temperature maps. The model is trained on data generated from a multi-physics simulation incorporating electrostatic field analysis, bio-heat transfer, and cell necrosis modeling, based on breast cancer MR images. It achieved a root mean square error (RMSE) of 0.7160 for temperature distribution and a Dice score of 94.38% for ablation regions on previously unseen MR anatomies, demonstrating strong generalization across anatomical variations. Furthermore, inference time was drastically reduced from 76.23 s, required by conventional numerical methods, to just 0.047 s, enabling real-time performance. These results demonstrate the feasibility of a digital twin for RFA, which holds promise for improving the safety and efficacy of personalized therapy. The code is available at: https://github.com/SeonAengCho/RFA-simulation-model.git .