This work presents a deep learning-based inverse design framework for rapid prototyping of fork antennas operating within the ultra-wideband (UWB) frequency. By integrating a pre-trained frozen forward model into the inverse training process, the proposed method with forward consistency loss ensures that predicted antenna geometries yield physically valid performance metrics. Three models were trained: Deep Neural Network (DNN), Long Short Term Memory (LSTM), and Echo State Network (ESN), to map performance targets to antenna dimensions. The results show that the models trained with forward consistency loss achieve a lower error in return loss, bandwidth, and frequency metrics. The proposed framework considering both forward loss and reverse loss outperforms inverse-only models, especially in complex high-frequency wideband scenarios. DNN and LSTM models demonstrated the best generalization at 3.5 GHz and 8 GHz, while all models significantly reduced the computational cost of antenna design compared to conventional simulation sweeps.