Electromagnetic look-ahead logging while drilling instruments detect the electrical characteristics of undrilled formations, enabling proactive decision-making. Real-time geological insight ahead of the drill bit is critical for effective geosteering. This study introduces a multi-task U-net neural network that simultaneously inverts multiple formation parameters real-time. Six datasets, each corresponding to different electromagnetic components, were used to train six specialized neural networks. All networks exhibited rapid convergence and successfully inverted 60,000 sample in 15 s, satisfying real-time requirements. Residual and relative error analyses reveal that the multi-component network delivers the highest accuracy. Sensitivity analysis shows that coaxial and coplanar components are more sensitive to conductivity variations, whereas coaxial and cross-components excel at resolving interface positions. The yy component displays the strongest sensitivity to anisotropy. Compared with the traditional Levenberg-Marquardt algorithm, the proposed method demonstrates improved accuracy and efficiency. Moreover, the Levenberg-Marquardt inversion with the neural network output as initial models further enhances accuracy. Benchmark comparisons reveal that the multi-task U-net outperforms various mainstream machine learning and deep learning models, including LSTM, FCN, ResNet, and XGBoost, in both inversion accuracy and generalization. Moreover, sensitivity analyses to noise and near-bit geological complexity reveal that, while the proposed model experiences some performance degradation under high noise levels or highly heterogeneous backgrounds, it maintains strong robustness under moderate noise conditions and achieves reliable inversion results in two-layer geological settings. These results establish the multi-task U-net as a fast, accurate, and robust tool for real-time electromagnetic look-ahead inversion in geosteering applications.