Abstract: Evaporation ducts over the sea confine electromagnetic energy to a near-surface layer and give rise to over-the-horizon propagation; therefore, accurate retrieval of duct characteristics is crucial for maritime radio-communication and radar detection systems. To address the high cost associated with direct evaporation duct measurements, this paper proposes a path-loss-driven hybrid learning framework for efficient and robust prediction of evaporation duct characteristics. Path loss data are first generated using the parabolic equation method. A set of heterogeneous deep neural networks is then constructed as base learners by performing global hyperparameter optimization with the tree-structured Parzen estimator. Learning rate scheduling and early stopping are further employed to enhance model stability and generalization. Finally, a Stacking meta-learner is introduced to adaptively fuse the outputs of the base learners, thereby improving prediction accuracy and robustness. Simulation experiments demonstrate that, under noise-free conditions, the proposed framework reduces the RMSE by approximately 88.9% and 49.9% compared with the multilayer perceptron and deep neural network methods, respectively, and exhibits superior robustness under noisy conditions. These results indicate that the proposed method can accurately predict evaporation duct characteristics under complex maritime environments and provides an effective solution for evaporation duct inversion at sea.