Vertical hydraulic conductivity of ephemeral streambeds is one of the key parameters in surface‑subsurface flow exchange and water resource management in arid and semi‑arid regions. However, its direct measurement is associated with high cost, time, and uncertainty. The aim of this study is to develop and evaluate a machine‑learning‑based framework for predicting vertical hydraulic conductivity of the bed, emphasizing the role of morphological heterogeneities such as the main channel and bars, to understand the behavior of ephemeral rivers. To this end, reinforcement‑based models were developed using grain‑size data, bed structural indices, and hydraulic parameters. The models performance was evaluated with three validation methods. Results show that the selected model achieved a coefficient of determination above 0.85 on the test data. Uncertainty analysis indicated that the 95 % confidence intervals of the predictions were narrow, and the model skill rate exceeded 85 % in most validation scenarios. Comparing model performance across different morphological units revealed that predicting hydraulic conductivity in bars was more stable than in the main channel, with the standard deviation of test results ranging from 10 % to 50 % lower depending on the validation method. Sensitivity analysis results indicate that the parameters mp and Ar, along with the 10th percentile grain size (d10), are the most important controlling factors for vertical hydraulic conductivity. These findings suggest that integrating machine learning with morphology‑based analysis can provide an efficient approach for estimating hydraulic conductivity of ephemeral streambeds and reducing uncertainty in hydrological studies.