Predicting daily flow duration curves (FDCs) in ungauged catchments remains a major challenge in hydrology and is critical for effective water resources management. The FDCs typical were predicted by relating FDC parameters or percentiles to catchment properties using statistical or machine learning-based models. Such models often suffer from limited interpretability and transferability across hydroclimatic conditions. In this study, we propose a process-informed, interpretable machine learning framework for predicting daily FDCs by integrating multivariate adaptive regression splines (MARS) with the Budyko theory, which provides a physically based representation of long-term water–energy constraints on catchment behaviour. Assuming the FDC follows a log-normal distribution determined by three parameters, MARS is used to explore relationships between FDC parameters and 19 catchment characteristics using data from 347 catchments across Australia. Results indicate that the proposed framework can satisfactorily predict the mean and deviation of daily streamflow, while prediction skill for ratio of non-zero flow days remains comparatively weaker. Incorporating hydrological constraints through Budyko theory improves the physical interpretability and robustness of model predictions, particularly in revealing the dominant hydroclimatic controls on streamflow regimes. Prediction performance is found generally higher in wetter catchments than in drier ones, mainly due to limitations of the models in predicting non-zero flow ratio and the lognormal assumption. To further improve FDC prediction in ungauged catchments, catchment characteristics more closely related to groundwater processes may be required, in addition to the adoption of more advanced modelling approaches.
Zhu et al. (2026) studied this question.