From Gauged to Ungauged: Using Catchment Morphology to Guide the Transfer of Physics‐Informed Neural Network Discharge Predictions

ABSTRACT Predicting discharge in ungauged catchments remains a fundamental challenge in hydrology, with direct consequences for flood risk management, water resource planning, and nature‐based flood mitigation. This study develops a transferability framework that combines ensemble clustering with hybrid Physics‐Informed Neural Network (PINN) and Gated Recurrent Unit (GRU) models. The framework uses the geomorphological similarity of donor catchments to transfer models to target catchments. The framework's limits were tested across three elements: the donor calibration, within‐cluster transfer with similar morphology, and outside‐cluster transfer where similarity was intentionally ignored. Ten catchment morphometrics, including catchment area, elongation ratio, form factor, relief ratio, drainage density, median elevation, longitudinal profile concavity, and percentages of confined, partly confined, and laterally unconfined valley settings, were used to identify morphologically similar clusters of catchments at 117 gauges in coastal New South Wales (NSW), Australia. An ensemble approach (K‐Means, Hierarchical, DBSCAN, GMM) was used for the clustering. PINN‐GRU models trained on high‐performing donor gauges within each cluster achieved strong within‐cluster transferability, with average R 2 and NSE values of approximately 0.70 across ungauged catchments. Conversely, transfer outside the cluster led to systematic performance decline, indicating that ignoring morphological similarity invalidates model transfer. Comprehensive‐input models (rainfall, temperature, static catchment attributes) outperformed simpler setups when transferring within clusters but showed the greatest vulnerability when transferring outside clusters. These findings emphasise that model transfer must occur between morphologically similar catchments and that input complexity benefits are conditional on preserving this transferability criteria between donor and target catchments. This framework offers a practical, scalable solution for discharge prediction in ungauged catchments, with direct applications in flood forecasting, historical flow reconstruction, and flood risk planning in data‐scarce regions.