Bhanu Magotra , Manabendra Saharia, C. T. Dhanya

Physically informed deep learning models, especially Long Short-Term Memory (LSTM) networks, have shown promise in large-scale streamflow simulations. However, an in-depth understanding of the relative contribution of physical information in deep learning models has been missing. Using a large-sample testbed of 220 catchments in hydrologically diverse regions of the Indian subcontinent, we quantify the impact of incremental addition of physical information on model performance using multiple variants of the LSTM model based on various combinations of static catchment attributes and simulated land surface states. We found that LSTM models trained with catchment geophysics as additional input outperformed the base LSTM model in terms of the nationwide median Kling-Gupta Efficiency (KGE) of in-sample catchments, increasing KGE from 0.32 to 0.60. Additionally, the model retained significant prediction skill in out-of-sample catchments, demonstrating that a pre-trained LSTM model can be a powerful tool to predict streamflow in data-scarce regions.