The use of atmospheric forcing data in hydrological models often introduces biases and errors, which can lead to inaccurate simulations of key fluxes such as streamflow and evapotranspiration (ET). To address these issues, data fusion using deep learning methods like long short-term memory (LSTM) networks has shown promise. However, the complex nature of these networks makes it difficult to interpret individual biases. In response to this challenge, we propose an innovative approach using a physics-informed, differentiable hydrological model. This model enhances learning for data fusion, improved parameterization, and correction of atmospheric forcing biases. It assigns static or time-dynamic weights to various datasets. These weights facilitate the fusion of multiple datasets or serve as dynamic correction factors to adjust inherent biases in individual datasets. Our approach achieved performance levels comparable to LSTM models and demonstrated improvements in high and low flow metrics, while maintaining ET prediction an accuracy. This indicates an overall enhancement in the hydrological model's structural integrity. We will also share some recent developmental updates on hybrid model structures for better capturing baseflow and how to account for the impacts of scale (basin area). Data characterizing basin attributes are crucial for enhancing the generalizability of the hydrological model. However, certain datasets, like land use information, present challenges in effective utilization due to their inherent high uncertainty, non-Gaussian distribution, and implicit functions. In this context, we have implemented a specialized framework within the differentiable hydrological model that concretely delineates the impact of temporal changes in land use by refining the parameterization process. Streamflow simulations conducted across approximately 3200 basins, guided by the PUB test, reveal that the integration of temporally coarse land use data through this innovative framework not only augments model performance but also provides a clear depiction of the influence of land use data on model parameters.The use of atmospheric forcing data in hydrological models often introduces biases and errors, which can lead to inaccurate simulations of key fluxes such as streamflow and evapotranspiration (ET). To address these issues, data fusion using deep learning methods like long short-term memory (LSTM) networks has shown promise. However, the complex nature of these networks makes it difficult to interpret individual biases. In response to this challenge, we propose an innovative approach using a physics-informed, differentiable hydrological model. This model enhances learning for data fusion, improved parameterization, and correction of atmospheric forcing biases. It assigns static or time-dynamic weights to various datasets. These weights facilitate the fusion of multiple datasets or serve as dynamic correction factors to adjust inherent biases in individual datasets. Our approach achieved performance levels comparable to LSTM models and demonstrated improvements in high and low flow metrics, while maintaining ET prediction accuracy. This indicates an overall enhancement in the hydrological model's structural integrity. We will also share some recent developmental updates on hybrid model structures for better capturing baseflow and how to account for the impacts of scale (basin area). Data characterizing basin attributes are crucial for enhancing the generalizability of the hydrological model. However, certain datasets, like land use information, present challenges in effective utilization due to their inherent high uncertainty, non-Gaussian distribution, and implicit functions. In this context, we have implemented a specialized framework within the differentiable hydrological model that concretely delineates the impact of temporal changes in land use by refining the parameterization process. Streamflow simulations conducted across approximately 3200 basins, guided by the PUB test, reveal that the integration of temporally coarse land use data through this innovative framework not only augments model performance but also provides a clear depiction of the influence of land use data on model parameters.