A Hybrid Differentiable Land Surface Model for Improved Land-Atmosphere Flux Predictions
Nabin Kalauni1 and Andrew Bennett1
1 Department of Hydrology and Atmospheric Sciences
The University of Arizona, Tucson, AZ
Accurate modeling of water and energy fluxes at the land-atmosphere interface is critical for understanding and predicting Earth system dynamics, including climate variability, ecosystem resilience, and water resource management. Land surface models (LSMs), based on parametrized differential equations, have traditionally been used to model water and energy exchanges, with parameters typically calibrated on a site-specific basis. However, the advent of recurrent neural networks, specifically the Long Short-Term Memory Networks (LSTMs), has opened the possibility of developing a globally calibrated model capable of generalizing across diverse sites. In this study, we first develop a conceptual LSM with coupled mass and energy balance. This LSM is dynamically parametrized by an LSTM network to model dominant hydrologic and thermodynamic processes. We train our hybrid model using the FLUXNET dataset, which provides high-quality, globally distributed observations of meteorology and land-atmosphere exchanges. By combining physics-based constraints from LSMs with the data-driven learning capabilities of LSTMs, we aim to enhance the accuracy, generalizability, and interpretability of latent and sensible heat flux predictions across diverse ecosystems and climatic conditions. The research will systematically evaluate the hybrid model's performance, comparing it to a standalone version of our LSM and a pure machine learning approach. Additionally, we will explore the model's capability to predict variables not included in training, such as runoff, and assess its inference of state variables. This comprehensive evaluation will help demonstrate the potential of hybrid modeling for improving Earth system predictions by advancing our understanding of land-atmosphere interactions.