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

Despite their physical complexity, Land Surface Models (LSMs) often fail to outperform simple statistical benchmarks in predicting turbulent heat fluxes. This points to a fundamental problem: traditional parameterization schemes are leaving information on the table. We address this by embedding a feedforward neural network directly into the governing equations of a mass- and energy-conserving LSM, replacing static flux parameterizations with data-driven representations learned from FLUXNET observations. The result is a hybrid differentiable model that both respects physical laws and leverages the pattern-recognition power of machine learning. We test the approach at two contrasting sites, a mature ponderosa pine forest in Oregon and a semi-arid grassland in Arizona. The hybrid model outperforms a calibrated physics-only baseline at both, with marked reductions in RMSE and bias in latent heat flux predictions. Ablation studies on neural network inputs reveal something telling. Providing the MLP with internal model states improves prediction accuracy of land surface fluxes, suggesting that physics-derived states also add information and aid our feedforward neural network. The results offer an optimistic view that physics and machine learning components could act synergistically in land surface modeling.